1 /*
2 * Copyright (c) 1998, 2019, 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 mcall->method_set((intptr_t)mcall->entry_point());
292
293 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
294 stub_size += CompiledStaticCall::to_interp_stub_size();
295 reloc_size += CompiledStaticCall::reloc_to_interp_stub();
296 #if INCLUDE_AOT
297 stub_size += CompiledStaticCall::to_aot_stub_size();
298 reloc_size += CompiledStaticCall::reloc_to_aot_stub();
299 #endif
300 }
301 } else if (mach->is_MachSafePoint()) {
302 // If call/safepoint are adjacent, account for possible
303 // nop to disambiguate the two safepoints.
304 // ScheduleAndBundle() can rearrange nodes in a block,
305 // check for all offsets inside this block.
306 if (last_call_adr >= blk_starts[i]) {
307 blk_size += nop_size;
308 }
309 }
310 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
311 // Nop is inserted between "avoid back to back" instructions.
312 // ScheduleAndBundle() can rearrange nodes in a block,
313 // check for all offsets inside this block.
314 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
315 blk_size += nop_size;
316 }
317 }
318 if (mach->may_be_short_branch()) {
319 if (!nj->is_MachBranch()) {
320 #ifndef PRODUCT
321 nj->dump(3);
322 #endif
323 Unimplemented();
324 }
325 assert(jmp_nidx[i] == -1, "block should have only one branch");
326 jmp_offset[i] = blk_size;
327 jmp_size[i] = nj->size(_regalloc);
328 jmp_nidx[i] = j;
329 has_short_branch_candidate = true;
330 }
331 }
332 blk_size += nj->size(_regalloc);
333 // Remember end of call offset
334 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
335 last_call_adr = blk_starts[i]+blk_size;
336 }
337 // Remember end of avoid_back_to_back offset
338 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
339 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
340 }
341 }
342
343 // When the next block starts a loop, we may insert pad NOP
344 // instructions. Since we cannot know our future alignment,
345 // assume the worst.
346 if (i < nblocks - 1) {
347 Block* nb = _cfg->get_block(i + 1);
348 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
349 if (max_loop_pad > 0) {
350 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
351 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
352 // If either is the last instruction in this block, bump by
353 // max_loop_pad in lock-step with blk_size, so sizing
354 // calculations in subsequent blocks still can conservatively
355 // detect that it may the last instruction in this block.
356 if (last_call_adr == blk_starts[i]+blk_size) {
357 last_call_adr += max_loop_pad;
358 }
359 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
360 last_avoid_back_to_back_adr += max_loop_pad;
361 }
362 blk_size += max_loop_pad;
363 block_worst_case_pad[i + 1] = max_loop_pad;
364 }
365 }
366
367 // Save block size; update total method size
368 blk_starts[i+1] = blk_starts[i]+blk_size;
369 }
370
371 // Step two, replace eligible long jumps.
372 bool progress = true;
373 uint last_may_be_short_branch_adr = max_juint;
374 while (has_short_branch_candidate && progress) {
375 progress = false;
376 has_short_branch_candidate = false;
377 int adjust_block_start = 0;
378 for (uint i = 0; i < nblocks; i++) {
379 Block* block = _cfg->get_block(i);
380 int idx = jmp_nidx[i];
381 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach();
382 if (mach != NULL && mach->may_be_short_branch()) {
383 #ifdef ASSERT
384 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
385 int j;
386 // Find the branch; ignore trailing NOPs.
387 for (j = block->number_of_nodes()-1; j>=0; j--) {
388 Node* n = block->get_node(j);
389 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
390 break;
391 }
392 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
393 #endif
394 int br_size = jmp_size[i];
395 int br_offs = blk_starts[i] + jmp_offset[i];
396
397 // This requires the TRUE branch target be in succs[0]
398 uint bnum = block->non_connector_successor(0)->_pre_order;
399 int offset = blk_starts[bnum] - br_offs;
400 if (bnum > i) { // adjust following block's offset
401 offset -= adjust_block_start;
402 }
403
404 // This block can be a loop header, account for the padding
405 // in the previous block.
406 int block_padding = block_worst_case_pad[i];
407 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
408 // In the following code a nop could be inserted before
409 // the branch which will increase the backward distance.
410 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
411 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
412
413 if (needs_padding && offset <= 0)
414 offset -= nop_size;
415
416 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
417 // We've got a winner. Replace this branch.
418 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
419
420 // Update the jmp_size.
421 int new_size = replacement->size(_regalloc);
422 int diff = br_size - new_size;
423 assert(diff >= (int)nop_size, "short_branch size should be smaller");
424 // Conservatively take into account padding between
425 // avoid_back_to_back branches. Previous branch could be
426 // converted into avoid_back_to_back branch during next
427 // rounds.
428 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
429 jmp_offset[i] += nop_size;
430 diff -= nop_size;
431 }
432 adjust_block_start += diff;
433 block->map_node(replacement, idx);
434 mach->subsume_by(replacement, C);
435 mach = replacement;
436 progress = true;
437
438 jmp_size[i] = new_size;
439 DEBUG_ONLY( jmp_target[i] = bnum; );
440 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
441 } else {
442 // The jump distance is not short, try again during next iteration.
443 has_short_branch_candidate = true;
444 }
445 } // (mach->may_be_short_branch())
446 if (mach != NULL && (mach->may_be_short_branch() ||
447 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
448 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
449 }
450 blk_starts[i+1] -= adjust_block_start;
451 }
452 }
453
454 #ifdef ASSERT
455 for (uint i = 0; i < nblocks; i++) { // For all blocks
456 if (jmp_target[i] != 0) {
457 int br_size = jmp_size[i];
458 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
459 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
460 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]);
461 }
462 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
463 }
464 }
465 #endif
466
467 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
468 // after ScheduleAndBundle().
469
470 // ------------------
471 // Compute size for code buffer
472 code_size = blk_starts[nblocks];
473
474 // Relocation records
475 reloc_size += 1; // Relo entry for exception handler
476
477 // Adjust reloc_size to number of record of relocation info
478 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
479 // a relocation index.
480 // The CodeBuffer will expand the locs array if this estimate is too low.
481 reloc_size *= 10 / sizeof(relocInfo);
482 }
483
484 //------------------------------FillLocArray-----------------------------------
485 // Create a bit of debug info and append it to the array. The mapping is from
486 // Java local or expression stack to constant, register or stack-slot. For
487 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
488 // entry has been taken care of and caller should skip it).
489 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
490 // This should never have accepted Bad before
491 assert(OptoReg::is_valid(regnum), "location must be valid");
492 return (OptoReg::is_reg(regnum))
493 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
494 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
495 }
496
497
498 ObjectValue*
499 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
500 for (int i = 0; i < objs->length(); i++) {
501 assert(objs->at(i)->is_object(), "corrupt object cache");
502 ObjectValue* sv = (ObjectValue*) objs->at(i);
503 if (sv->id() == id) {
504 return sv;
505 }
506 }
507 // Otherwise..
508 return NULL;
509 }
510
511 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
512 ObjectValue* sv ) {
513 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
514 objs->append(sv);
515 }
516
517
518 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
519 GrowableArray<ScopeValue*> *array,
520 GrowableArray<ScopeValue*> *objs ) {
521 assert( local, "use _top instead of null" );
522 if (array->length() != idx) {
523 assert(array->length() == idx + 1, "Unexpected array count");
524 // Old functionality:
525 // return
526 // New functionality:
527 // Assert if the local is not top. In product mode let the new node
528 // override the old entry.
529 assert(local == top(), "LocArray collision");
530 if (local == top()) {
531 return;
532 }
533 array->pop();
534 }
535 const Type *t = local->bottom_type();
536
537 // Is it a safepoint scalar object node?
538 if (local->is_SafePointScalarObject()) {
539 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
540
541 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
542 if (sv == NULL) {
543 ciKlass* cik = t->is_oopptr()->klass();
544 assert(cik->is_instance_klass() ||
545 cik->is_array_klass(), "Not supported allocation.");
546 sv = new ObjectValue(spobj->_idx,
547 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
548 Compile::set_sv_for_object_node(objs, sv);
549
550 uint first_ind = spobj->first_index(sfpt->jvms());
551 for (uint i = 0; i < spobj->n_fields(); i++) {
552 Node* fld_node = sfpt->in(first_ind+i);
553 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
554 }
555 }
556 array->append(sv);
557 return;
558 }
559
560 // Grab the register number for the local
561 OptoReg::Name regnum = _regalloc->get_reg_first(local);
562 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
563 // Record the double as two float registers.
564 // The register mask for such a value always specifies two adjacent
565 // float registers, with the lower register number even.
566 // Normally, the allocation of high and low words to these registers
567 // is irrelevant, because nearly all operations on register pairs
568 // (e.g., StoreD) treat them as a single unit.
569 // Here, we assume in addition that the words in these two registers
570 // stored "naturally" (by operations like StoreD and double stores
571 // within the interpreter) such that the lower-numbered register
572 // is written to the lower memory address. This may seem like
573 // a machine dependency, but it is not--it is a requirement on
574 // the author of the <arch>.ad file to ensure that, for every
575 // even/odd double-register pair to which a double may be allocated,
576 // the word in the even single-register is stored to the first
577 // memory word. (Note that register numbers are completely
578 // arbitrary, and are not tied to any machine-level encodings.)
579 #ifdef _LP64
580 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
581 array->append(new ConstantIntValue((jint)0));
582 array->append(new_loc_value( _regalloc, regnum, Location::dbl ));
583 } else if ( t->base() == Type::Long ) {
584 array->append(new ConstantIntValue((jint)0));
585 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
586 } else if ( t->base() == Type::RawPtr ) {
587 // jsr/ret return address which must be restored into a the full
588 // width 64-bit stack slot.
589 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
590 }
591 #else //_LP64
592 #ifdef SPARC
593 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {
594 // For SPARC we have to swap high and low words for
595 // long values stored in a single-register (g0-g7).
596 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
597 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
598 } else
599 #endif //SPARC
600 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
601 // Repack the double/long as two jints.
602 // The convention the interpreter uses is that the second local
603 // holds the first raw word of the native double representation.
604 // This is actually reasonable, since locals and stack arrays
605 // grow downwards in all implementations.
606 // (If, on some machine, the interpreter's Java locals or stack
607 // were to grow upwards, the embedded doubles would be word-swapped.)
608 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
609 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
610 }
611 #endif //_LP64
612 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
613 OptoReg::is_reg(regnum) ) {
614 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()
615 ? Location::float_in_dbl : Location::normal ));
616 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
617 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long
618 ? Location::int_in_long : Location::normal ));
619 } else if( t->base() == Type::NarrowOop ) {
620 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));
621 } else {
622 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));
623 }
624 return;
625 }
626
627 // No register. It must be constant data.
