1 /*
2 * Copyright (c) 1998, 2014, 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 "code/compiledIC.hpp"
28 #include "code/debugInfo.hpp"
29 #include "code/debugInfoRec.hpp"
30 #include "compiler/compileBroker.hpp"
31 #include "compiler/compilerDirectives.hpp"
32 #include "compiler/oopMap.hpp"
33 #include "memory/allocation.inline.hpp"
34 #include "opto/ad.hpp"
35 #include "opto/callnode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/locknode.hpp"
38 #include "opto/machnode.hpp"
39 #include "opto/optoreg.hpp"
40 #include "opto/output.hpp"
41 #include "opto/regalloc.hpp"
42 #include "opto/runtime.hpp"
43 #include "opto/subnode.hpp"
44 #include "opto/type.hpp"
45 #include "runtime/handles.inline.hpp"
46 #include "utilities/xmlstream.hpp"
47
48 #ifndef PRODUCT
49 #define DEBUG_ARG(x) , x
50 #else
51 #define DEBUG_ARG(x)
52 #endif
53
54 // Convert Nodes to instruction bits and pass off to the VM
55 void Compile::Output() {
56 // RootNode goes
57 assert( _cfg->get_root_block()->number_of_nodes() == 0, "" );
58
59 // The number of new nodes (mostly MachNop) is proportional to
60 // the number of java calls and inner loops which are aligned.
61 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
62 C->inner_loops()*(OptoLoopAlignment-1)),
63 "out of nodes before code generation" ) ) {
64 return;
65 }
66 // Make sure I can find the Start Node
67 Block *entry = _cfg->get_block(1);
68 Block *broot = _cfg->get_root_block();
69
70 const StartNode *start = entry->head()->as_Start();
71
72 // Replace StartNode with prolog
73 MachPrologNode *prolog = new MachPrologNode();
74 entry->map_node(prolog, 0);
75 _cfg->map_node_to_block(prolog, entry);
76 _cfg->unmap_node_from_block(start); // start is no longer in any block
77
78 // Virtual methods need an unverified entry point
79
80 if( is_osr_compilation() ) {
81 if( PoisonOSREntry ) {
82 // TODO: Should use a ShouldNotReachHereNode...
83 _cfg->insert( broot, 0, new MachBreakpointNode() );
84 }
85 } else {
86 if( _method && !_method->flags().is_static() ) {
87 // Insert unvalidated entry point
88 _cfg->insert( broot, 0, new MachUEPNode() );
89 }
90
91 }
92
93 // Break before main entry point
94 if( (_method && C->directive()->BreakAtExecuteOption)
95 #ifndef PRODUCT
96 ||(OptoBreakpoint && is_method_compilation())
97 ||(OptoBreakpointOSR && is_osr_compilation())
98 ||(OptoBreakpointC2R && !_method)
99 #endif
100 ) {
101 // checking for _method means that OptoBreakpoint does not apply to
102 // runtime stubs or frame converters
103 _cfg->insert( entry, 1, new MachBreakpointNode() );
104 }
105
106 // Insert epilogs before every return
107 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
108 Block* block = _cfg->get_block(i);
109 if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point?
110 Node* m = block->end();
111 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
112 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
113 block->add_inst(epilog);
114 _cfg->map_node_to_block(epilog, block);
115 }
116 }
117 }
118
119 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1);
120 blk_starts[0] = 0;
121
122 // Initialize code buffer and process short branches.
123 CodeBuffer* cb = init_buffer(blk_starts);
124
125 if (cb == NULL || failing()) {
126 return;
127 }
128
129 ScheduleAndBundle();
130
131 #ifndef PRODUCT
132 if (trace_opto_output()) {
133 tty->print("\n---- After ScheduleAndBundle ----\n");
134 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
135 tty->print("\nBB#%03d:\n", i);
136 Block* block = _cfg->get_block(i);
137 for (uint j = 0; j < block->number_of_nodes(); j++) {
138 Node* n = block->get_node(j);
139 OptoReg::Name reg = _regalloc->get_reg_first(n);
140 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
141 n->dump();
142 }
143 }
144 }
145 #endif
146
147 if (failing()) {
148 return;
149 }
150
151 BuildOopMaps();
152
153 if (failing()) {
154 return;
155 }
156
157 fill_buffer(cb, blk_starts);
158 }
159
160 bool Compile::need_stack_bang(int frame_size_in_bytes) const {
161 // Determine if we need to generate a stack overflow check.
162 // Do it if the method is not a stub function and
163 // has java calls or has frame size > vm_page_size/8.
164 // The debug VM checks that deoptimization doesn't trigger an
165 // unexpected stack overflow (compiled method stack banging should
166 // guarantee it doesn't happen) so we always need the stack bang in
167 // a debug VM.
168 return (UseStackBanging && stub_function() == NULL &&
169 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3
170 DEBUG_ONLY(|| true)));
171 }
172
173 bool Compile::need_register_stack_bang() const {
174 // Determine if we need to generate a register stack overflow check.
175 // This is only used on architectures which have split register
176 // and memory stacks (ie. IA64).
177 // Bang if the method is not a stub function and has java calls
178 return (stub_function() == NULL && has_java_calls());
179 }
180
181
182 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
183 // of a loop. When aligning a loop we need to provide enough instructions
184 // in cpu's fetch buffer to feed decoders. The loop alignment could be
185 // avoided if we have enough instructions in fetch buffer at the head of a loop.
186 // By default, the size is set to 999999 by Block's constructor so that
187 // a loop will be aligned if the size is not reset here.
188 //
189 // Note: Mach instructions could contain several HW instructions
190 // so the size is estimated only.
191 //
192 void Compile::compute_loop_first_inst_sizes() {
193 // The next condition is used to gate the loop alignment optimization.
194 // Don't aligned a loop if there are enough instructions at the head of a loop
195 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
196 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
197 // equal to 11 bytes which is the largest address NOP instruction.
198 if (MaxLoopPad < OptoLoopAlignment - 1) {
199 uint last_block = _cfg->number_of_blocks() - 1;
200 for (uint i = 1; i <= last_block; i++) {
201 Block* block = _cfg->get_block(i);
202 // Check the first loop's block which requires an alignment.
203 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
204 uint sum_size = 0;
205 uint inst_cnt = NumberOfLoopInstrToAlign;
206 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
207
208 // Check subsequent fallthrough blocks if the loop's first
209 // block(s) does not have enough instructions.
210 Block *nb = block;
211 while(inst_cnt > 0 &&
212 i < last_block &&
213 !_cfg->get_block(i + 1)->has_loop_alignment() &&
214 !nb->has_successor(block)) {
215 i++;
216 nb = _cfg->get_block(i);
217 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
218 } // while( inst_cnt > 0 && i < last_block )
219
220 block->set_first_inst_size(sum_size);
221 } // f( b->head()->is_Loop() )
222 } // for( i <= last_block )
223 } // if( MaxLoopPad < OptoLoopAlignment-1 )
224 }
225
226 // The architecture description provides short branch variants for some long
227 // branch instructions. Replace eligible long branches with short branches.
228 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) {
229 // Compute size of each block, method size, and relocation information size
230 uint nblocks = _cfg->number_of_blocks();
231
232 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
233 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
234 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
235
236 // Collect worst case block paddings
237 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
238 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
239
240 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
241 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
242
243 bool has_short_branch_candidate = false;
244
245 // Initialize the sizes to 0
246 code_size = 0; // Size in bytes of generated code
247 stub_size = 0; // Size in bytes of all stub entries
248 // Size in bytes of all relocation entries, including those in local stubs.
249 // Start with 2-bytes of reloc info for the unvalidated entry point
250 reloc_size = 1; // Number of relocation entries
251
252 // Make three passes. The first computes pessimistic blk_starts,
253 // relative jmp_offset and reloc_size information. The second performs
254 // short branch substitution using the pessimistic sizing. The
255 // third inserts nops where needed.
256
257 // Step one, perform a pessimistic sizing pass.
258 uint last_call_adr = max_juint;
259 uint last_avoid_back_to_back_adr = max_juint;
260 uint nop_size = (new MachNopNode())->size(_regalloc);
261 for (uint i = 0; i < nblocks; i++) { // For all blocks
262 Block* block = _cfg->get_block(i);
263
264 // During short branch replacement, we store the relative (to blk_starts)
265 // offset of jump in jmp_offset, rather than the absolute offset of jump.
266 // This is so that we do not need to recompute sizes of all nodes when
267 // we compute correct blk_starts in our next sizing pass.