628 switch (t->base()) {
629 case Type::Half: // Second half of a double
630 ShouldNotReachHere(); // Caller should skip 2nd halves
631 break;
632 case Type::AnyPtr:
633 array->append(new ConstantOopWriteValue(NULL));
634 break;
635 case Type::AryPtr:
636 case Type::InstPtr: // fall through
637 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
638 break;
639 case Type::NarrowOop:
640 if (t == TypeNarrowOop::NULL_PTR) {
641 array->append(new ConstantOopWriteValue(NULL));
642 } else {
643 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
644 }
645 break;
646 case Type::Int:
647 array->append(new ConstantIntValue(t->is_int()->get_con()));
648 break;
649 case Type::RawPtr:
650 // A return address (T_ADDRESS).
651 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
652 #ifdef _LP64
653 // Must be restored to the full-width 64-bit stack slot.
654 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
655 #else
656 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
657 #endif
658 break;
659 case Type::FloatCon: {
660 float f = t->is_float_constant()->getf();
661 array->append(new ConstantIntValue(jint_cast(f)));
662 break;
663 }
664 case Type::DoubleCon: {
665 jdouble d = t->is_double_constant()->getd();
666 #ifdef _LP64
667 array->append(new ConstantIntValue((jint)0));
668 array->append(new ConstantDoubleValue(d));
669 #else
670 // Repack the double as two jints.
671 // The convention the interpreter uses is that the second local
672 // holds the first raw word of the native double representation.
673 // This is actually reasonable, since locals and stack arrays
674 // grow downwards in all implementations.
675 // (If, on some machine, the interpreter's Java locals or stack
676 // were to grow upwards, the embedded doubles would be word-swapped.)
677 jlong_accessor acc;
678 acc.long_value = jlong_cast(d);
679 array->append(new ConstantIntValue(acc.words[1]));
680 array->append(new ConstantIntValue(acc.words[0]));
681 #endif
682 break;
683 }
684 case Type::Long: {
685 jlong d = t->is_long()->get_con();
686 #ifdef _LP64
687 array->append(new ConstantIntValue((jint)0));
688 array->append(new ConstantLongValue(d));
689 #else
690 // Repack the long as two jints.
691 // The convention the interpreter uses is that the second local
692 // holds the first raw word of the native double representation.
693 // This is actually reasonable, since locals and stack arrays
694 // grow downwards in all implementations.
695 // (If, on some machine, the interpreter's Java locals or stack
696 // were to grow upwards, the embedded doubles would be word-swapped.)
697 jlong_accessor acc;
698 acc.long_value = d;
699 array->append(new ConstantIntValue(acc.words[1]));
700 array->append(new ConstantIntValue(acc.words[0]));
701 #endif
702 break;
703 }
704 case Type::Top: // Add an illegal value here
705 array->append(new LocationValue(Location()));
706 break;
707 default:
708 ShouldNotReachHere();
709 break;
710 }
711 }
712
713 // Determine if this node starts a bundle
714 bool Compile::starts_bundle(const Node *n) const {
715 return (_node_bundling_limit > n->_idx &&
716 _node_bundling_base[n->_idx].starts_bundle());
717 }
718
719 //--------------------------Process_OopMap_Node--------------------------------
720 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
721
722 // Handle special safepoint nodes for synchronization
723 MachSafePointNode *sfn = mach->as_MachSafePoint();
724 MachCallNode *mcall;
725
726 int safepoint_pc_offset = current_offset;
727 bool is_method_handle_invoke = false;
728 bool return_oop = false;
729 bool not_global_escape_in_scope = sfn->_not_global_escape_in_scope;
730 bool arg_escape = false;
731
732 // Add the safepoint in the DebugInfoRecorder
733 if( !mach->is_MachCall() ) {
734 mcall = NULL;
735 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
736 } else {
737 mcall = mach->as_MachCall();
738
739 // Is the call a MethodHandle call?
740 if (mcall->is_MachCallJava()) {
741 if (mcall->as_MachCallJava()->_method_handle_invoke) {
742 assert(has_method_handle_invokes(), "must have been set during call generation");
743 is_method_handle_invoke = true;
744 }
745 arg_escape = mcall->as_MachCallJava()->_arg_escape;
746 }
747
748 // Check if a call returns an object.
749 if (mcall->returns_pointer()) {
750 return_oop = true;
751 }
752 safepoint_pc_offset += mcall->ret_addr_offset();
753 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
754 }
755
756 // Loop over the JVMState list to add scope information
757 // Do not skip safepoints with a NULL method, they need monitor info
758 JVMState* youngest_jvms = sfn->jvms();
759 int max_depth = youngest_jvms->depth();
760
761 // Allocate the object pool for scalar-replaced objects -- the map from
762 // small-integer keys (which can be recorded in the local and ostack
763 // arrays) to descriptions of the object state.
764 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
765
766 // Visit scopes from oldest to youngest.
767 for (int depth = 1; depth <= max_depth; depth++) {
768 JVMState* jvms = youngest_jvms->of_depth(depth);
769 int idx;
770 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
771 // Safepoints that do not have method() set only provide oop-map and monitor info
772 // to support GC; these do not support deoptimization.
773 int num_locs = (method == NULL) ? 0 : jvms->loc_size();
774 int num_exps = (method == NULL) ? 0 : jvms->stk_size();
775 int num_mon = jvms->nof_monitors();
776 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
777 "JVMS local count must match that of the method");
778
779 // Add Local and Expression Stack Information
780
781 // Insert locals into the locarray
782 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
783 for( idx = 0; idx < num_locs; idx++ ) {
784 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
785 }
786
787 // Insert expression stack entries into the exparray
788 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
789 for( idx = 0; idx < num_exps; idx++ ) {
790 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
791 }
792
793 // Add in mappings of the monitors
794 assert( !method ||
795 !method->is_synchronized() ||
796 method->is_native() ||
797 num_mon > 0 ||
798 !GenerateSynchronizationCode,
799 "monitors must always exist for synchronized methods");
800
801 // Build the growable array of ScopeValues for exp stack
802 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
803
804 // Loop over monitors and insert into array
805 for (idx = 0; idx < num_mon; idx++) {
806 // Grab the node that defines this monitor
807 Node* box_node = sfn->monitor_box(jvms, idx);
808 Node* obj_node = sfn->monitor_obj(jvms, idx);
809
810 // Create ScopeValue for object
811 ScopeValue *scval = NULL;
812
813 if (obj_node->is_SafePointScalarObject()) {
814 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
815 scval = Compile::sv_for_node_id(objs, spobj->_idx);
816 if (scval == NULL) {
817 const Type *t = spobj->bottom_type();
818 ciKlass* cik = t->is_oopptr()->klass();
819 assert(cik->is_instance_klass() ||
820 cik->is_array_klass(), "Not supported allocation.");
821 ObjectValue* sv = new ObjectValue(spobj->_idx,
822 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
823 Compile::set_sv_for_object_node(objs, sv);
824
825 uint first_ind = spobj->first_index(youngest_jvms);
826 for (uint i = 0; i < spobj->n_fields(); i++) {
827 Node* fld_node = sfn->in(first_ind+i);
828 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
829 }
830 scval = sv;
831 }
832 } else if (!obj_node->is_Con()) {
833 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);
834 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
835 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );
836 } else {
837 scval = new_loc_value( _regalloc, obj_reg, Location::oop );
838 }
839 } else {
840 const TypePtr *tp = obj_node->get_ptr_type();
841 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
842 }
843
844 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
845 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
846 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
847 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
848 }
849
850 // We dump the object pool first, since deoptimization reads it in first.
851 debug_info()->dump_object_pool(objs);
852
853 // Build first class objects to pass to scope
854 DebugToken *locvals = debug_info()->create_scope_values(locarray);
855 DebugToken *expvals = debug_info()->create_scope_values(exparray);
856 DebugToken *monvals = debug_info()->create_monitor_values(monarray);
857
858 // Make method available for all Safepoints
859 ciMethod* scope_method = method ? method : _method;
860 // Describe the scope here
861 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
862 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
863 // Now we can describe the scope.
864 methodHandle null_mh;
865 bool rethrow_exception = false;
866 debug_info()->describe_scope(safepoint_pc_offset, null_mh, scope_method, jvms->bci(), jvms->should_reexecute(), rethrow_exception, is_method_handle_invoke, return_oop,
867 not_global_escape_in_scope, arg_escape,
868 locvals, expvals, monvals);
869 } // End jvms loop
870
871 // Mark the end of the scope set.
872 debug_info()->end_safepoint(safepoint_pc_offset);
873 }
874
875
876
877 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
878 class NonSafepointEmitter {
879 Compile* C;
880 JVMState* _pending_jvms;
881 int _pending_offset;
882
883 void emit_non_safepoint();
884
885 public:
886 NonSafepointEmitter(Compile* compile) {
887 this->C = compile;
888 _pending_jvms = NULL;
889 _pending_offset = 0;
890 }
891
892 void observe_instruction(Node* n, int pc_offset) {
893 if (!C->debug_info()->recording_non_safepoints()) return;
894
895 Node_Notes* nn = C->node_notes_at(n->_idx);
896 if (nn == NULL || nn->jvms() == NULL) return;
897 if (_pending_jvms != NULL &&
898 _pending_jvms->same_calls_as(nn->jvms())) {
899 // Repeated JVMS? Stretch it up here.
900 _pending_offset = pc_offset;
901 } else {
902 if (_pending_jvms != NULL &&
903 _pending_offset < pc_offset) {
904 emit_non_safepoint();
905 }
906 _pending_jvms = NULL;
907 if (pc_offset > C->debug_info()->last_pc_offset()) {
908 // This is the only way _pending_jvms can become non-NULL:
909 _pending_jvms = nn->jvms();
910 _pending_offset = pc_offset;
911 }
912 }
913 }
914
915 // Stay out of the way of real safepoints:
916 void observe_safepoint(JVMState* jvms, int pc_offset) {
917 if (_pending_jvms != NULL &&
918 !_pending_jvms->same_calls_as(jvms) &&
919 _pending_offset < pc_offset) {
920 emit_non_safepoint();
921 }
922 _pending_jvms = NULL;
923 }
924
925 void flush_at_end() {
926 if (_pending_jvms != NULL) {
927 emit_non_safepoint();
928 }
929 _pending_jvms = NULL;
930 }
931 };
932
933 void NonSafepointEmitter::emit_non_safepoint() {
934 JVMState* youngest_jvms = _pending_jvms;
935 int pc_offset = _pending_offset;
936
937 // Clear it now:
938 _pending_jvms = NULL;
939
940 DebugInformationRecorder* debug_info = C->debug_info();
941 assert(debug_info->recording_non_safepoints(), "sanity");
942
943 debug_info->add_non_safepoint(pc_offset);
944 int max_depth = youngest_jvms->depth();
945
946 // Visit scopes from oldest to youngest.
947 for (int depth = 1; depth <= max_depth; depth++) {
948 JVMState* jvms = youngest_jvms->of_depth(depth);
949 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
950 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
951 methodHandle null_mh;
952 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
953 }
954
955 // Mark the end of the scope set.