268 jmp_offset[i] = 0;
269 jmp_size[i] = 0;
270 jmp_nidx[i] = -1;
271 DEBUG_ONLY( jmp_target[i] = 0; )
272 DEBUG_ONLY( jmp_rule[i] = 0; )
273
274 // Sum all instruction sizes to compute block size
275 uint last_inst = block->number_of_nodes();
276 uint blk_size = 0;
277 for (uint j = 0; j < last_inst; j++) {
278 Node* nj = block->get_node(j);
279 // Handle machine instruction nodes
280 if (nj->is_Mach()) {
281 MachNode *mach = nj->as_Mach();
282 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
283 reloc_size += mach->reloc();
284 if (mach->is_MachCall()) {
285 // add size information for trampoline stub
286 // class CallStubImpl is platform-specific and defined in the *.ad files.
287 stub_size += CallStubImpl::size_call_trampoline();
288 reloc_size += CallStubImpl::reloc_call_trampoline();
289
290 MachCallNode *mcall = mach->as_MachCall();
291 // This destination address is NOT PC-relative
292
293 mcall->method_set((intptr_t)mcall->entry_point());
294
295 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
296 stub_size += CompiledStaticCall::to_interp_stub_size();
297 reloc_size += CompiledStaticCall::reloc_to_interp_stub();
298 }
299 } else if (mach->is_MachSafePoint()) {
300 // If call/safepoint are adjacent, account for possible
301 // nop to disambiguate the two safepoints.
302 // ScheduleAndBundle() can rearrange nodes in a block,
303 // check for all offsets inside this block.
304 if (last_call_adr >= blk_starts[i]) {
305 blk_size += nop_size;
306 }
307 }
308 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
309 // Nop is inserted between "avoid back to back" instructions.
310 // ScheduleAndBundle() can rearrange nodes in a block,
311 // check for all offsets inside this block.
312 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
313 blk_size += nop_size;
314 }
315 }
316 if (mach->may_be_short_branch()) {
317 if (!nj->is_MachBranch()) {
318 #ifndef PRODUCT
319 nj->dump(3);
320 #endif
321 Unimplemented();
322 }
323 assert(jmp_nidx[i] == -1, "block should have only one branch");
324 jmp_offset[i] = blk_size;
325 jmp_size[i] = nj->size(_regalloc);
326 jmp_nidx[i] = j;
327 has_short_branch_candidate = true;
328 }
329 }
330 blk_size += nj->size(_regalloc);
331 // Remember end of call offset
332 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
333 last_call_adr = blk_starts[i]+blk_size;
334 }
335 // Remember end of avoid_back_to_back offset
336 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
337 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
338 }
339 }
340
341 // When the next block starts a loop, we may insert pad NOP
342 // instructions. Since we cannot know our future alignment,
343 // assume the worst.
344 if (i < nblocks - 1) {
345 Block* nb = _cfg->get_block(i + 1);
346 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
347 if (max_loop_pad > 0) {
348 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
349 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
350 // If either is the last instruction in this block, bump by
351 // max_loop_pad in lock-step with blk_size, so sizing
352 // calculations in subsequent blocks still can conservatively
353 // detect that it may the last instruction in this block.
354 if (last_call_adr == blk_starts[i]+blk_size) {
355 last_call_adr += max_loop_pad;
356 }
357 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
358 last_avoid_back_to_back_adr += max_loop_pad;
359 }
360 blk_size += max_loop_pad;
361 block_worst_case_pad[i + 1] = max_loop_pad;
362 }
363 }
364
365 // Save block size; update total method size
366 blk_starts[i+1] = blk_starts[i]+blk_size;
367 }
368
369 // Step two, replace eligible long jumps.
370 bool progress = true;
371 uint last_may_be_short_branch_adr = max_juint;
372 while (has_short_branch_candidate && progress) {
373 progress = false;
374 has_short_branch_candidate = false;
375 int adjust_block_start = 0;
376 for (uint i = 0; i < nblocks; i++) {
377 Block* block = _cfg->get_block(i);
378 int idx = jmp_nidx[i];
379 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach();
380 if (mach != NULL && mach->may_be_short_branch()) {
381 #ifdef ASSERT
382 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
383 int j;
384 // Find the branch; ignore trailing NOPs.
385 for (j = block->number_of_nodes()-1; j>=0; j--) {
386 Node* n = block->get_node(j);
387 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
388 break;
389 }
390 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
391 #endif
392 int br_size = jmp_size[i];
393 int br_offs = blk_starts[i] + jmp_offset[i];
394
395 // This requires the TRUE branch target be in succs[0]
396 uint bnum = block->non_connector_successor(0)->_pre_order;
397 int offset = blk_starts[bnum] - br_offs;
398 if (bnum > i) { // adjust following block's offset
399 offset -= adjust_block_start;
400 }
401
402 // This block can be a loop header, account for the padding
403 // in the previous block.
404 int block_padding = block_worst_case_pad[i];
405 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
406 // In the following code a nop could be inserted before
407 // the branch which will increase the backward distance.
408 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
409 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
410
411 if (needs_padding && offset <= 0)
412 offset -= nop_size;
413
414 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
415 // We've got a winner. Replace this branch.
416 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
417
418 // Update the jmp_size.
419 int new_size = replacement->size(_regalloc);
420 int diff = br_size - new_size;
421 assert(diff >= (int)nop_size, "short_branch size should be smaller");
422 // Conservatively take into account padding between
423 // avoid_back_to_back branches. Previous branch could be
424 // converted into avoid_back_to_back branch during next
425 // rounds.
426 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
427 jmp_offset[i] += nop_size;
428 diff -= nop_size;
429 }
430 adjust_block_start += diff;
431 block->map_node(replacement, idx);
432 mach->subsume_by(replacement, C);
433 mach = replacement;
434 progress = true;
435
436 jmp_size[i] = new_size;
437 DEBUG_ONLY( jmp_target[i] = bnum; );
438 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
439 } else {
440 // The jump distance is not short, try again during next iteration.
441 has_short_branch_candidate = true;
442 }
443 } // (mach->may_be_short_branch())
444 if (mach != NULL && (mach->may_be_short_branch() ||
445 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
446 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
447 }
448 blk_starts[i+1] -= adjust_block_start;
449 }
450 }
451
452 #ifdef ASSERT
453 for (uint i = 0; i < nblocks; i++) { // For all blocks
454 if (jmp_target[i] != 0) {
455 int br_size = jmp_size[i];
456 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
457 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
458 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]);
459 }
460 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
461 }
462 }
463 #endif
464
465 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
466 // after ScheduleAndBundle().
467
468 // ------------------
469 // Compute size for code buffer
470 code_size = blk_starts[nblocks];
471
472 // Relocation records
473 reloc_size += 1; // Relo entry for exception handler
474
475 // Adjust reloc_size to number of record of relocation info
476 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
477 // a relocation index.
478 // The CodeBuffer will expand the locs array if this estimate is too low.
479 reloc_size *= 10 / sizeof(relocInfo);
480 }
481
482 //------------------------------FillLocArray-----------------------------------
483 // Create a bit of debug info and append it to the array. The mapping is from
484 // Java local or expression stack to constant, register or stack-slot. For
485 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
486 // entry has been taken care of and caller should skip it).
487 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
488 // This should never have accepted Bad before
489 assert(OptoReg::is_valid(regnum), "location must be valid");
490 return (OptoReg::is_reg(regnum))
491 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
492 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
493 }
494
495
496 ObjectValue*
497 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
498 for (int i = 0; i < objs->length(); i++) {
499 assert(objs->at(i)->is_object(), "corrupt object cache");
500 ObjectValue* sv = (ObjectValue*) objs->at(i);
501 if (sv->id() == id) {
502 return sv;
503 }
504 }
505 // Otherwise..
506 return NULL;
507 }
508
509 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
510 ObjectValue* sv ) {
511 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
512 objs->append(sv);
513 }
514
515
516 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
517 GrowableArray<ScopeValue*> *array,
518 GrowableArray<ScopeValue*> *objs ) {
519 assert( local, "use _top instead of null" );
520 if (array->length() != idx) {
521 assert(array->length() == idx + 1, "Unexpected array count");
522 // Old functionality:
523 // return
524 // New functionality:
525 // Assert if the local is not top. In product mode let the new node
526 // override the old entry.
527 assert(local == top(), "LocArray collision");
528 if (local == top()) {
529 return;
530 }
531 array->pop();
532 }
533 const Type *t = local->bottom_type();
534
535 // Is it a safepoint scalar object node?
536 if (local->is_SafePointScalarObject()) {
537 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
538
539 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
540 if (sv == NULL) {
541 ciKlass* cik = t->is_oopptr()->klass();
542 assert(cik->is_instance_klass() ||
543 cik->is_array_klass(), "Not supported allocation.");
544 sv = new ObjectValue(spobj->_idx,
545 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
546 Compile::set_sv_for_object_node(objs, sv);
547
548 uint first_ind = spobj->first_index(sfpt->jvms());
549 for (uint i = 0; i < spobj->n_fields(); i++) {
550 Node* fld_node = sfpt->in(first_ind+i);
551 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
552 }
553 }
554 array->append(sv);
555 return;
556 }
557
558 // Grab the register number for the local
559 OptoReg::Name regnum = _regalloc->get_reg_first(local);
560 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
561 // Record the double as two float registers.
562 // The register mask for such a value always specifies two adjacent
563 // float registers, with the lower register number even.
564 // Normally, the allocation of high and low words to these registers
565 // is irrelevant, because nearly all operations on register pairs
566 // (e.g., StoreD) treat them as a single unit.
567 // Here, we assume in addition that the words in these two registers
568 // stored "naturally" (by operations like StoreD and double stores
569 // within the interpreter) such that the lower-numbered register
570 // is written to the lower memory address. This may seem like
571 // a machine dependency, but it is not--it is a requirement on
572 // the author of the <arch>.ad file to ensure that, for every
573 // even/odd double-register pair to which a double may be allocated,
574 // the word in the even single-register is stored to the first
575 // memory word. (Note that register numbers are completely
576 // arbitrary, and are not tied to any machine-level encodings.)
577 #ifdef _LP64
578 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
579 array->append(new ConstantIntValue(0));
580 array->append(new_loc_value( _regalloc, regnum, Location::dbl ));
581 } else if ( t->base() == Type::Long ) {
582 array->append(new ConstantIntValue(0));
583 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
584 } else if ( t->base() == Type::RawPtr ) {
585 // jsr/ret return address which must be restored into a the full
586 // width 64-bit stack slot.
587 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
588 }
589 #else //_LP64
590 #ifdef SPARC
591 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {
592 // For SPARC we have to swap high and low words for
593 // long values stored in a single-register (g0-g7).
594 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
595 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
596 } else
597 #endif //SPARC
598 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
599 // Repack the double/long as two jints.
600 // The convention the interpreter uses is that the second local
601 // holds the first raw word of the native double representation.
602 // This is actually reasonable, since locals and stack arrays
603 // grow downwards in all implementations.
604 // (If, on some machine, the interpreter's Java locals or stack
605 // were to grow upwards, the embedded doubles would be word-swapped.)
606 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
607 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
608 }
609 #endif //_LP64
610 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
611 OptoReg::is_reg(regnum) ) {
612 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()
613 ? Location::float_in_dbl : Location::normal ));
614 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
615 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long
616 ? Location::int_in_long : Location::normal ));
617 } else if( t->base() == Type::NarrowOop ) {
618 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));
619 } else {
620 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));
621 }
622 return;
623 }
624
625 // No register. It must be constant data.
626 switch (t->base()) {
627 case Type::Half: // Second half of a double
628 ShouldNotReachHere(); // Caller should skip 2nd halves
629 break;
630 case Type::AnyPtr:
631 array->append(new ConstantOopWriteValue(NULL));
632 break;
633 case Type::AryPtr:
634 case Type::InstPtr: // fall through
635 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
636 break;
637 case Type::NarrowOop:
638 if (t == TypeNarrowOop::NULL_PTR) {
639 array->append(new ConstantOopWriteValue(NULL));
640 } else {
641 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
642 }
643 break;
644 case Type::Int:
645 array->append(new ConstantIntValue(t->is_int()->get_con()));
646 break;
647 case Type::RawPtr:
648 // A return address (T_ADDRESS).
649 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
650 #ifdef _LP64
651 // Must be restored to the full-width 64-bit stack slot.
652 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
653 #else
654 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
655 #endif
656 break;
657 case Type::FloatCon: {
658 float f = t->is_float_constant()->getf();
659 array->append(new ConstantIntValue(jint_cast(f)));
660 break;
661 }
662 case Type::DoubleCon: {
663 jdouble d = t->is_double_constant()->getd();
664 #ifdef _LP64
665 array->append(new ConstantIntValue(0));
666 array->append(new ConstantDoubleValue(d));
667 #else
668 // Repack the double as two jints.
669 // The convention the interpreter uses is that the second local
670 // holds the first raw word of the native double representation.
671 // This is actually reasonable, since locals and stack arrays
672 // grow downwards in all implementations.
673 // (If, on some machine, the interpreter's Java locals or stack
674 // were to grow upwards, the embedded doubles would be word-swapped.)
675 jlong_accessor acc;
676 acc.long_value = jlong_cast(d);
677 array->append(new ConstantIntValue(acc.words[1]));
678 array->append(new ConstantIntValue(acc.words[0]));
679 #endif
680 break;
681 }
682 case Type::Long: {
683 jlong d = t->is_long()->get_con();
684 #ifdef _LP64
685 array->append(new ConstantIntValue(0));
686 array->append(new ConstantLongValue(d));
687 #else
688 // Repack the long as two jints.
689 // The convention the interpreter uses is that the second local
690 // holds the first raw word of the native double representation.
691 // This is actually reasonable, since locals and stack arrays
692 // grow downwards in all implementations.
693 // (If, on some machine, the interpreter's Java locals or stack
694 // were to grow upwards, the embedded doubles would be word-swapped.)
695 jlong_accessor acc;
696 acc.long_value = d;
697 array->append(new ConstantIntValue(acc.words[1]));
698 array->append(new ConstantIntValue(acc.words[0]));
699 #endif
700 break;
701 }
702 case Type::Top: // Add an illegal value here
703 array->append(new LocationValue(Location()));
704 break;
705 default:
706 ShouldNotReachHere();
707 break;
708 }
709 }
710
711 // Determine if this node starts a bundle
712 bool Compile::starts_bundle(const Node *n) const {
713 return (_node_bundling_limit > n->_idx &&
714 _node_bundling_base[n->_idx].starts_bundle());
715 }
716
717 //--------------------------Process_OopMap_Node--------------------------------
718 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
719
720 // Handle special safepoint nodes for synchronization
721 MachSafePointNode *sfn = mach->as_MachSafePoint();
722 MachCallNode *mcall;
723
724 int safepoint_pc_offset = current_offset;
725 bool is_method_handle_invoke = false;
726 bool return_oop = false;
727
728 // Add the safepoint in the DebugInfoRecorder
729 if( !mach->is_MachCall() ) {
730 mcall = NULL;
731 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
732 } else {
733 mcall = mach->as_MachCall();
734
735 // Is the call a MethodHandle call?
736 if (mcall->is_MachCallJava()) {
737 if (mcall->as_MachCallJava()->_method_handle_invoke) {
738 assert(has_method_handle_invokes(), "must have been set during call generation");
739 is_method_handle_invoke = true;
740 }
741 }
742
743 // Check if a call returns an object.
744 if (mcall->returns_pointer()) {
745 return_oop = true;
746 }
747 safepoint_pc_offset += mcall->ret_addr_offset();
748 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
749 }
750
751 // Loop over the JVMState list to add scope information
752 // Do not skip safepoints with a NULL method, they need monitor info
753 JVMState* youngest_jvms = sfn->jvms();
754 int max_depth = youngest_jvms->depth();
755
756 // Allocate the object pool for scalar-replaced objects -- the map from
757 // small-integer keys (which can be recorded in the local and ostack
758 // arrays) to descriptions of the object state.
759 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
760
761 // Visit scopes from oldest to youngest.
762 for (int depth = 1; depth <= max_depth; depth++) {
763 JVMState* jvms = youngest_jvms->of_depth(depth);
764 int idx;
765 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
766 // Safepoints that do not have method() set only provide oop-map and monitor info
767 // to support GC; these do not support deoptimization.