956 debug_info->end_non_safepoint(pc_offset);
957 }
958
959 //------------------------------init_buffer------------------------------------
960 CodeBuffer* Compile::init_buffer(uint* blk_starts) {
961
962 // Set the initially allocated size
963 int code_req = initial_code_capacity;
964 int locs_req = initial_locs_capacity;
965 int stub_req = initial_stub_capacity;
966 int const_req = initial_const_capacity;
967
968 int pad_req = NativeCall::instruction_size;
969 // The extra spacing after the code is necessary on some platforms.
970 // Sometimes we need to patch in a jump after the last instruction,
971 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
972
973 // Compute the byte offset where we can store the deopt pc.
974 if (fixed_slots() != 0) {
975 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
976 }
977
978 // Compute prolog code size
979 _method_size = 0;
980 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;
981 #if defined(IA64) && !defined(AIX)
982 if (save_argument_registers()) {
983 // 4815101: this is a stub with implicit and unknown precision fp args.
984 // The usual spill mechanism can only generate stfd's in this case, which
985 // doesn't work if the fp reg to spill contains a single-precision denorm.
986 // Instead, we hack around the normal spill mechanism using stfspill's and
987 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate
988 // space here for the fp arg regs (f8-f15) we're going to thusly spill.
989 //
990 // If we ever implement 16-byte 'registers' == stack slots, we can
991 // get rid of this hack and have SpillCopy generate stfspill/ldffill
992 // instead of stfd/stfs/ldfd/ldfs.
993 _frame_slots += 8*(16/BytesPerInt);
994 }
995 #endif
996 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
997
998 if (has_mach_constant_base_node()) {
999 uint add_size = 0;
1000 // Fill the constant table.
1001 // Note: This must happen before shorten_branches.
1002 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1003 Block* b = _cfg->get_block(i);
1004
1005 for (uint j = 0; j < b->number_of_nodes(); j++) {
1006 Node* n = b->get_node(j);
1007
1008 // If the node is a MachConstantNode evaluate the constant
1009 // value section.
1010 if (n->is_MachConstant()) {
1011 MachConstantNode* machcon = n->as_MachConstant();
1012 machcon->eval_constant(C);
1013 } else if (n->is_Mach()) {
1014 // On Power there are more nodes that issue constants.
1015 add_size += (n->as_Mach()->ins_num_consts() * 8);
1016 }
1017 }
1018 }
1019
1020 // Calculate the offsets of the constants and the size of the
1021 // constant table (including the padding to the next section).
1022 constant_table().calculate_offsets_and_size();
1023 const_req = constant_table().size() + add_size;
1024 }
1025
1026 // Initialize the space for the BufferBlob used to find and verify
1027 // instruction size in MachNode::emit_size()
1028 init_scratch_buffer_blob(const_req);
1029 if (failing()) return NULL; // Out of memory
1030
1031 // Pre-compute the length of blocks and replace
1032 // long branches with short if machine supports it.
1033 shorten_branches(blk_starts, code_req, locs_req, stub_req);
1034
1035 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1036 // class HandlerImpl is platform-specific and defined in the *.ad files.
1037 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1038 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1039 stub_req += MAX_stubs_size; // ensure per-stub margin
1040 code_req += MAX_inst_size; // ensure per-instruction margin
1041
1042 if (StressCodeBuffers)
1043 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1044
1045 int total_req =
1046 const_req +
1047 code_req +
1048 pad_req +
1049 stub_req +
1050 exception_handler_req +
1051 deopt_handler_req; // deopt handler
1052
1053 if (has_method_handle_invokes())
1054 total_req += deopt_handler_req; // deopt MH handler
1055
1056 CodeBuffer* cb = code_buffer();
1057 cb->initialize(total_req, locs_req);
1058
1059 // Have we run out of code space?
1060 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1061 C->record_failure("CodeCache is full");
1062 return NULL;
1063 }
1064 // Configure the code buffer.
1065 cb->initialize_consts_size(const_req);
1066 cb->initialize_stubs_size(stub_req);
1067 cb->initialize_oop_recorder(env()->oop_recorder());
1068
1069 // fill in the nop array for bundling computations
1070 MachNode *_nop_list[Bundle::_nop_count];
1071 Bundle::initialize_nops(_nop_list);
1072
1073 return cb;
1074 }
1075
1076 //------------------------------fill_buffer------------------------------------
1077 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1078 // blk_starts[] contains offsets calculated during short branches processing,
1079 // offsets should not be increased during following steps.
1080
1081 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1082 // of a loop. It is used to determine the padding for loop alignment.
1083 compute_loop_first_inst_sizes();
1084
1085 // Create oopmap set.
1086 _oop_map_set = new OopMapSet();
1087
1088 // !!!!! This preserves old handling of oopmaps for now
1089 debug_info()->set_oopmaps(_oop_map_set);
1090
1091 uint nblocks = _cfg->number_of_blocks();
1092 // Count and start of implicit null check instructions
1093 uint inct_cnt = 0;
1094 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1095
1096 // Count and start of calls
1097 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1098
1099 uint return_offset = 0;
1100 int nop_size = (new MachNopNode())->size(_regalloc);
1101
1102 int previous_offset = 0;
1103 int current_offset = 0;
1104 int last_call_offset = -1;
1105 int last_avoid_back_to_back_offset = -1;
1106 #ifdef ASSERT
1107 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1108 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1109 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1110 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1111 #endif
1112
1113 // Create an array of unused labels, one for each basic block, if printing is enabled
1114 #if defined(SUPPORT_OPTO_ASSEMBLY)
1115 int *node_offsets = NULL;
1116 uint node_offset_limit = unique();
1117
1118 if (print_assembly()) {
1119 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1120 }
1121 if (node_offsets != NULL) {
1122 // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly.
1123 memset(node_offsets, 0, node_offset_limit*sizeof(int));
1124 }
1125 #endif
1126
1127 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily
1128
1129 // Emit the constant table.
1130 if (has_mach_constant_base_node()) {
1131 constant_table().emit(*cb);
1132 }
1133
1134 // Create an array of labels, one for each basic block
1135 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1136 for (uint i=0; i <= nblocks; i++) {
1137 blk_labels[i].init();
1138 }
1139
1140 // ------------------
1141 // Now fill in the code buffer
1142 Node *delay_slot = NULL;
1143
1144 for (uint i = 0; i < nblocks; i++) {
1145 Block* block = _cfg->get_block(i);
1146 Node* head = block->head();
1147
1148 // If this block needs to start aligned (i.e, can be reached other
1149 // than by falling-thru from the previous block), then force the
1150 // start of a new bundle.
1151 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1152 cb->flush_bundle(true);
1153 }
1154
1155 #ifdef ASSERT
1156 if (!block->is_connector()) {
1157 stringStream st;
1158 block->dump_head(_cfg, &st);
1159 MacroAssembler(cb).block_comment(st.as_string());
1160 }
1161 jmp_target[i] = 0;
1162 jmp_offset[i] = 0;
1163 jmp_size[i] = 0;
1164 jmp_rule[i] = 0;
1165 #endif
1166 int blk_offset = current_offset;
1167
1168 // Define the label at the beginning of the basic block
1169 MacroAssembler(cb).bind(blk_labels[block->_pre_order]);
1170
1171 uint last_inst = block->number_of_nodes();
1172
1173 // Emit block normally, except for last instruction.
1174 // Emit means "dump code bits into code buffer".
1175 for (uint j = 0; j<last_inst; j++) {
1176
1177 // Get the node
1178 Node* n = block->get_node(j);
1179
1180 // See if delay slots are supported
1181 if (valid_bundle_info(n) &&
1182 node_bundling(n)->used_in_unconditional_delay()) {
1183 assert(delay_slot == NULL, "no use of delay slot node");
1184 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1185
1186 delay_slot = n;
1187 continue;
1188 }
1189
1190 // If this starts a new instruction group, then flush the current one
1191 // (but allow split bundles)
1192 if (Pipeline::requires_bundling() && starts_bundle(n))
1193 cb->flush_bundle(false);
1194
1195 // Special handling for SafePoint/Call Nodes
1196 bool is_mcall = false;
1197 if (n->is_Mach()) {
1198 MachNode *mach = n->as_Mach();
1199 is_mcall = n->is_MachCall();
1200 bool is_sfn = n->is_MachSafePoint();
1201
1202 // If this requires all previous instructions be flushed, then do so
1203 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1204 cb->flush_bundle(true);
1205 current_offset = cb->insts_size();
1206 }
1207
1208 // A padding may be needed again since a previous instruction
1209 // could be moved to delay slot.
1210
1211 // align the instruction if necessary
1212 int padding = mach->compute_padding(current_offset);
1213 // Make sure safepoint node for polling is distinct from a call's
1214 // return by adding a nop if needed.
1215 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1216 padding = nop_size;
1217 }
1218 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1219 current_offset == last_avoid_back_to_back_offset) {
1220 // Avoid back to back some instructions.
1221 padding = nop_size;
1222 }
1223
1224 if (padding > 0) {
1225 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1226 int nops_cnt = padding / nop_size;
1227 MachNode *nop = new MachNopNode(nops_cnt);
1228 block->insert_node(nop, j++);
1229 last_inst++;
1230 _cfg->map_node_to_block(nop, block);
1231 // Ensure enough space.
1232 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1233 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1234 C->record_failure("CodeCache is full");
1235 return;
1236 }
1237 nop->emit(*cb, _regalloc);
1238 cb->flush_bundle(true);
1239 current_offset = cb->insts_size();
1240 }
1241
1242 // Remember the start of the last call in a basic block
1243 if (is_mcall) {
1244 MachCallNode *mcall = mach->as_MachCall();
1245
1246 // This destination address is NOT PC-relative
1247 mcall->method_set((intptr_t)mcall->entry_point());
1248
1249 // Save the return address
1250 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1251
1252 if (mcall->is_MachCallLeaf()) {
1253 is_mcall = false;
1254 is_sfn = false;
1255 }
1256 }
1257
1258 // sfn will be valid whenever mcall is valid now because of inheritance
1259 if (is_sfn || is_mcall) {
1260
1261 // Handle special safepoint nodes for synchronization
1262 if (!is_mcall) {
1263 MachSafePointNode *sfn = mach->as_MachSafePoint();
1264 // !!!!! Stubs only need an oopmap right now, so bail out
1265 if (sfn->jvms()->method() == NULL) {
1266 // Write the oopmap directly to the code blob??!!
1267 continue;
1268 }
1269 } // End synchronization
1270
1271 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1272 current_offset);
1273 Process_OopMap_Node(mach, current_offset);
1274 } // End if safepoint
1275
1276 // If this is a null check, then add the start of the previous instruction to the list
1277 else if( mach->is_MachNullCheck() ) {
1278 inct_starts[inct_cnt++] = previous_offset;
1279 }
1280
1281 // If this is a branch, then fill in the label with the target BB's label
1282 else if (mach->is_MachBranch()) {
1283 // This requires the TRUE branch target be in succs[0]
1284 uint block_num = block->non_connector_successor(0)->_pre_order;
1285
1286 // Try to replace long branch if delay slot is not used,
1287 // it is mostly for back branches since forward branch's
1288 // distance is not updated yet.