768 int num_locs = (method == NULL) ? 0 : jvms->loc_size();
769 int num_exps = (method == NULL) ? 0 : jvms->stk_size();
770 int num_mon = jvms->nof_monitors();
771 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
772 "JVMS local count must match that of the method");
773
774 // Add Local and Expression Stack Information
775
776 // Insert locals into the locarray
777 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
778 for( idx = 0; idx < num_locs; idx++ ) {
779 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
780 }
781
782 // Insert expression stack entries into the exparray
783 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
784 for( idx = 0; idx < num_exps; idx++ ) {
785 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
786 }
787
788 // Add in mappings of the monitors
789 assert( !method ||
790 !method->is_synchronized() ||
791 method->is_native() ||
792 num_mon > 0 ||
793 !GenerateSynchronizationCode,
794 "monitors must always exist for synchronized methods");
795
796 // Build the growable array of ScopeValues for exp stack
797 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
798
799 // Loop over monitors and insert into array
800 for (idx = 0; idx < num_mon; idx++) {
801 // Grab the node that defines this monitor
802 Node* box_node = sfn->monitor_box(jvms, idx);
803 Node* obj_node = sfn->monitor_obj(jvms, idx);
804
805 // Create ScopeValue for object
806 ScopeValue *scval = NULL;
807
808 if (obj_node->is_SafePointScalarObject()) {
809 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
810 scval = Compile::sv_for_node_id(objs, spobj->_idx);
811 if (scval == NULL) {
812 const Type *t = spobj->bottom_type();
813 ciKlass* cik = t->is_oopptr()->klass();
814 assert(cik->is_instance_klass() ||
815 cik->is_array_klass(), "Not supported allocation.");
816 ObjectValue* sv = new ObjectValue(spobj->_idx,
817 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
818 Compile::set_sv_for_object_node(objs, sv);
819
820 uint first_ind = spobj->first_index(youngest_jvms);
821 for (uint i = 0; i < spobj->n_fields(); i++) {
822 Node* fld_node = sfn->in(first_ind+i);
823 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
824 }
825 scval = sv;
826 }
827 } else if (!obj_node->is_Con()) {
828 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);
829 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
830 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );
831 } else {
832 scval = new_loc_value( _regalloc, obj_reg, Location::oop );
833 }
834 } else {
835 const TypePtr *tp = obj_node->get_ptr_type();
836 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
837 }
838
839 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
840 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
841 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
842 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
843 }
844
845 // We dump the object pool first, since deoptimization reads it in first.
846 debug_info()->dump_object_pool(objs);
847
848 // Build first class objects to pass to scope
849 DebugToken *locvals = debug_info()->create_scope_values(locarray);
850 DebugToken *expvals = debug_info()->create_scope_values(exparray);
851 DebugToken *monvals = debug_info()->create_monitor_values(monarray);
852
853 // Make method available for all Safepoints
854 ciMethod* scope_method = method ? method : _method;
855 // Describe the scope here
856 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
857 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
858 // Now we can describe the scope.
859 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals);
860 } // End jvms loop
861
862 // Mark the end of the scope set.
863 debug_info()->end_safepoint(safepoint_pc_offset);
864 }
865
866
867
868 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
869 class NonSafepointEmitter {
870 Compile* C;
871 JVMState* _pending_jvms;
872 int _pending_offset;
873
874 void emit_non_safepoint();
875
876 public:
877 NonSafepointEmitter(Compile* compile) {
878 this->C = compile;
879 _pending_jvms = NULL;
880 _pending_offset = 0;
881 }
882
883 void observe_instruction(Node* n, int pc_offset) {
884 if (!C->debug_info()->recording_non_safepoints()) return;
885
886 Node_Notes* nn = C->node_notes_at(n->_idx);
887 if (nn == NULL || nn->jvms() == NULL) return;
888 if (_pending_jvms != NULL &&
889 _pending_jvms->same_calls_as(nn->jvms())) {
890 // Repeated JVMS? Stretch it up here.
891 _pending_offset = pc_offset;
892 } else {
893 if (_pending_jvms != NULL &&
894 _pending_offset < pc_offset) {
895 emit_non_safepoint();
896 }
897 _pending_jvms = NULL;
898 if (pc_offset > C->debug_info()->last_pc_offset()) {
899 // This is the only way _pending_jvms can become non-NULL:
900 _pending_jvms = nn->jvms();
901 _pending_offset = pc_offset;
902 }
903 }
904 }
905
906 // Stay out of the way of real safepoints:
907 void observe_safepoint(JVMState* jvms, int pc_offset) {
908 if (_pending_jvms != NULL &&
909 !_pending_jvms->same_calls_as(jvms) &&
910 _pending_offset < pc_offset) {
911 emit_non_safepoint();
912 }
913 _pending_jvms = NULL;
914 }
915
916 void flush_at_end() {
917 if (_pending_jvms != NULL) {
918 emit_non_safepoint();
919 }
920 _pending_jvms = NULL;
921 }
922 };
923
924 void NonSafepointEmitter::emit_non_safepoint() {
925 JVMState* youngest_jvms = _pending_jvms;
926 int pc_offset = _pending_offset;
927
928 // Clear it now:
929 _pending_jvms = NULL;
930
931 DebugInformationRecorder* debug_info = C->debug_info();
932 assert(debug_info->recording_non_safepoints(), "sanity");
933
934 debug_info->add_non_safepoint(pc_offset);
935 int max_depth = youngest_jvms->depth();
936
937 // Visit scopes from oldest to youngest.
938 for (int depth = 1; depth <= max_depth; depth++) {
939 JVMState* jvms = youngest_jvms->of_depth(depth);
940 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
941 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
942 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute());
943 }
944
945 // Mark the end of the scope set.
946 debug_info->end_non_safepoint(pc_offset);
947 }
948
949 //------------------------------init_buffer------------------------------------
950 CodeBuffer* Compile::init_buffer(uint* blk_starts) {
951
952 // Set the initially allocated size
953 int code_req = initial_code_capacity;
954 int locs_req = initial_locs_capacity;
955 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity;
956 int const_req = initial_const_capacity;
957
958 int pad_req = NativeCall::instruction_size;
959 // The extra spacing after the code is necessary on some platforms.
960 // Sometimes we need to patch in a jump after the last instruction,
961 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
962
963 // Compute the byte offset where we can store the deopt pc.
964 if (fixed_slots() != 0) {
965 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
966 }
967
968 // Compute prolog code size
969 _method_size = 0;
970 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;
971 #if defined(IA64) && !defined(AIX)
972 if (save_argument_registers()) {
973 // 4815101: this is a stub with implicit and unknown precision fp args.
974 // The usual spill mechanism can only generate stfd's in this case, which
975 // doesn't work if the fp reg to spill contains a single-precision denorm.
976 // Instead, we hack around the normal spill mechanism using stfspill's and
977 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate
978 // space here for the fp arg regs (f8-f15) we're going to thusly spill.
979 //
980 // If we ever implement 16-byte 'registers' == stack slots, we can
981 // get rid of this hack and have SpillCopy generate stfspill/ldffill
982 // instead of stfd/stfs/ldfd/ldfs.
983 _frame_slots += 8*(16/BytesPerInt);
984 }
985 #endif
986 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
987
988 if (has_mach_constant_base_node()) {
989 uint add_size = 0;
990 // Fill the constant table.
991 // Note: This must happen before shorten_branches.
992 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
993 Block* b = _cfg->get_block(i);
994
995 for (uint j = 0; j < b->number_of_nodes(); j++) {
996 Node* n = b->get_node(j);
997
998 // If the node is a MachConstantNode evaluate the constant
999 // value section.
1000 if (n->is_MachConstant()) {
1001 MachConstantNode* machcon = n->as_MachConstant();
1002 machcon->eval_constant(C);
1003 } else if (n->is_Mach()) {
1004 // On Power there are more nodes that issue constants.
1005 add_size += (n->as_Mach()->ins_num_consts() * 8);
1006 }
1007 }
1008 }
1009
1010 // Calculate the offsets of the constants and the size of the
1011 // constant table (including the padding to the next section).
1012 constant_table().calculate_offsets_and_size();
1013 const_req = constant_table().size() + add_size;
1014 }
1015
1016 // Initialize the space for the BufferBlob used to find and verify
1017 // instruction size in MachNode::emit_size()
1018 init_scratch_buffer_blob(const_req);
1019 if (failing()) return NULL; // Out of memory
1020
1021 // Pre-compute the length of blocks and replace
1022 // long branches with short if machine supports it.
1023 shorten_branches(blk_starts, code_req, locs_req, stub_req);
1024
1025 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1026 // class HandlerImpl is platform-specific and defined in the *.ad files.
1027 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1028 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1029 stub_req += MAX_stubs_size; // ensure per-stub margin
1030 code_req += MAX_inst_size; // ensure per-instruction margin
1031
1032 if (StressCodeBuffers)
1033 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1034
1035 int total_req =
1036 const_req +
1037 code_req +
1038 pad_req +
1039 stub_req +
1040 exception_handler_req +
1041 deopt_handler_req; // deopt handler
1042
1043 if (has_method_handle_invokes())
1044 total_req += deopt_handler_req; // deopt MH handler
1045
1046 CodeBuffer* cb = code_buffer();
1047 cb->initialize(total_req, locs_req);
1048
1049 // Have we run out of code space?
1050 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1051 C->record_failure("CodeCache is full");
1052 return NULL;
1053 }
1054 // Configure the code buffer.
1055 cb->initialize_consts_size(const_req);
1056 cb->initialize_stubs_size(stub_req);
1057 cb->initialize_oop_recorder(env()->oop_recorder());
1058
1059 // fill in the nop array for bundling computations
1060 MachNode *_nop_list[Bundle::_nop_count];
1061 Bundle::initialize_nops(_nop_list);
1062
1063 return cb;
1064 }
1065
1066 //------------------------------fill_buffer------------------------------------
1067 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1068 // blk_starts[] contains offsets calculated during short branches processing,
1069 // offsets should not be increased during following steps.
1070
1071 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1072 // of a loop. It is used to determine the padding for loop alignment.
1073 compute_loop_first_inst_sizes();
1074
1075 // Create oopmap set.
1076 _oop_map_set = new OopMapSet();
1077
1078 // !!!!! This preserves old handling of oopmaps for now
1079 debug_info()->set_oopmaps(_oop_map_set);
1080
1081 uint nblocks = _cfg->number_of_blocks();
1082 // Count and start of implicit null check instructions
1083 uint inct_cnt = 0;
1084 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1085
1086 // Count and start of calls
1087 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1088
1089 uint return_offset = 0;
1090 int nop_size = (new MachNopNode())->size(_regalloc);
1091
1092 int previous_offset = 0;
1093 int current_offset = 0;
1094 int last_call_offset = -1;
1095 int last_avoid_back_to_back_offset = -1;
1096 #ifdef ASSERT
1097 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1098 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1099 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1100 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1101 #endif
1102
1103 // Create an array of unused labels, one for each basic block, if printing is enabled
1104 #ifndef PRODUCT
1105 int *node_offsets = NULL;
1106 uint node_offset_limit = unique();
1107
1108 if (print_assembly())
1109 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1110 #endif
1111
1112 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily
1113
1114 // Emit the constant table.
1115 if (has_mach_constant_base_node()) {
1116 constant_table().emit(*cb);
1117 }
1118
1119 // Create an array of labels, one for each basic block
1120 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1121 for (uint i=0; i <= nblocks; i++) {
1122 blk_labels[i].init();
1123 }
1124
1125 // ------------------
1126 // Now fill in the code buffer
1127 Node *delay_slot = NULL;
1128
1129 for (uint i = 0; i < nblocks; i++) {
1130 Block* block = _cfg->get_block(i);
1131 Node* head = block->head();
1132
1133 // If this block needs to start aligned (i.e, can be reached other
1134 // than by falling-thru from the previous block), then force the
1135 // start of a new bundle.
1136 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1137 cb->flush_bundle(true);
1138 }
1139
1140 #ifdef ASSERT
1141 if (!block->is_connector()) {
1142 stringStream st;
1143 block->dump_head(_cfg, &st);
1144 MacroAssembler(cb).block_comment(st.as_string());
1145 }
1146 jmp_target[i] = 0;
1147 jmp_offset[i] = 0;
1148 jmp_size[i] = 0;
1149 jmp_rule[i] = 0;
1150 #endif
1151 int blk_offset = current_offset;
1152
1153 // Define the label at the beginning of the basic block
1154 MacroAssembler(cb).bind(blk_labels[block->_pre_order]);
1155
1156 uint last_inst = block->number_of_nodes();
1157
1158 // Emit block normally, except for last instruction.
1159 // Emit means "dump code bits into code buffer".
1160 for (uint j = 0; j<last_inst; j++) {
1161
1162 // Get the node
1163 Node* n = block->get_node(j);
1164
1165 // See if delay slots are supported
1166 if (valid_bundle_info(n) &&
1167 node_bundling(n)->used_in_unconditional_delay()) {
1168 assert(delay_slot == NULL, "no use of delay slot node");
1169 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1170
1171 delay_slot = n;
1172 continue;
1173 }
1174
1175 // If this starts a new instruction group, then flush the current one
1176 // (but allow split bundles)
1177 if (Pipeline::requires_bundling() && starts_bundle(n))
1178 cb->flush_bundle(false);
1179
1180 // Special handling for SafePoint/Call Nodes
1181 bool is_mcall = false;
1182 if (n->is_Mach()) {
1183 MachNode *mach = n->as_Mach();
1184 is_mcall = n->is_MachCall();
1185 bool is_sfn = n->is_MachSafePoint();
1186
1187 // If this requires all previous instructions be flushed, then do so
1188 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1189 cb->flush_bundle(true);
1190 current_offset = cb->insts_size();
1191 }
1192
1193 // A padding may be needed again since a previous instruction
1194 // could be moved to delay slot.
1195
1196 // align the instruction if necessary
1197 int padding = mach->compute_padding(current_offset);
1198 // Make sure safepoint node for polling is distinct from a call's
1199 // return by adding a nop if needed.
1200 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1201 padding = nop_size;
1202 }
1203 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1204 current_offset == last_avoid_back_to_back_offset) {
1205 // Avoid back to back some instructions.
1206 padding = nop_size;
1207 }
1208
1209 if(padding > 0) {
1210 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1211 int nops_cnt = padding / nop_size;
1212 MachNode *nop = new MachNopNode(nops_cnt);
1213 block->insert_node(nop, j++);
1214 last_inst++;
1215 _cfg->map_node_to_block(nop, block);
1216 nop->emit(*cb, _regalloc);
1217 cb->flush_bundle(true);
1218 current_offset = cb->insts_size();
1219 }
1220
1221 // Remember the start of the last call in a basic block
1222 if (is_mcall) {
1223 MachCallNode *mcall = mach->as_MachCall();
1224
1225 // This destination address is NOT PC-relative
1226 mcall->method_set((intptr_t)mcall->entry_point());
1227
1228 // Save the return address
1229 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1230
1231 if (mcall->is_MachCallLeaf()) {
1232 is_mcall = false;
1233 is_sfn = false;
1234 }
1235 }
1236
1237 // sfn will be valid whenever mcall is valid now because of inheritance
1238 if (is_sfn || is_mcall) {
1239
1240 // Handle special safepoint nodes for synchronization
1241 if (!is_mcall) {
1242 MachSafePointNode *sfn = mach->as_MachSafePoint();
1243 // !!!!! Stubs only need an oopmap right now, so bail out
1244 if (sfn->jvms()->method() == NULL) {
1245 // Write the oopmap directly to the code blob??!!
1246 continue;
1247 }
1248 } // End synchronization
1249
1250 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1251 current_offset);
1252 Process_OopMap_Node(mach, current_offset);
1253 } // End if safepoint
1254
1255 // If this is a null check, then add the start of the previous instruction to the list
1256 else if( mach->is_MachNullCheck() ) {
1257 inct_starts[inct_cnt++] = previous_offset;
1258 }
1259
1260 // If this is a branch, then fill in the label with the target BB's label
1261 else if (mach->is_MachBranch()) {
1262 // This requires the TRUE branch target be in succs[0]
1263 uint block_num = block->non_connector_successor(0)->_pre_order;
1264
1265 // Try to replace long branch if delay slot is not used,
1266 // it is mostly for back branches since forward branch's
1267 // distance is not updated yet.
1268 bool delay_slot_is_used = valid_bundle_info(n) &&
1269 node_bundling(n)->use_unconditional_delay();
1270 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1271 assert(delay_slot == NULL, "not expecting delay slot node");
1272 int br_size = n->size(_regalloc);
1273 int offset = blk_starts[block_num] - current_offset;
1274 if (block_num >= i) {
1275 // Current and following block's offset are not
1276 // finalized yet, adjust distance by the difference
1277 // between calculated and final offsets of current block.
1278 offset -= (blk_starts[i] - blk_offset);
1279 }
1280 // In the following code a nop could be inserted before
1281 // the branch which will increase the backward distance.
1282 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1283 if (needs_padding && offset <= 0)
1284 offset -= nop_size;
1285
1286 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
1287 // We've got a winner. Replace this branch.
1288 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1289
1290 // Update the jmp_size.
1291 int new_size = replacement->size(_regalloc);
1292 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1293 // Insert padding between avoid_back_to_back branches.