1289 bool delay_slot_is_used = valid_bundle_info(n) &&
1290 node_bundling(n)->use_unconditional_delay();
1291 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1292 assert(delay_slot == NULL, "not expecting delay slot node");
1293 int br_size = n->size(_regalloc);
1294 int offset = blk_starts[block_num] - current_offset;
1295 if (block_num >= i) {
1296 // Current and following block's offset are not
1297 // finalized yet, adjust distance by the difference
1298 // between calculated and final offsets of current block.
1299 offset -= (blk_starts[i] - blk_offset);
1300 }
1301 // In the following code a nop could be inserted before
1302 // the branch which will increase the backward distance.
1303 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1304 if (needs_padding && offset <= 0)
1305 offset -= nop_size;
1306
1307 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
1308 // We've got a winner. Replace this branch.
1309 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1310
1311 // Update the jmp_size.
1312 int new_size = replacement->size(_regalloc);
1313 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1314 // Insert padding between avoid_back_to_back branches.
1315 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1316 MachNode *nop = new MachNopNode();
1317 block->insert_node(nop, j++);
1318 _cfg->map_node_to_block(nop, block);
1319 last_inst++;
1320 nop->emit(*cb, _regalloc);
1321 cb->flush_bundle(true);
1322 current_offset = cb->insts_size();
1323 }
1324 #ifdef ASSERT
1325 jmp_target[i] = block_num;
1326 jmp_offset[i] = current_offset - blk_offset;
1327 jmp_size[i] = new_size;
1328 jmp_rule[i] = mach->rule();
1329 #endif
1330 block->map_node(replacement, j);
1331 mach->subsume_by(replacement, C);
1332 n = replacement;
1333 mach = replacement;
1334 }
1335 }
1336 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1337 } else if (mach->ideal_Opcode() == Op_Jump) {
1338 for (uint h = 0; h < block->_num_succs; h++) {
1339 Block* succs_block = block->_succs[h];
1340 for (uint j = 1; j < succs_block->num_preds(); j++) {
1341 Node* jpn = succs_block->pred(j);
1342 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1343 uint block_num = succs_block->non_connector()->_pre_order;
1344 Label *blkLabel = &blk_labels[block_num];
1345 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1346 }
1347 }
1348 }
1349 }
1350 #ifdef ASSERT
1351 // Check that oop-store precedes the card-mark
1352 else if (mach->ideal_Opcode() == Op_StoreCM) {
1353 uint storeCM_idx = j;
1354 int count = 0;
1355 for (uint prec = mach->req(); prec < mach->len(); prec++) {
1356 Node *oop_store = mach->in(prec); // Precedence edge
1357 if (oop_store == NULL) continue;
1358 count++;
1359 uint i4;
1360 for (i4 = 0; i4 < last_inst; ++i4) {
1361 if (block->get_node(i4) == oop_store) {
1362 break;
1363 }
1364 }
1365 // Note: This test can provide a false failure if other precedence
1366 // edges have been added to the storeCMNode.
1367 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1368 }
1369 assert(count > 0, "storeCM expects at least one precedence edge");
1370 }
1371 #endif
1372 else if (!n->is_Proj()) {
1373 // Remember the beginning of the previous instruction, in case
1374 // it's followed by a flag-kill and a null-check. Happens on
1375 // Intel all the time, with add-to-memory kind of opcodes.
1376 previous_offset = current_offset;
1377 }
1378
1379 // Not an else-if!
1380 // If this is a trap based cmp then add its offset to the list.
1381 if (mach->is_TrapBasedCheckNode()) {
1382 inct_starts[inct_cnt++] = current_offset;
1383 }
1384 }
1385
1386 // Verify that there is sufficient space remaining
1387 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1388 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1389 C->record_failure("CodeCache is full");
1390 return;
1391 }
1392
1393 // Save the offset for the listing
1394 #if defined(SUPPORT_OPTO_ASSEMBLY)
1395 if ((node_offsets != NULL) && (n->_idx < node_offset_limit)) {
1396 node_offsets[n->_idx] = cb->insts_size();
1397 }
1398 #endif
1399
1400 // "Normal" instruction case
1401 DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
1402 n->emit(*cb, _regalloc);
1403 current_offset = cb->insts_size();
1404
1405 // Above we only verified that there is enough space in the instruction section.
1406 // However, the instruction may emit stubs that cause code buffer expansion.
1407 // Bail out here if expansion failed due to a lack of code cache space.
1408 if (failing()) {
1409 return;
1410 }
1411
1412 #ifdef ASSERT
1413 if (n->size(_regalloc) < (current_offset-instr_offset)) {
1414 n->dump();
1415 assert(false, "wrong size of mach node");
1416 }
1417 #endif
1418 non_safepoints.observe_instruction(n, current_offset);
1419
1420 // mcall is last "call" that can be a safepoint
1421 // record it so we can see if a poll will directly follow it
1422 // in which case we'll need a pad to make the PcDesc sites unique
1423 // see 5010568. This can be slightly inaccurate but conservative
1424 // in the case that return address is not actually at current_offset.
1425 // This is a small price to pay.
1426
1427 if (is_mcall) {
1428 last_call_offset = current_offset;
1429 }
1430
1431 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1432 // Avoid back to back some instructions.
1433 last_avoid_back_to_back_offset = current_offset;
1434 }
1435
1436 // See if this instruction has a delay slot
1437 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1438 guarantee(delay_slot != NULL, "expecting delay slot node");
1439
1440 // Back up 1 instruction
1441 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1442
1443 // Save the offset for the listing
1444 #if defined(SUPPORT_OPTO_ASSEMBLY)
1445 if ((node_offsets != NULL) && (delay_slot->_idx < node_offset_limit)) {
1446 node_offsets[delay_slot->_idx] = cb->insts_size();
1447 }
1448 #endif
1449
1450 // Support a SafePoint in the delay slot
1451 if (delay_slot->is_MachSafePoint()) {
1452 MachNode *mach = delay_slot->as_Mach();
1453 // !!!!! Stubs only need an oopmap right now, so bail out
1454 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
1455 // Write the oopmap directly to the code blob??!!
1456 delay_slot = NULL;
1457 continue;
1458 }
1459
1460 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1461 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1462 adjusted_offset);
1463 // Generate an OopMap entry
1464 Process_OopMap_Node(mach, adjusted_offset);
1465 }
1466
1467 // Insert the delay slot instruction
1468 delay_slot->emit(*cb, _regalloc);
1469
1470 // Don't reuse it
1471 delay_slot = NULL;
1472 }
1473
1474 } // End for all instructions in block
1475
1476 // If the next block is the top of a loop, pad this block out to align
1477 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1478 if (i < nblocks-1) {
1479 Block *nb = _cfg->get_block(i + 1);
1480 int padding = nb->alignment_padding(current_offset);
1481 if( padding > 0 ) {
1482 MachNode *nop = new MachNopNode(padding / nop_size);
1483 block->insert_node(nop, block->number_of_nodes());
1484 _cfg->map_node_to_block(nop, block);
1485 nop->emit(*cb, _regalloc);
1486 current_offset = cb->insts_size();
1487 }
1488 }
1489 // Verify that the distance for generated before forward
1490 // short branches is still valid.
1491 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1492
1493 // Save new block start offset
1494 blk_starts[i] = blk_offset;
1495 } // End of for all blocks
1496 blk_starts[nblocks] = current_offset;
1497
1498 non_safepoints.flush_at_end();
1499
1500 // Offset too large?
1501 if (failing()) return;
1502
1503 // Define a pseudo-label at the end of the code
1504 MacroAssembler(cb).bind( blk_labels[nblocks] );
1505
1506 // Compute the size of the first block
1507 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1508
1509 #ifdef ASSERT
1510 for (uint i = 0; i < nblocks; i++) { // For all blocks
1511 if (jmp_target[i] != 0) {
1512 int br_size = jmp_size[i];
1513 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1514 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1515 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]);
1516 assert(false, "Displacement too large for short jmp");
1517 }
1518 }
1519 }
1520 #endif
1521
1522 #ifndef PRODUCT
1523 // Information on the size of the method, without the extraneous code
1524 Scheduling::increment_method_size(cb->insts_size());
1525 #endif
1526
1527 // ------------------
1528 // Fill in exception table entries.
1529 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1530
1531 // Only java methods have exception handlers and deopt handlers
1532 // class HandlerImpl is platform-specific and defined in the *.ad files.
1533 if (_method) {
1534 // Emit the exception handler code.
1535 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb));
1536 if (failing()) {
1537 return; // CodeBuffer::expand failed
1538 }
1539 // Emit the deopt handler code.
1540 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb));
1541
1542 // Emit the MethodHandle deopt handler code (if required).
1543 if (has_method_handle_invokes() && !failing()) {
1544 // We can use the same code as for the normal deopt handler, we
1545 // just need a different entry point address.
1546 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb));
1547 }
1548 }
1549
1550 // One last check for failed CodeBuffer::expand:
1551 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1552 C->record_failure("CodeCache is full");
1553 return;
1554 }
1555
1556 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY)
1557 if (print_assembly()) {
1558 tty->cr();
1559 tty->print_cr("============================= C2-compiled nmethod ==============================");
1560 }
1561 #endif
1562
1563 #if defined(SUPPORT_OPTO_ASSEMBLY)
1564 // Dump the assembly code, including basic-block numbers
1565 if (print_assembly()) {
1566 ttyLocker ttyl; // keep the following output all in one block
1567 if (!VMThread::should_terminate()) { // test this under the tty lock
1568 // This output goes directly to the tty, not the compiler log.
1569 // To enable tools to match it up with the compilation activity,
1570 // be sure to tag this tty output with the compile ID.
1571 if (xtty != NULL) {
1572 xtty->head("opto_assembly compile_id='%d'%s", compile_id(),
1573 is_osr_compilation() ? " compile_kind='osr'" :
1574 "");
1575 }
1576 if (method() != NULL) {
1577 tty->print_cr("----------------------------------- MetaData -----------------------------------");
1578 method()->print_metadata();
1579 } else if (stub_name() != NULL) {
1580 tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", stub_name());
1581 }
1582 tty->cr();
1583 tty->print_cr("--------------------------------- OptoAssembly ---------------------------------");
1584 dump_asm(node_offsets, node_offset_limit);
1585 tty->print_cr("--------------------------------------------------------------------------------");
1586 if (xtty != NULL) {
1587 // print_metadata and dump_asm above may safepoint which makes us loose the ttylock.
1588 // Retake lock too make sure the end tag is coherent, and that xmlStream->pop_tag is done
1589 // thread safe
1590 ttyLocker ttyl2;
1591 xtty->tail("opto_assembly");
1592 }
1593 }
1594 }
1595 #endif
1596 }
1597
1598 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1599 _inc_table.set_size(cnt);
1600
1601 uint inct_cnt = 0;
1602 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1603 Block* block = _cfg->get_block(i);
1604 Node *n = NULL;
1605 int j;
1606
1607 // Find the branch; ignore trailing NOPs.