1294 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1295 MachNode *nop = new MachNopNode();
1296 block->insert_node(nop, j++);
1297 _cfg->map_node_to_block(nop, block);
1298 last_inst++;
1299 nop->emit(*cb, _regalloc);
1300 cb->flush_bundle(true);
1301 current_offset = cb->insts_size();
1302 }
1303 #ifdef ASSERT
1304 jmp_target[i] = block_num;
1305 jmp_offset[i] = current_offset - blk_offset;
1306 jmp_size[i] = new_size;
1307 jmp_rule[i] = mach->rule();
1308 #endif
1309 block->map_node(replacement, j);
1310 mach->subsume_by(replacement, C);
1311 n = replacement;
1312 mach = replacement;
1313 }
1314 }
1315 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1316 } else if (mach->ideal_Opcode() == Op_Jump) {
1317 for (uint h = 0; h < block->_num_succs; h++) {
1318 Block* succs_block = block->_succs[h];
1319 for (uint j = 1; j < succs_block->num_preds(); j++) {
1320 Node* jpn = succs_block->pred(j);
1321 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1322 uint block_num = succs_block->non_connector()->_pre_order;
1323 Label *blkLabel = &blk_labels[block_num];
1324 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1325 }
1326 }
1327 }
1328 }
1329 #ifdef ASSERT
1330 // Check that oop-store precedes the card-mark
1331 else if (mach->ideal_Opcode() == Op_StoreCM) {
1332 uint storeCM_idx = j;
1333 int count = 0;
1334 for (uint prec = mach->req(); prec < mach->len(); prec++) {
1335 Node *oop_store = mach->in(prec); // Precedence edge
1336 if (oop_store == NULL) continue;
1337 count++;
1338 uint i4;
1339 for (i4 = 0; i4 < last_inst; ++i4) {
1340 if (block->get_node(i4) == oop_store) {
1341 break;
1342 }
1343 }
1344 // Note: This test can provide a false failure if other precedence
1345 // edges have been added to the storeCMNode.
1346 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1347 }
1348 assert(count > 0, "storeCM expects at least one precedence edge");
1349 }
1350 #endif
1351 else if (!n->is_Proj()) {
1352 // Remember the beginning of the previous instruction, in case
1353 // it's followed by a flag-kill and a null-check. Happens on
1354 // Intel all the time, with add-to-memory kind of opcodes.
1355 previous_offset = current_offset;
1356 }
1357
1358 // Not an else-if!
1359 // If this is a trap based cmp then add its offset to the list.
1360 if (mach->is_TrapBasedCheckNode()) {
1361 inct_starts[inct_cnt++] = current_offset;
1362 }
1363 }
1364
1365 // Verify that there is sufficient space remaining
1366 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1367 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1368 C->record_failure("CodeCache is full");
1369 return;
1370 }
1371
1372 // Save the offset for the listing
1373 #ifndef PRODUCT
1374 if (node_offsets && n->_idx < node_offset_limit)
1375 node_offsets[n->_idx] = cb->insts_size();
1376 #endif
1377
1378 // "Normal" instruction case
1379 DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
1380 n->emit(*cb, _regalloc);
1381 current_offset = cb->insts_size();
1382
1383 // Above we only verified that there is enough space in the instruction section.
1384 // However, the instruction may emit stubs that cause code buffer expansion.
1385 // Bail out here if expansion failed due to a lack of code cache space.
1386 if (failing()) {
1387 return;
1388 }
1389
1390 #ifdef ASSERT
1391 if (n->size(_regalloc) < (current_offset-instr_offset)) {
1392 n->dump();
1393 assert(false, "wrong size of mach node");
1394 }
1395 #endif
1396 non_safepoints.observe_instruction(n, current_offset);
1397
1398 // mcall is last "call" that can be a safepoint
1399 // record it so we can see if a poll will directly follow it
1400 // in which case we'll need a pad to make the PcDesc sites unique
1401 // see 5010568. This can be slightly inaccurate but conservative
1402 // in the case that return address is not actually at current_offset.
1403 // This is a small price to pay.
1404
1405 if (is_mcall) {
1406 last_call_offset = current_offset;
1407 }
1408
1409 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1410 // Avoid back to back some instructions.
1411 last_avoid_back_to_back_offset = current_offset;
1412 }
1413
1414 // See if this instruction has a delay slot
1415 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1416 assert(delay_slot != NULL, "expecting delay slot node");
1417
1418 // Back up 1 instruction
1419 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1420
1421 // Save the offset for the listing
1422 #ifndef PRODUCT
1423 if (node_offsets && delay_slot->_idx < node_offset_limit)
1424 node_offsets[delay_slot->_idx] = cb->insts_size();
1425 #endif
1426
1427 // Support a SafePoint in the delay slot
1428 if (delay_slot->is_MachSafePoint()) {
1429 MachNode *mach = delay_slot->as_Mach();
1430 // !!!!! Stubs only need an oopmap right now, so bail out
1431 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
1432 // Write the oopmap directly to the code blob??!!
1433 delay_slot = NULL;
1434 continue;
1435 }
1436
1437 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1438 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1439 adjusted_offset);
1440 // Generate an OopMap entry
1441 Process_OopMap_Node(mach, adjusted_offset);
1442 }
1443
1444 // Insert the delay slot instruction
1445 delay_slot->emit(*cb, _regalloc);
1446
1447 // Don't reuse it
1448 delay_slot = NULL;
1449 }
1450
1451 } // End for all instructions in block
1452
1453 // If the next block is the top of a loop, pad this block out to align
1454 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1455 if (i < nblocks-1) {
1456 Block *nb = _cfg->get_block(i + 1);
1457 int padding = nb->alignment_padding(current_offset);
1458 if( padding > 0 ) {
1459 MachNode *nop = new MachNopNode(padding / nop_size);
1460 block->insert_node(nop, block->number_of_nodes());
1461 _cfg->map_node_to_block(nop, block);
1462 nop->emit(*cb, _regalloc);
1463 current_offset = cb->insts_size();
1464 }
1465 }
1466 // Verify that the distance for generated before forward
1467 // short branches is still valid.
1468 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1469
1470 // Save new block start offset
1471 blk_starts[i] = blk_offset;
1472 } // End of for all blocks
1473 blk_starts[nblocks] = current_offset;
1474
1475 non_safepoints.flush_at_end();
1476
1477 // Offset too large?
1478 if (failing()) return;
1479
1480 // Define a pseudo-label at the end of the code
1481 MacroAssembler(cb).bind( blk_labels[nblocks] );
1482
1483 // Compute the size of the first block
1484 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1485
1486 assert(cb->insts_size() < 500000, "method is unreasonably large");
1487
1488 #ifdef ASSERT
1489 for (uint i = 0; i < nblocks; i++) { // For all blocks
1490 if (jmp_target[i] != 0) {
1491 int br_size = jmp_size[i];
1492 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1493 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1494 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]);
1495 assert(false, "Displacement too large for short jmp");
1496 }
1497 }
1498 }
1499 #endif
1500
1501 #ifndef PRODUCT
1502 // Information on the size of the method, without the extraneous code
1503 Scheduling::increment_method_size(cb->insts_size());
1504 #endif
1505
1506 // ------------------
1507 // Fill in exception table entries.
1508 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1509
1510 // Only java methods have exception handlers and deopt handlers
1511 // class HandlerImpl is platform-specific and defined in the *.ad files.
1512 if (_method) {
1513 // Emit the exception handler code.
1514 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb));
1515 if (failing()) {
1516 return; // CodeBuffer::expand failed
1517 }
1518 // Emit the deopt handler code.
1519 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb));
1520
1521 // Emit the MethodHandle deopt handler code (if required).
1522 if (has_method_handle_invokes() && !failing()) {
1523 // We can use the same code as for the normal deopt handler, we
1524 // just need a different entry point address.
1525 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb));
1526 }
1527 }
1528
1529 // One last check for failed CodeBuffer::expand:
1530 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1531 C->record_failure("CodeCache is full");
1532 return;
1533 }
1534
1535 #ifndef PRODUCT
1536 // Dump the assembly code, including basic-block numbers
1537 if (print_assembly()) {
1538 ttyLocker ttyl; // keep the following output all in one block
1539 if (!VMThread::should_terminate()) { // test this under the tty lock
1540 // This output goes directly to the tty, not the compiler log.
1541 // To enable tools to match it up with the compilation activity,
1542 // be sure to tag this tty output with the compile ID.
1543 if (xtty != NULL) {
1544 xtty->head("opto_assembly compile_id='%d'%s", compile_id(),
1545 is_osr_compilation() ? " compile_kind='osr'" :
1546 "");
1547 }
1548 if (method() != NULL) {
1549 method()->print_metadata();
1550 }
1551 dump_asm(node_offsets, node_offset_limit);
1552 if (xtty != NULL) {
1553 xtty->tail("opto_assembly");
1554 }
1555 }
1556 }
1557 #endif
1558
1559 }
1560
1561 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1562 _inc_table.set_size(cnt);
1563
1564 uint inct_cnt = 0;
1565 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1566 Block* block = _cfg->get_block(i);
1567 Node *n = NULL;
1568 int j;
1569
1570 // Find the branch; ignore trailing NOPs.
1571 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1572 n = block->get_node(j);
1573 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1574 break;
1575 }
1576 }
1577
1578 // If we didn't find anything, continue
1579 if (j < 0) {
1580 continue;
1581 }
1582
1583 // Compute ExceptionHandlerTable subtable entry and add it
1584 // (skip empty blocks)
1585 if (n->is_Catch()) {
1586
1587 // Get the offset of the return from the call
1588 uint call_return = call_returns[block->_pre_order];
1589 #ifdef ASSERT
1590 assert( call_return > 0, "no call seen for this basic block" );
1591 while (block->get_node(--j)->is_MachProj()) ;
1592 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1593 #endif
1594 // last instruction is a CatchNode, find it's CatchProjNodes
1595 int nof_succs = block->_num_succs;
1596 // allocate space
1597 GrowableArray<intptr_t> handler_bcis(nof_succs);
1598 GrowableArray<intptr_t> handler_pcos(nof_succs);
1599 // iterate through all successors
1600 for (int j = 0; j < nof_succs; j++) {
1601 Block* s = block->_succs[j];
1602 bool found_p = false;
1603 for (uint k = 1; k < s->num_preds(); k++) {
1604 Node* pk = s->pred(k);
1605 if (pk->is_CatchProj() && pk->in(0) == n) {
1606 const CatchProjNode* p = pk->as_CatchProj();
1607 found_p = true;
1608 // add the corresponding handler bci & pco information
1609 if (p->_con != CatchProjNode::fall_through_index) {
1610 // p leads to an exception handler (and is not fall through)
1611 assert(s == _cfg->get_block(s->_pre_order), "bad numbering");
1612 // no duplicates, please
1613 if (!handler_bcis.contains(p->handler_bci())) {
1614 uint block_num = s->non_connector()->_pre_order;
1615 handler_bcis.append(p->handler_bci());
1616 handler_pcos.append(blk_labels[block_num].loc_pos());
1617 }
1618 }
1619 }
1620 }
1621 assert(found_p, "no matching predecessor found");
1622 // Note: Due to empty block removal, one block may have
1623 // several CatchProj inputs, from the same Catch.
1624 }
1625
1626 // Set the offset of the return from the call
1627 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
1628 continue;
1629 }
1630
1631 // Handle implicit null exception table updates
1632 if (n->is_MachNullCheck()) {
1633 uint block_num = block->non_connector_successor(0)->_pre_order;
1634 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1635 continue;
1636 }
1637 // Handle implicit exception table updates: trap instructions.
1638 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
1639 uint block_num = block->non_connector_successor(0)->_pre_order;
1640 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1641 continue;
1642 }
1643 } // End of for all blocks fill in exception table entries
1644 }
1645
1646 // Static Variables
1647 #ifndef PRODUCT
1648 uint Scheduling::_total_nop_size = 0;
1649 uint Scheduling::_total_method_size = 0;
1650 uint Scheduling::_total_branches = 0;
1651 uint Scheduling::_total_unconditional_delays = 0;
1652 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1653 #endif
1654
1655 // Initializer for class Scheduling
1656
1657 Scheduling::Scheduling(Arena *arena, Compile &compile)
1658 : _arena(arena),
1659 _cfg(compile.cfg()),
1660 _regalloc(compile.regalloc()),
1661 _reg_node(arena),
1662 _bundle_instr_count(0),
1663 _bundle_cycle_number(0),
1664 _scheduled(arena),
1665 _available(arena),
1666 _next_node(NULL),
1667 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]),
1668 _pinch_free_list(arena)
1669 #ifndef PRODUCT
1670 , _branches(0)
1671 , _unconditional_delays(0)
1672 #endif
1673 {
1674 // Create a MachNopNode
1675 _nop = new MachNopNode();
1676
1677 // Now that the nops are in the array, save the count
1678 // (but allow entries for the nops)
1679 _node_bundling_limit = compile.unique();
1680 uint node_max = _regalloc->node_regs_max_index();
1681
1682 compile.set_node_bundling_limit(_node_bundling_limit);
1683
1684 // This one is persistent within the Compile class
1685 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1686
1687 // Allocate space for fixed-size arrays
1688 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1689 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
1690 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1691
1692 // Clear the arrays
1693 memset(_node_bundling_base, 0, node_max * sizeof(Bundle));
1694 memset(_node_latency, 0, node_max * sizeof(unsigned short));
1695 memset(_uses, 0, node_max * sizeof(short));
1696 memset(_current_latency, 0, node_max * sizeof(unsigned short));
1697
1698 // Clear the bundling information
1699 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
1700
1701 // Get the last node
1702 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
1703
1704 _next_node = block->get_node(block->number_of_nodes() - 1);
1705 }
1706
1707 #ifndef PRODUCT
1708 // Scheduling destructor
1709 Scheduling::~Scheduling() {
1710 _total_branches += _branches;
1711 _total_unconditional_delays += _unconditional_delays;
1712 }
1713 #endif
1714
1715 // Step ahead "i" cycles
1716 void Scheduling::step(uint i) {
1717
1718 Bundle *bundle = node_bundling(_next_node);
1719 bundle->set_starts_bundle();
1720
1721 // Update the bundle record, but leave the flags information alone
1722 if (_bundle_instr_count > 0) {
1723 bundle->set_instr_count(_bundle_instr_count);
1724 bundle->set_resources_used(_bundle_use.resourcesUsed());
1725 }
1726
1727 // Update the state information
1728 _bundle_instr_count = 0;
1729 _bundle_cycle_number += i;
1730 _bundle_use.step(i);
1731 }
1732
1733 void Scheduling::step_and_clear() {
1734 Bundle *bundle = node_bundling(_next_node);
1735 bundle->set_starts_bundle();
1736
1737 // Update the bundle record
1738 if (_bundle_instr_count > 0) {
1739 bundle->set_instr_count(_bundle_instr_count);
1740 bundle->set_resources_used(_bundle_use.resourcesUsed());
1741
1742 _bundle_cycle_number += 1;
1743 }
1744
1745 // Clear the bundling information
1746 _bundle_instr_count = 0;
1747 _bundle_use.reset();
1748
1749 memcpy(_bundle_use_elements,
1750 Pipeline_Use::elaborated_elements,
1751 sizeof(Pipeline_Use::elaborated_elements));
1752 }
1753
1754 // Perform instruction scheduling and bundling over the sequence of
1755 // instructions in backwards order.
1756 void Compile::ScheduleAndBundle() {
1757
1758 // Don't optimize this if it isn't a method
1759 if (!_method)
1760 return;
1761
1762 // Don't optimize this if scheduling is disabled
1763 if (!do_scheduling())
1764 return;
1765
1766 // Scheduling code works only with pairs (16 bytes) maximum.
1767 if (max_vector_size() > 16)
1768 return;
1769
1770 TracePhase tp("isched", &timers[_t_instrSched]);
1771
1772 // Create a data structure for all the scheduling information
1773 Scheduling scheduling(Thread::current()->resource_area(), *this);
1774
1775 // Walk backwards over each basic block, computing the needed alignment
1776 // Walk over all the basic blocks
1777 scheduling.DoScheduling();
1778 }
1779
1780 // Compute the latency of all the instructions. This is fairly simple,
1781 // because we already have a legal ordering. Walk over the instructions
1782 // from first to last, and compute the latency of the instruction based
1783 // on the latency of the preceding instruction(s).
1784 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
1785 #ifndef PRODUCT
1786 if (_cfg->C->trace_opto_output())
1787 tty->print("# -> ComputeLocalLatenciesForward\n");
1788 #endif
1789
1790 // Walk over all the schedulable instructions
1791 for( uint j=_bb_start; j < _bb_end; j++ ) {
1792
1793 // This is a kludge, forcing all latency calculations to start at 1.