1608 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1609 n = block->get_node(j);
1610 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1611 break;
1612 }
1613 }
1614
1615 // If we didn't find anything, continue
1616 if (j < 0) {
1617 continue;
1618 }
1619
1620 // Compute ExceptionHandlerTable subtable entry and add it
1621 // (skip empty blocks)
1622 if (n->is_Catch()) {
1623
1624 // Get the offset of the return from the call
1625 uint call_return = call_returns[block->_pre_order];
1626 #ifdef ASSERT
1627 assert( call_return > 0, "no call seen for this basic block" );
1628 while (block->get_node(--j)->is_MachProj()) ;
1629 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1630 #endif
1631 // last instruction is a CatchNode, find it's CatchProjNodes
1632 int nof_succs = block->_num_succs;
1633 // allocate space
1634 GrowableArray<intptr_t> handler_bcis(nof_succs);
1635 GrowableArray<intptr_t> handler_pcos(nof_succs);
1636 // iterate through all successors
1637 for (int j = 0; j < nof_succs; j++) {
1638 Block* s = block->_succs[j];
1639 bool found_p = false;
1640 for (uint k = 1; k < s->num_preds(); k++) {
1641 Node* pk = s->pred(k);
1642 if (pk->is_CatchProj() && pk->in(0) == n) {
1643 const CatchProjNode* p = pk->as_CatchProj();
1644 found_p = true;
1645 // add the corresponding handler bci & pco information
1646 if (p->_con != CatchProjNode::fall_through_index) {
1647 // p leads to an exception handler (and is not fall through)
1648 assert(s == _cfg->get_block(s->_pre_order), "bad numbering");
1649 // no duplicates, please
1650 if (!handler_bcis.contains(p->handler_bci())) {
1651 uint block_num = s->non_connector()->_pre_order;
1652 handler_bcis.append(p->handler_bci());
1653 handler_pcos.append(blk_labels[block_num].loc_pos());
1654 }
1655 }
1656 }
1657 }
1658 assert(found_p, "no matching predecessor found");
1659 // Note: Due to empty block removal, one block may have
1660 // several CatchProj inputs, from the same Catch.
1661 }
1662
1663 // Set the offset of the return from the call
1664 assert(handler_bcis.find(-1) != -1, "must have default handler");
1665 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
1666 continue;
1667 }
1668
1669 // Handle implicit null exception table updates
1670 if (n->is_MachNullCheck()) {
1671 uint block_num = block->non_connector_successor(0)->_pre_order;
1672 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1673 continue;
1674 }
1675 // Handle implicit exception table updates: trap instructions.
1676 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
1677 uint block_num = block->non_connector_successor(0)->_pre_order;
1678 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1679 continue;
1680 }
1681 } // End of for all blocks fill in exception table entries
1682 }
1683
1684 // Static Variables
1685 #ifndef PRODUCT
1686 uint Scheduling::_total_nop_size = 0;
1687 uint Scheduling::_total_method_size = 0;
1688 uint Scheduling::_total_branches = 0;
1689 uint Scheduling::_total_unconditional_delays = 0;
1690 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1691 #endif
1692
1693 // Initializer for class Scheduling
1694
1695 Scheduling::Scheduling(Arena *arena, Compile &compile)
1696 : _arena(arena),
1697 _cfg(compile.cfg()),
1698 _regalloc(compile.regalloc()),
1699 _scheduled(arena),
1700 _available(arena),
1701 _reg_node(arena),
1702 _pinch_free_list(arena),
1703 _next_node(NULL),
1704 _bundle_instr_count(0),
1705 _bundle_cycle_number(0),
1706 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
1707 #ifndef PRODUCT
1708 , _branches(0)
1709 , _unconditional_delays(0)
1710 #endif
1711 {
1712 // Create a MachNopNode
1713 _nop = new MachNopNode();
1714
1715 // Now that the nops are in the array, save the count
1716 // (but allow entries for the nops)
1717 _node_bundling_limit = compile.unique();
1718 uint node_max = _regalloc->node_regs_max_index();
1719
1720 compile.set_node_bundling_limit(_node_bundling_limit);
1721
1722 // This one is persistent within the Compile class
1723 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1724
1725 // Allocate space for fixed-size arrays
1726 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1727 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
1728 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1729
1730 // Clear the arrays
1731 for (uint i = 0; i < node_max; i++) {
1732 ::new (&_node_bundling_base[i]) Bundle();
1733 }
1734 memset(_node_latency, 0, node_max * sizeof(unsigned short));
1735 memset(_uses, 0, node_max * sizeof(short));
1736 memset(_current_latency, 0, node_max * sizeof(unsigned short));
1737
1738 // Clear the bundling information
1739 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
1740
1741 // Get the last node
1742 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
1743
1744 _next_node = block->get_node(block->number_of_nodes() - 1);
1745 }
1746
1747 #ifndef PRODUCT
1748 // Scheduling destructor
1749 Scheduling::~Scheduling() {
1750 _total_branches += _branches;
1751 _total_unconditional_delays += _unconditional_delays;
1752 }
1753 #endif
1754
1755 // Step ahead "i" cycles
1756 void Scheduling::step(uint i) {
1757
1758 Bundle *bundle = node_bundling(_next_node);
1759 bundle->set_starts_bundle();
1760
1761 // Update the bundle record, but leave the flags information alone
1762 if (_bundle_instr_count > 0) {
1763 bundle->set_instr_count(_bundle_instr_count);
1764 bundle->set_resources_used(_bundle_use.resourcesUsed());
1765 }
1766
1767 // Update the state information
1768 _bundle_instr_count = 0;
1769 _bundle_cycle_number += i;
1770 _bundle_use.step(i);
1771 }
1772
1773 void Scheduling::step_and_clear() {
1774 Bundle *bundle = node_bundling(_next_node);
1775 bundle->set_starts_bundle();
1776
1777 // Update the bundle record
1778 if (_bundle_instr_count > 0) {
1779 bundle->set_instr_count(_bundle_instr_count);
1780 bundle->set_resources_used(_bundle_use.resourcesUsed());
1781
1782 _bundle_cycle_number += 1;
1783 }
1784
1785 // Clear the bundling information
1786 _bundle_instr_count = 0;
1787 _bundle_use.reset();
1788
1789 memcpy(_bundle_use_elements,
1790 Pipeline_Use::elaborated_elements,
1791 sizeof(Pipeline_Use::elaborated_elements));
1792 }
1793
1794 // Perform instruction scheduling and bundling over the sequence of
1795 // instructions in backwards order.
1796 void Compile::ScheduleAndBundle() {
1797
1798 // Don't optimize this if it isn't a method
1799 if (!_method)
1800 return;
1801
1802 // Don't optimize this if scheduling is disabled
1803 if (!do_scheduling())
1804 return;
1805
1806 // Scheduling code works only with pairs (16 bytes) maximum.
1807 if (max_vector_size() > 16)
1808 return;
1809
1810 TracePhase tp("isched", &timers[_t_instrSched]);
1811
1812 // Create a data structure for all the scheduling information
1813 Scheduling scheduling(Thread::current()->resource_area(), *this);
1814
1815 // Walk backwards over each basic block, computing the needed alignment
1816 // Walk over all the basic blocks
1817 scheduling.DoScheduling();
1818 }
1819
1820 // Compute the latency of all the instructions. This is fairly simple,
1821 // because we already have a legal ordering. Walk over the instructions
1822 // from first to last, and compute the latency of the instruction based
1823 // on the latency of the preceding instruction(s).
1824 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
1825 #ifndef PRODUCT
1826 if (_cfg->C->trace_opto_output())
1827 tty->print("# -> ComputeLocalLatenciesForward\n");
1828 #endif
1829
1830 // Walk over all the schedulable instructions
1831 for( uint j=_bb_start; j < _bb_end; j++ ) {
1832
1833 // This is a kludge, forcing all latency calculations to start at 1.
1834 // Used to allow latency 0 to force an instruction to the beginning
1835 // of the bb
1836 uint latency = 1;
1837 Node *use = bb->get_node(j);
1838 uint nlen = use->len();
1839
1840 // Walk over all the inputs
1841 for ( uint k=0; k < nlen; k++ ) {
1842 Node *def = use->in(k);
1843 if (!def)
1844 continue;
1845
1846 uint l = _node_latency[def->_idx] + use->latency(k);
1847 if (latency < l)
1848 latency = l;
1849 }
1850
1851 _node_latency[use->_idx] = latency;
1852
1853 #ifndef PRODUCT
1854 if (_cfg->C->trace_opto_output()) {
1855 tty->print("# latency %4d: ", latency);
1856 use->dump();
1857 }
1858 #endif
1859 }
1860
1861 #ifndef PRODUCT
1862 if (_cfg->C->trace_opto_output())
1863 tty->print("# <- ComputeLocalLatenciesForward\n");
1864 #endif
1865
1866 } // end ComputeLocalLatenciesForward
1867
1868 // See if this node fits into the present instruction bundle
1869 bool Scheduling::NodeFitsInBundle(Node *n) {
1870 uint n_idx = n->_idx;
1871
1872 // If this is the unconditional delay instruction, then it fits
1873 if (n == _unconditional_delay_slot) {
1874 #ifndef PRODUCT
1875 if (_cfg->C->trace_opto_output())
1876 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
1877 #endif
1878 return (true);
1879 }
1880
1881 // If the node cannot be scheduled this cycle, skip it
1882 if (_current_latency[n_idx] > _bundle_cycle_number) {
1883 #ifndef PRODUCT
1884 if (_cfg->C->trace_opto_output())
1885 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
1886 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
1887 #endif
1888 return (false);
1889 }
1890
1891 const Pipeline *node_pipeline = n->pipeline();
1892
1893 uint instruction_count = node_pipeline->instructionCount();
1894 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
1895 instruction_count = 0;
1896 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
1897 instruction_count++;
1898
1899 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
1900 #ifndef PRODUCT
1901 if (_cfg->C->trace_opto_output())
1902 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
1903 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
1904 #endif
1905 return (false);
1906 }
1907
1908 // Don't allow non-machine nodes to be handled this way
1909 if (!n->is_Mach() && instruction_count == 0)
1910 return (false);
1911
1912 // See if there is any overlap
1913 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
1914
1915 if (delay > 0) {
1916 #ifndef PRODUCT
1917 if (_cfg->C->trace_opto_output())
1918 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
1919 #endif
1920 return false;
1921 }
1922
1923 #ifndef PRODUCT
1924 if (_cfg->C->trace_opto_output())
1925 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
1926 #endif
1927
1928 return true;
1929 }
1930
1931 Node * Scheduling::ChooseNodeToBundle() {
1932 uint siz = _available.size();
1933
1934 if (siz == 0) {
1935
1936 #ifndef PRODUCT
1937 if (_cfg->C->trace_opto_output())
1938 tty->print("# ChooseNodeToBundle: NULL\n");
1939 #endif
1940 return (NULL);
1941 }
1942
1943 // Fast path, if only 1 instruction in the bundle
1944 if (siz == 1) {
1945 #ifndef PRODUCT
1946 if (_cfg->C->trace_opto_output()) {
1947 tty->print("# ChooseNodeToBundle (only 1): ");
1948 _available[0]->dump();
1949 }
1950 #endif
1951 return (_available[0]);
1952 }
1953
1954 // Don't bother, if the bundle is already full
1955 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
1956 for ( uint i = 0; i < siz; i++ ) {
1957 Node *n = _available[i];
1958
1959 // Skip projections, we'll handle them another way
1960 if (n->is_Proj())
1961 continue;
1962
1963 // This presupposed that instructions are inserted into the
1964 // available list in a legality order; i.e. instructions that
1965 // must be inserted first are at the head of the list
1966 if (NodeFitsInBundle(n)) {
1967 #ifndef PRODUCT
1968 if (_cfg->C->trace_opto_output()) {
1969 tty->print("# ChooseNodeToBundle: ");
1970 n->dump();
1971 }
1972 #endif
1973 return (n);
1974 }
1975 }
1976 }
1977
1978 // Nothing fits in this bundle, choose the highest priority
1979 #ifndef PRODUCT
1980 if (_cfg->C->trace_opto_output()) {
1981 tty->print("# ChooseNodeToBundle: ");
1982 _available[0]->dump();
1983 }
1984 #endif
1985
1986 return _available[0];
1987 }
1988
1989 void Scheduling::AddNodeToAvailableList(Node *n) {
1990 assert( !n->is_Proj(), "projections never directly made available" );
1991 #ifndef PRODUCT
1992 if (_cfg->C->trace_opto_output()) {
1993 tty->print("# AddNodeToAvailableList: ");
1994 n->dump();
1995 }
1996 #endif
1997
1998 int latency = _current_latency[n->_idx];
1999
2000 // Insert in latency order (insertion sort)
2001 uint i;
2002 for ( i=0; i < _available.size(); i++ )
2003 if (_current_latency[_available[i]->_idx] > latency)
2004 break;
2005
2006 // Special Check for compares following branches
2007 if( n->is_Mach() && _scheduled.size() > 0 ) {
2008 int op = n->as_Mach()->ideal_Opcode();
2009 Node *last = _scheduled[0];
2010 if( last->is_MachIf() && last->in(1) == n &&
2011 ( op == Op_CmpI ||
2012 op == Op_CmpU ||
2013 op == Op_CmpUL ||
2014 op == Op_CmpP ||
2015 op == Op_CmpF ||
2016 op == Op_CmpD ||
2017 op == Op_CmpL ) ) {
2018
2019 // Recalculate position, moving to front of same latency
2020 for ( i=0 ; i < _available.size(); i++ )
2021 if (_current_latency[_available[i]->_idx] >= latency)
2022 break;
2023 }
2024 }
2025
2026 // Insert the node in the available list
2027 _available.insert(i, n);
2028
2029 #ifndef PRODUCT
2030 if (_cfg->C->trace_opto_output())
2031 dump_available();
2032 #endif
2033 }
2034
2035 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2036 for ( uint i=0; i < n->len(); i++ ) {
2037 Node *def = n->in(i);
2038 if (!def) continue;
2039 if( def->is_Proj() ) // If this is a machine projection, then
2040 def = def->in(0); // propagate usage thru to the base instruction
2041
2042 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2043 continue;
2044 }
2045
2046 // Compute the latency
2047 uint l = _bundle_cycle_number + n->latency(i);
2048 if (_current_latency[def->_idx] < l)
2049 _current_latency[def->_idx] = l;
2050
2051 // If this does not have uses then schedule it
2052 if ((--_uses[def->_idx]) == 0)
2053 AddNodeToAvailableList(def);
2054 }
2055 }
2056
2057 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2058 #ifndef PRODUCT
2059 if (_cfg->C->trace_opto_output()) {
2060 tty->print("# AddNodeToBundle: ");
2061 n->dump();
2062 }
2063 #endif
2064
2065 // Remove this from the available list
2066 uint i;
2067 for (i = 0; i < _available.size(); i++)
2068 if (_available[i] == n)
2069 break;
2070 assert(i < _available.size(), "entry in _available list not found");
2071 _available.remove(i);
2072
2073 // See if this fits in the current bundle
2074 const Pipeline *node_pipeline = n->pipeline();
2075 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2076
2077 // Check for instructions to be placed in the delay slot. We
2078 // do this before we actually schedule the current instruction,
2079 // because the delay slot follows the current instruction.
2080 if (Pipeline::_branch_has_delay_slot &&
2081 node_pipeline->hasBranchDelay() &&
2082 !_unconditional_delay_slot) {
2083
2084 uint siz = _available.size();
2085
2086 // Conditional branches can support an instruction that
2087 // is unconditionally executed and not dependent by the
2088 // branch, OR a conditionally executed instruction if
2089 // the branch is taken. In practice, this means that
2090 // the first instruction at the branch target is
2091 // copied to the delay slot, and the branch goes to
2092 // the instruction after that at the branch target
2093 if ( n->is_MachBranch() ) {
2094
2095 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2096 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2097
2098 #ifndef PRODUCT
2099 _branches++;
2100 #endif
2101
2102 // At least 1 instruction is on the available list
2103 // that is not dependent on the branch
2104 for (uint i = 0; i < siz; i++) {
2105 Node *d = _available[i];
2106 const Pipeline *avail_pipeline = d->pipeline();
2107
2108 // Don't allow safepoints in the branch shadow, that will
2109 // cause a number of difficulties
2110 if ( avail_pipeline->instructionCount() == 1 &&
2111 !avail_pipeline->hasMultipleBundles() &&
2112 !avail_pipeline->hasBranchDelay() &&
2113 Pipeline::instr_has_unit_size() &&
2114 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2115 NodeFitsInBundle(d) &&
2116 !node_bundling(d)->used_in_delay()) {
2117
2118 if (d->is_Mach() && !d->is_MachSafePoint()) {
2119 // A node that fits in the delay slot was found, so we need to
2120 // set the appropriate bits in the bundle pipeline information so
2121 // that it correctly indicates resource usage. Later, when we
2122 // attempt to add this instruction to the bundle, we will skip
2123 // setting the resource usage.
2124 _unconditional_delay_slot = d;
2125 node_bundling(n)->set_use_unconditional_delay();
2126 node_bundling(d)->set_used_in_unconditional_delay();
2127 _bundle_use.add_usage(avail_pipeline->resourceUse());
2128 _current_latency[d->_idx] = _bundle_cycle_number;
2129 _next_node = d;
2130 ++_bundle_instr_count;
2131 #ifndef PRODUCT
2132 _unconditional_delays++;
2133 #endif
2134 break;
2135 }
2136 }
2137 }
2138 }
2139
2140 // No delay slot, add a nop to the usage
2141 if (!_unconditional_delay_slot) {
2142 // See if adding an instruction in the delay slot will overflow
2143 // the bundle.
2144 if (!NodeFitsInBundle(_nop)) {
2145 #ifndef PRODUCT
2146 if (_cfg->C->trace_opto_output())
2147 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2148 #endif
2149 step(1);
2150 }
2151
2152 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2153 _next_node = _nop;
2154 ++_bundle_instr_count;
2155 }
2156
2157 // See if the instruction in the delay slot requires a
2158 // step of the bundles
2159 if (!NodeFitsInBundle(n)) {
2160 #ifndef PRODUCT
2161 if (_cfg->C->trace_opto_output())
2162 tty->print("# *** STEP(branch won't fit) ***\n");
2163 #endif
2164 // Update the state information
2165 _bundle_instr_count = 0;
2166 _bundle_cycle_number += 1;
2167 _bundle_use.step(1);
2168 }
2169 }
2170
2171 // Get the number of instructions
2172 uint instruction_count = node_pipeline->instructionCount();
2173 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2174 instruction_count = 0;
2175
2176 // Compute the latency information
2177 uint delay = 0;
2178
2179 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2180 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2181 if (relative_latency < 0)
2182 relative_latency = 0;
2183
2184 delay = _bundle_use.full_latency(relative_latency, node_usage);
2185
2186 // Does not fit in this bundle, start a new one
2187 if (delay > 0) {
2188 step(delay);
2189
2190 #ifndef PRODUCT
2191 if (_cfg->C->trace_opto_output())
2192 tty->print("# *** STEP(%d) ***\n", delay);
2193 #endif
2194 }
2195 }
2196
2197 // If this was placed in the delay slot, ignore it
2198 if (n != _unconditional_delay_slot) {
2199
2200 if (delay == 0) {
2201 if (node_pipeline->hasMultipleBundles()) {
2202 #ifndef PRODUCT
2203 if (_cfg->C->trace_opto_output())
2204 tty->print("# *** STEP(multiple instructions) ***\n");
2205 #endif
2206 step(1);
2207 }
2208
2209 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2210 #ifndef PRODUCT
2211 if (_cfg->C->trace_opto_output())
2212 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2213 instruction_count + _bundle_instr_count,
2214 Pipeline::_max_instrs_per_cycle);
2215 #endif
2216 step(1);
2217 }
2218 }
2219
2220 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2221 _bundle_instr_count++;
2222
2223 // Set the node's latency
2224 _current_latency[n->_idx] = _bundle_cycle_number;
2225
2226 // Now merge the functional unit information
2227 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2228 _bundle_use.add_usage(node_usage);
2229
2230 // Increment the number of instructions in this bundle
2231 _bundle_instr_count += instruction_count;
2232
2233 // Remember this node for later
2234 if (n->is_Mach())
2235 _next_node = n;
2236 }
2237
2238 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2239 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2240 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2241 // into the block. All other scheduled nodes get put in the schedule here.
2242 int op = n->Opcode();
2243 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2244 (op != Op_Node && // Not an unused antidepedence node and
2245 // not an unallocated boxlock
2246 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2247
2248 // Push any trailing projections
2249 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2250 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2251 Node *foi = n->fast_out(i);
2252 if( foi->is_Proj() )
2253 _scheduled.push(foi);
2254 }
2255 }
2256
2257 // Put the instruction in the schedule list
2258 _scheduled.push(n);
2259 }
2260
2261 #ifndef PRODUCT
2262 if (_cfg->C->trace_opto_output())
2263 dump_available();
2264 #endif
2265
2266 // Walk all the definitions, decrementing use counts, and
2267 // if a definition has a 0 use count, place it in the available list.
2268 DecrementUseCounts(n,bb);
2269 }
2270
2271 // This method sets the use count within a basic block. We will ignore all
2272 // uses outside the current basic block. As we are doing a backwards walk,
2273 // any node we reach that has a use count of 0 may be scheduled. This also
2274 // avoids the problem of cyclic references from phi nodes, as long as phi
2275 // nodes are at the front of the basic block. This method also initializes
2276 // the available list to the set of instructions that have no uses within this
2277 // basic block.
2278 void Scheduling::ComputeUseCount(const Block *bb) {
2279 #ifndef PRODUCT
2280 if (_cfg->C->trace_opto_output())
2281 tty->print("# -> ComputeUseCount\n");
2282 #endif
2283
2284 // Clear the list of available and scheduled instructions, just in case
2285 _available.clear();
2286 _scheduled.clear();
2287
2288 // No delay slot specified
2289 _unconditional_delay_slot = NULL;
2290
2291 #ifdef ASSERT
2292 for( uint i=0; i < bb->number_of_nodes(); i++ )
2293 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2294 #endif
2295
2296 // Force the _uses count to never go to zero for unscheduable pieces
2297 // of the block
2298 for( uint k = 0; k < _bb_start; k++ )
2299 _uses[bb->get_node(k)->_idx] = 1;
2300 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2301 _uses[bb->get_node(l)->_idx] = 1;
2302
2303 // Iterate backwards over the instructions in the block. Don't count the
2304 // branch projections at end or the block header instructions.
2305 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2306 Node *n = bb->get_node(j);
2307 if( n->is_Proj() ) continue; // Projections handled another way
2308
2309 // Account for all uses
2310 for ( uint k = 0; k < n->len(); k++ ) {
2311 Node *inp = n->in(k);
2312 if (!inp) continue;
2313 assert(inp != n, "no cycles allowed" );
2314 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2315 if (inp->is_Proj()) { // Skip through Proj's
2316 inp = inp->in(0);
2317 }
2318 ++_uses[inp->_idx]; // Count 1 block-local use
2319 }
2320 }
2321
2322 // If this instruction has a 0 use count, then it is available
2323 if (!_uses[n->_idx]) {
2324 _current_latency[n->_idx] = _bundle_cycle_number;
2325 AddNodeToAvailableList(n);
2326 }
2327
2328 #ifndef PRODUCT
2329 if (_cfg->C->trace_opto_output()) {
2330 tty->print("# uses: %3d: ", _uses[n->_idx]);
2331 n->dump();
2332 }
2333 #endif
2334 }
2335
2336 #ifndef PRODUCT
2337 if (_cfg->C->trace_opto_output())
2338 tty->print("# <- ComputeUseCount\n");
2339 #endif
2340 }
2341
2342 // This routine performs scheduling on each basic block in reverse order,
2343 // using instruction latencies and taking into account function unit
2344 // availability.
2345 void Scheduling::DoScheduling() {
2346 #ifndef PRODUCT
2347 if (_cfg->C->trace_opto_output())
2348 tty->print("# -> DoScheduling\n");
2349 #endif
2350
2351 Block *succ_bb = NULL;
2352 Block *bb;
2353
2354 // Walk over all the basic blocks in reverse order
2355 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2356 bb = _cfg->get_block(i);
2357
2358 #ifndef PRODUCT
2359 if (_cfg->C->trace_opto_output()) {
2360 tty->print("# Schedule BB#%03d (initial)\n", i);
2361 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2362 bb->get_node(j)->dump();
2363 }
2364 }
2365 #endif
2366
2367 // On the head node, skip processing
2368 if (bb == _cfg->get_root_block()) {
2369 continue;
2370 }
2371
2372 // Skip empty, connector blocks
2373 if (bb->is_connector())
2374 continue;
2375
2376 // If the following block is not the sole successor of
2377 // this one, then reset the pipeline information
2378 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2379 #ifndef PRODUCT
2380 if (_cfg->C->trace_opto_output()) {
2381 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2382 _next_node->_idx, _bundle_instr_count);
2383 }
2384 #endif
2385 step_and_clear();
2386 }
2387
2388 // Leave untouched the starting instruction, any Phis, a CreateEx node
2389 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2390 _bb_end = bb->number_of_nodes()-1;
2391 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2392 Node *n = bb->get_node(_bb_start);
2393 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2394 // Also, MachIdealNodes do not get scheduled
2395 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2396 MachNode *mach = n->as_Mach();
2397 int iop = mach->ideal_Opcode();
2398 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2399 if( iop == Op_Con ) continue; // Do not schedule Top
2400 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2401 mach->pipeline() == MachNode::pipeline_class() &&
2402 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
2403 continue;
2404 break; // Funny loop structure to be sure...
2405 }
2406 // Compute last "interesting" instruction in block - last instruction we
2407 // might schedule. _bb_end points just after last schedulable inst. We
2408 // normally schedule conditional branches (despite them being forced last
2409 // in the block), because they have delay slots we can fill. Calls all
2410 // have their delay slots filled in the template expansions, so we don't
2411 // bother scheduling them.
2412 Node *last = bb->get_node(_bb_end);
2413 // Ignore trailing NOPs.
2414 while (_bb_end > 0 && last->is_Mach() &&
2415 last->as_Mach()->ideal_Opcode() == Op_Con) {
2416 last = bb->get_node(--_bb_end);
2417 }
2418 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2419 if( last->is_Catch() ||
2420 (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2421 // There might be a prior call. Skip it.
2422 while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj());
2423 } else if( last->is_MachNullCheck() ) {
2424 // Backup so the last null-checked memory instruction is
2425 // outside the schedulable range. Skip over the nullcheck,
2426 // projection, and the memory nodes.
2427 Node *mem = last->in(1);
2428 do {
2429 _bb_end--;
2430 } while (mem != bb->get_node(_bb_end));
2431 } else {
2432 // Set _bb_end to point after last schedulable inst.
2433 _bb_end++;
2434 }
2435
2436 assert( _bb_start <= _bb_end, "inverted block ends" );
2437
2438 // Compute the register antidependencies for the basic block
2439 ComputeRegisterAntidependencies(bb);
2440 if (_cfg->C->failing()) return; // too many D-U pinch points
2441
2442 // Compute intra-bb latencies for the nodes
2443 ComputeLocalLatenciesForward(bb);
2444
2445 // Compute the usage within the block, and set the list of all nodes
2446 // in the block that have no uses within the block.
2447 ComputeUseCount(bb);
2448
2449 // Schedule the remaining instructions in the block
2450 while ( _available.size() > 0 ) {
2451 Node *n = ChooseNodeToBundle();
2452 guarantee(n != NULL, "no nodes available");
2453 AddNodeToBundle(n,bb);
2454 }
2455
2456 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2457 #ifdef ASSERT
2458 for( uint l = _bb_start; l < _bb_end; l++ ) {
2459 Node *n = bb->get_node(l);
2460 uint m;
2461 for( m = 0; m < _bb_end-_bb_start; m++ )
2462 if( _scheduled[m] == n )
2463 break;
2464 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2465 }
2466 #endif
2467
2468 // Now copy the instructions (in reverse order) back to the block
2469 for ( uint k = _bb_start; k < _bb_end; k++ )
2470 bb->map_node(_scheduled[_bb_end-k-1], k);
2471
2472 #ifndef PRODUCT
2473 if (_cfg->C->trace_opto_output()) {
2474 tty->print("# Schedule BB#%03d (final)\n", i);
2475 uint current = 0;
2476 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2477 Node *n = bb->get_node(j);
2478 if( valid_bundle_info(n) ) {
2479 Bundle *bundle = node_bundling(n);
2480 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2481 tty->print("*** Bundle: ");
2482 bundle->dump();
2483 }
2484 n->dump();
2485 }
2486 }
2487 }
2488 #endif
2489 #ifdef ASSERT
2490 verify_good_schedule(bb,"after block local scheduling");
2491 #endif
2492 }
2493
2494 #ifndef PRODUCT
2495 if (_cfg->C->trace_opto_output())
2496 tty->print("# <- DoScheduling\n");
2497 #endif
2498
2499 // Record final node-bundling array location
2500 _regalloc->C->set_node_bundling_base(_node_bundling_base);
2501
2502 } // end DoScheduling
2503
2504 // Verify that no live-range used in the block is killed in the block by a
2505 // wrong DEF. This doesn't verify live-ranges that span blocks.
2506
2507 // Check for edge existence. Used to avoid adding redundant precedence edges.
2508 static bool edge_from_to( Node *from, Node *to ) {
2509 for( uint i=0; i<from->len(); i++ )
2510 if( from->in(i) == to )
2511 return true;
2512 return false;
2513 }
2514
2515 #ifdef ASSERT
2516 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2517 // Check for bad kills
2518 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2519 Node *prior_use = _reg_node[def];
2520 if( prior_use && !edge_from_to(prior_use,n) ) {
2521 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2522 n->dump();
2523 tty->print_cr("...");
2524 prior_use->dump();
2525 assert(edge_from_to(prior_use,n), "%s", msg);
2526 }
2527 _reg_node.map(def,NULL); // Kill live USEs
2528 }
2529 }
2530
2531 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2532
2533 // Zap to something reasonable for the verify code
2534 _reg_node.clear();
2535
2536 // Walk over the block backwards. Check to make sure each DEF doesn't
2537 // kill a live value (other than the one it's supposed to). Add each
2538 // USE to the live set.
2539 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2540 Node *n = b->get_node(i);
2541 int n_op = n->Opcode();
2542 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2543 // Fat-proj kills a slew of registers
2544 RegMask rm = n->out_RegMask();// Make local copy
2545 while( rm.is_NotEmpty() ) {
2546 OptoReg::Name kill = rm.find_first_elem();
2547 rm.Remove(kill);
2548 verify_do_def( n, kill, msg );
2549 }
2550 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2551 // Get DEF'd registers the normal way
2552 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2553 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2554 }
2555
2556 // Now make all USEs live
2557 for( uint i=1; i<n->req(); i++ ) {
2558 Node *def = n->in(i);
2559 assert(def != 0, "input edge required");
2560 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2561 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2562 if( OptoReg::is_valid(reg_lo) ) {
2563 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
2564 _reg_node.map(reg_lo,n);
2565 }
2566 if( OptoReg::is_valid(reg_hi) ) {
2567 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
2568 _reg_node.map(reg_hi,n);
2569 }
2570 }
2571
2572 }
2573
2574 // Zap to something reasonable for the Antidependence code
2575 _reg_node.clear();
2576 }
2577 #endif
2578
2579 // Conditionally add precedence edges. Avoid putting edges on Projs.
2580 static void add_prec_edge_from_to( Node *from, Node *to ) {
2581 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2582 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2583 from = from->in(0);
2584 }
2585 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2586 !edge_from_to( from, to ) ) // Avoid duplicate edge
2587 from->add_prec(to);
2588 }
2589
2590 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2591 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2592 return;
2593
2594 Node *pinch = _reg_node[def_reg]; // Get pinch point
2595 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2596 is_def ) { // Check for a true def (not a kill)
2597 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2598 return;
2599 }
2600
2601 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2602 debug_only( def = (Node*)((intptr_t)0xdeadbeef); )
2603
2604 // After some number of kills there _may_ be a later def
2605 Node *later_def = NULL;
2606
2607 // Finding a kill requires a real pinch-point.
2608 // Check for not already having a pinch-point.
2609 // Pinch points are Op_Node's.
2610 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2611 later_def = pinch; // Must be def/kill as optimistic pinch-point
2612 if ( _pinch_free_list.size() > 0) {
2613 pinch = _pinch_free_list.pop();
2614 } else {
2615 pinch = new Node(1); // Pinch point to-be
2616 }
2617 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2618 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2619 return;
2620 }
2621 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
2622 _reg_node.map(def_reg,pinch); // Record pinch-point
2623 //_regalloc->set_bad(pinch->_idx); // Already initialized this way.
2624 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2625 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call
2626 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2627 later_def = NULL; // and no later def
2628 }
2629 pinch->set_req(0,later_def); // Hook later def so we can find it
2630 } else { // Else have valid pinch point
2631 if( pinch->in(0) ) // If there is a later-def
2632 later_def = pinch->in(0); // Get it
2633 }
2634
2635 // Add output-dependence edge from later def to kill
2636 if( later_def ) // If there is some original def
2637 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2638
2639 // See if current kill is also a use, and so is forced to be the pinch-point.
2640 if( pinch->Opcode() == Op_Node ) {
2641 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2642 for( uint i=1; i<uses->req(); i++ ) {
2643 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2644 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2645 // Yes, found a use/kill pinch-point
2646 pinch->set_req(0,NULL); //
2647 pinch->replace_by(kill); // Move anti-dep edges up
2648 pinch = kill;
2649 _reg_node.map(def_reg,pinch);
2650 return;
2651 }
2652 }
2653 }
2654
2655 // Add edge from kill to pinch-point
2656 add_prec_edge_from_to(kill,pinch);
2657 }
2658
2659 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2660 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2661 return;
2662 Node *pinch = _reg_node[use_reg]; // Get pinch point
2663 // Check for no later def_reg/kill in block
2664 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b &&
2665 // Use has to be block-local as well
2666 _cfg->get_block_for_node(use) == b) {
2667 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2668 pinch->req() == 1 ) { // pinch not yet in block?
2669 pinch->del_req(0); // yank pointer to later-def, also set flag
2670 // Insert the pinch-point in the block just after the last use
2671 b->insert_node(pinch, b->find_node(use) + 1);
2672 _bb_end++; // Increase size scheduled region in block
2673 }
2674
2675 add_prec_edge_from_to(pinch,use);
2676 }
2677 }
2678
2679 // We insert antidependences between the reads and following write of
2680 // allocated registers to prevent illegal code motion. Hopefully, the
2681 // number of added references should be fairly small, especially as we
2682 // are only adding references within the current basic block.
2683 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2684
2685 #ifdef ASSERT
2686 verify_good_schedule(b,"before block local scheduling");
2687 #endif
2688
2689 // A valid schedule, for each register independently, is an endless cycle
2690 // of: a def, then some uses (connected to the def by true dependencies),
2691 // then some kills (defs with no uses), finally the cycle repeats with a new
2692 // def. The uses are allowed to float relative to each other, as are the
2693 // kills. No use is allowed to slide past a kill (or def). This requires
2694 // antidependencies between all uses of a single def and all kills that
2695 // follow, up to the next def. More edges are redundant, because later defs
2696 // & kills are already serialized with true or antidependencies. To keep
2697 // the edge count down, we add a 'pinch point' node if there's more than
2698 // one use or more than one kill/def.
2699
2700 // We add dependencies in one bottom-up pass.
2701
2702 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2703
2704 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2705 // register. If not, we record the DEF/KILL in _reg_node, the
2706 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
2707 // "pinch point", a new Node that's in the graph but not in the block.
2708 // We put edges from the prior and current DEF/KILLs to the pinch point.
2709 // We put the pinch point in _reg_node. If there's already a pinch point
2710 // we merely add an edge from the current DEF/KILL to the pinch point.
2711
2712 // After doing the DEF/KILLs, we handle USEs. For each used register, we
2713 // put an edge from the pinch point to the USE.
2714
2715 // To be expedient, the _reg_node array is pre-allocated for the whole
2716 // compilation. _reg_node is lazily initialized; it either contains a NULL,
2717 // or a valid def/kill/pinch-point, or a leftover node from some prior
2718 // block. Leftover node from some prior block is treated like a NULL (no
2719 // prior def, so no anti-dependence needed). Valid def is distinguished by
2720 // it being in the current block.
2721 bool fat_proj_seen = false;
2722 uint last_safept = _bb_end-1;
2723 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL;
2724 Node* last_safept_node = end_node;
2725 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2726 Node *n = b->get_node(i);
2727 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
2728 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
2729 // Fat-proj kills a slew of registers
2730 // This can add edges to 'n' and obscure whether or not it was a def,
2731 // hence the is_def flag.
2732 fat_proj_seen = true;
2733 RegMask rm = n->out_RegMask();// Make local copy
2734 while( rm.is_NotEmpty() ) {
2735 OptoReg::Name kill = rm.find_first_elem();
2736 rm.Remove(kill);
2737 anti_do_def( b, n, kill, is_def );
2738 }
2739 } else {
2740 // Get DEF'd registers the normal way
2741 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2742 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2743 }
2744
2745 // Kill projections on a branch should appear to occur on the
2746 // branch, not afterwards, so grab the masks from the projections
2747 // and process them.
2748 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
2749 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2750 Node* use = n->fast_out(i);
2751 if (use->is_Proj()) {
2752 RegMask rm = use->out_RegMask();// Make local copy
2753 while( rm.is_NotEmpty() ) {
2754 OptoReg::Name kill = rm.find_first_elem();
2755 rm.Remove(kill);
2756 anti_do_def( b, n, kill, false );
2757 }
2758 }
2759 }
2760 }
2761
2762 // Check each register used by this instruction for a following DEF/KILL
2763 // that must occur afterward and requires an anti-dependence edge.
2764 for( uint j=0; j<n->req(); j++ ) {
2765 Node *def = n->in(j);
2766 if( def ) {
2767 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
2768 anti_do_use( b, n, _regalloc->get_reg_first(def) );
2769 anti_do_use( b, n, _regalloc->get_reg_second(def) );
2770 }
2771 }
2772 // Do not allow defs of new derived values to float above GC
2773 // points unless the base is definitely available at the GC point.
2774
2775 Node *m = b->get_node(i);
2776
2777 // Add precedence edge from following safepoint to use of derived pointer
2778 if( last_safept_node != end_node &&
2779 m != last_safept_node) {
2780 for (uint k = 1; k < m->req(); k++) {
2781 const Type *t = m->in(k)->bottom_type();
2782 if( t->isa_oop_ptr() &&
2783 t->is_ptr()->offset() != 0 ) {
2784 last_safept_node->add_prec( m );
2785 break;
2786 }
2787 }
2788 }
2789
2790 if( n->jvms() ) { // Precedence edge from derived to safept
2791 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2792 if( b->get_node(last_safept) != last_safept_node ) {
2793 last_safept = b->find_node(last_safept_node);
2794 }
2795 for( uint j=last_safept; j > i; j-- ) {
2796 Node *mach = b->get_node(j);
2797 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2798 mach->add_prec( n );
2799 }
2800 last_safept = i;
2801 last_safept_node = m;
2802 }
2803 }
2804
2805 if (fat_proj_seen) {
2806 // Garbage collect pinch nodes that were not consumed.
2807 // They are usually created by a fat kill MachProj for a call.
2808 garbage_collect_pinch_nodes();
2809 }
2810 }
2811
2812 // Garbage collect pinch nodes for reuse by other blocks.
2813 //
2814 // The block scheduler's insertion of anti-dependence
2815 // edges creates many pinch nodes when the block contains
2816 // 2 or more Calls. A pinch node is used to prevent a
2817 // combinatorial explosion of edges. If a set of kills for a
2818 // register is anti-dependent on a set of uses (or defs), rather
2819 // than adding an edge in the graph between each pair of kill
2820 // and use (or def), a pinch is inserted between them:
2821 //
2822 // use1 use2 use3
2823 // \ | /
2824 // \ | /
2825 // pinch
2826 // / | \
2827 // / | \
2828 // kill1 kill2 kill3
2829 //
2830 // One pinch node is created per register killed when
2831 // the second call is encountered during a backwards pass
2832 // over the block. Most of these pinch nodes are never
2833 // wired into the graph because the register is never
2834 // used or def'ed in the block.
2835 //
2836 void Scheduling::garbage_collect_pinch_nodes() {
2837 #ifndef PRODUCT
2838 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
2839 #endif
2840 int trace_cnt = 0;
2841 for (uint k = 0; k < _reg_node.Size(); k++) {
2842 Node* pinch = _reg_node[k];
2843 if ((pinch != NULL) && pinch->Opcode() == Op_Node &&
2844 // no predecence input edges
2845 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
2846 cleanup_pinch(pinch);
2847 _pinch_free_list.push(pinch);
2848 _reg_node.map(k, NULL);
2849 #ifndef PRODUCT
2850 if (_cfg->C->trace_opto_output()) {
2851 trace_cnt++;
2852 if (trace_cnt > 40) {
2853 tty->print("\n");
2854 trace_cnt = 0;
2855 }
2856 tty->print(" %d", pinch->_idx);
2857 }
2858 #endif
2859 }
2860 }
2861 #ifndef PRODUCT
2862 if (_cfg->C->trace_opto_output()) tty->print("\n");
2863 #endif
2864 }
2865
2866 // Clean up a pinch node for reuse.
2867 void Scheduling::cleanup_pinch( Node *pinch ) {
2868 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
2869
2870 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
2871 Node* use = pinch->last_out(i);
2872 uint uses_found = 0;
2873 for (uint j = use->req(); j < use->len(); j++) {
2874 if (use->in(j) == pinch) {
2875 use->rm_prec(j);
2876 uses_found++;
2877 }
2878 }
2879 assert(uses_found > 0, "must be a precedence edge");
2880 i -= uses_found; // we deleted 1 or more copies of this edge
2881 }
2882 // May have a later_def entry
2883 pinch->set_req(0, NULL);
2884 }
2885
2886 #ifndef PRODUCT
2887
2888 void Scheduling::dump_available() const {
2889 tty->print("#Availist ");
2890 for (uint i = 0; i < _available.size(); i++)
2891 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
2892 tty->cr();
2893 }
2894
2895 // Print Scheduling Statistics
2896 void Scheduling::print_statistics() {
2897 // Print the size added by nops for bundling
2898 tty->print("Nops added %d bytes to total of %d bytes",
2899 _total_nop_size, _total_method_size);
2900 if (_total_method_size > 0)
2901 tty->print(", for %.2f%%",
2902 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
2903 tty->print("\n");
2904
2905 // Print the number of branch shadows filled
2906 if (Pipeline::_branch_has_delay_slot) {
2907 tty->print("Of %d branches, %d had unconditional delay slots filled",
2908 _total_branches, _total_unconditional_delays);
2909 if (_total_branches > 0)
2910 tty->print(", for %.2f%%",
2911 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
2912 tty->print("\n");
2913 }
2914
2915 uint total_instructions = 0, total_bundles = 0;
2916
2917 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
2918 uint bundle_count = _total_instructions_per_bundle[i];
2919 total_instructions += bundle_count * i;
2920 total_bundles += bundle_count;
2921 }
2922
2923 if (total_bundles > 0)
2924 tty->print("Average ILP (excluding nops) is %.2f\n",
2925 ((double)total_instructions) / ((double)total_bundles));
2926 }
2927 #endif