1794 // Used to allow latency 0 to force an instruction to the beginning
1795 // of the bb
1796 uint latency = 1;
1797 Node *use = bb->get_node(j);
1798 uint nlen = use->len();
1799
1800 // Walk over all the inputs
1801 for ( uint k=0; k < nlen; k++ ) {
1802 Node *def = use->in(k);
1803 if (!def)
1804 continue;
1805
1806 uint l = _node_latency[def->_idx] + use->latency(k);
1807 if (latency < l)
1808 latency = l;
1809 }
1810
1811 _node_latency[use->_idx] = latency;
1812
1813 #ifndef PRODUCT
1814 if (_cfg->C->trace_opto_output()) {
1815 tty->print("# latency %4d: ", latency);
1816 use->dump();
1817 }
1818 #endif
1819 }
1820
1821 #ifndef PRODUCT
1822 if (_cfg->C->trace_opto_output())
1823 tty->print("# <- ComputeLocalLatenciesForward\n");
1824 #endif
1825
1826 } // end ComputeLocalLatenciesForward
1827
1828 // See if this node fits into the present instruction bundle
1829 bool Scheduling::NodeFitsInBundle(Node *n) {
1830 uint n_idx = n->_idx;
1831
1832 // If this is the unconditional delay instruction, then it fits
1833 if (n == _unconditional_delay_slot) {
1834 #ifndef PRODUCT
1835 if (_cfg->C->trace_opto_output())
1836 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
1837 #endif
1838 return (true);
1839 }
1840
1841 // If the node cannot be scheduled this cycle, skip it
1842 if (_current_latency[n_idx] > _bundle_cycle_number) {
1843 #ifndef PRODUCT
1844 if (_cfg->C->trace_opto_output())
1845 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
1846 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
1847 #endif
1848 return (false);
1849 }
1850
1851 const Pipeline *node_pipeline = n->pipeline();
1852
1853 uint instruction_count = node_pipeline->instructionCount();
1854 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
1855 instruction_count = 0;
1856 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
1857 instruction_count++;
1858
1859 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
1860 #ifndef PRODUCT
1861 if (_cfg->C->trace_opto_output())
1862 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
1863 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
1864 #endif
1865 return (false);
1866 }
1867
1868 // Don't allow non-machine nodes to be handled this way
1869 if (!n->is_Mach() && instruction_count == 0)
1870 return (false);
1871
1872 // See if there is any overlap
1873 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
1874
1875 if (delay > 0) {
1876 #ifndef PRODUCT
1877 if (_cfg->C->trace_opto_output())
1878 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
1879 #endif
1880 return false;
1881 }
1882
1883 #ifndef PRODUCT
1884 if (_cfg->C->trace_opto_output())
1885 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
1886 #endif
1887
1888 return true;
1889 }
1890
1891 Node * Scheduling::ChooseNodeToBundle() {
1892 uint siz = _available.size();
1893
1894 if (siz == 0) {
1895
1896 #ifndef PRODUCT
1897 if (_cfg->C->trace_opto_output())
1898 tty->print("# ChooseNodeToBundle: NULL\n");
1899 #endif
1900 return (NULL);
1901 }
1902
1903 // Fast path, if only 1 instruction in the bundle
1904 if (siz == 1) {
1905 #ifndef PRODUCT
1906 if (_cfg->C->trace_opto_output()) {
1907 tty->print("# ChooseNodeToBundle (only 1): ");
1908 _available[0]->dump();
1909 }
1910 #endif
1911 return (_available[0]);
1912 }
1913
1914 // Don't bother, if the bundle is already full
1915 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
1916 for ( uint i = 0; i < siz; i++ ) {
1917 Node *n = _available[i];
1918
1919 // Skip projections, we'll handle them another way
1920 if (n->is_Proj())
1921 continue;
1922
1923 // This presupposed that instructions are inserted into the
1924 // available list in a legality order; i.e. instructions that
1925 // must be inserted first are at the head of the list
1926 if (NodeFitsInBundle(n)) {
1927 #ifndef PRODUCT
1928 if (_cfg->C->trace_opto_output()) {
1929 tty->print("# ChooseNodeToBundle: ");
1930 n->dump();
1931 }
1932 #endif
1933 return (n);
1934 }
1935 }
1936 }
1937
1938 // Nothing fits in this bundle, choose the highest priority
1939 #ifndef PRODUCT
1940 if (_cfg->C->trace_opto_output()) {
1941 tty->print("# ChooseNodeToBundle: ");
1942 _available[0]->dump();
1943 }
1944 #endif
1945
1946 return _available[0];
1947 }
1948
1949 void Scheduling::AddNodeToAvailableList(Node *n) {
1950 assert( !n->is_Proj(), "projections never directly made available" );
1951 #ifndef PRODUCT
1952 if (_cfg->C->trace_opto_output()) {
1953 tty->print("# AddNodeToAvailableList: ");
1954 n->dump();
1955 }
1956 #endif
1957
1958 int latency = _current_latency[n->_idx];
1959
1960 // Insert in latency order (insertion sort)
1961 uint i;
1962 for ( i=0; i < _available.size(); i++ )
1963 if (_current_latency[_available[i]->_idx] > latency)
1964 break;
1965
1966 // Special Check for compares following branches
1967 if( n->is_Mach() && _scheduled.size() > 0 ) {
1968 int op = n->as_Mach()->ideal_Opcode();
1969 Node *last = _scheduled[0];
1970 if( last->is_MachIf() && last->in(1) == n &&
1971 ( op == Op_CmpI ||
1972 op == Op_CmpU ||
1973 op == Op_CmpP ||
1974 op == Op_CmpF ||
1975 op == Op_CmpD ||
1976 op == Op_CmpL ) ) {
1977
1978 // Recalculate position, moving to front of same latency
1979 for ( i=0 ; i < _available.size(); i++ )
1980 if (_current_latency[_available[i]->_idx] >= latency)
1981 break;
1982 }
1983 }
1984
1985 // Insert the node in the available list
1986 _available.insert(i, n);
1987
1988 #ifndef PRODUCT
1989 if (_cfg->C->trace_opto_output())
1990 dump_available();
1991 #endif
1992 }
1993
1994 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
1995 for ( uint i=0; i < n->len(); i++ ) {
1996 Node *def = n->in(i);
1997 if (!def) continue;
1998 if( def->is_Proj() ) // If this is a machine projection, then
1999 def = def->in(0); // propagate usage thru to the base instruction
2000
2001 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2002 continue;
2003 }
2004
2005 // Compute the latency
2006 uint l = _bundle_cycle_number + n->latency(i);
2007 if (_current_latency[def->_idx] < l)
2008 _current_latency[def->_idx] = l;
2009
2010 // If this does not have uses then schedule it
2011 if ((--_uses[def->_idx]) == 0)
2012 AddNodeToAvailableList(def);
2013 }
2014 }
2015
2016 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2017 #ifndef PRODUCT
2018 if (_cfg->C->trace_opto_output()) {
2019 tty->print("# AddNodeToBundle: ");
2020 n->dump();
2021 }
2022 #endif
2023
2024 // Remove this from the available list
2025 uint i;
2026 for (i = 0; i < _available.size(); i++)
2027 if (_available[i] == n)
2028 break;
2029 assert(i < _available.size(), "entry in _available list not found");
2030 _available.remove(i);
2031
2032 // See if this fits in the current bundle
2033 const Pipeline *node_pipeline = n->pipeline();
2034 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2035
2036 // Check for instructions to be placed in the delay slot. We
2037 // do this before we actually schedule the current instruction,
2038 // because the delay slot follows the current instruction.
2039 if (Pipeline::_branch_has_delay_slot &&
2040 node_pipeline->hasBranchDelay() &&
2041 !_unconditional_delay_slot) {
2042
2043 uint siz = _available.size();
2044
2045 // Conditional branches can support an instruction that
2046 // is unconditionally executed and not dependent by the
2047 // branch, OR a conditionally executed instruction if
2048 // the branch is taken. In practice, this means that
2049 // the first instruction at the branch target is
2050 // copied to the delay slot, and the branch goes to
2051 // the instruction after that at the branch target
2052 if ( n->is_MachBranch() ) {
2053
2054 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2055 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2056
2057 #ifndef PRODUCT
2058 _branches++;
2059 #endif
2060
2061 // At least 1 instruction is on the available list
2062 // that is not dependent on the branch
2063 for (uint i = 0; i < siz; i++) {
2064 Node *d = _available[i];
2065 const Pipeline *avail_pipeline = d->pipeline();
2066
2067 // Don't allow safepoints in the branch shadow, that will
2068 // cause a number of difficulties
2069 if ( avail_pipeline->instructionCount() == 1 &&
2070 !avail_pipeline->hasMultipleBundles() &&
2071 !avail_pipeline->hasBranchDelay() &&
2072 Pipeline::instr_has_unit_size() &&
2073 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2074 NodeFitsInBundle(d) &&
2075 !node_bundling(d)->used_in_delay()) {
2076
2077 if (d->is_Mach() && !d->is_MachSafePoint()) {
2078 // A node that fits in the delay slot was found, so we need to
2079 // set the appropriate bits in the bundle pipeline information so
2080 // that it correctly indicates resource usage. Later, when we
2081 // attempt to add this instruction to the bundle, we will skip
2082 // setting the resource usage.
2083 _unconditional_delay_slot = d;
2084 node_bundling(n)->set_use_unconditional_delay();
2085 node_bundling(d)->set_used_in_unconditional_delay();
2086 _bundle_use.add_usage(avail_pipeline->resourceUse());
2087 _current_latency[d->_idx] = _bundle_cycle_number;
2088 _next_node = d;
2089 ++_bundle_instr_count;
2090 #ifndef PRODUCT
2091 _unconditional_delays++;
2092 #endif
2093 break;
2094 }
2095 }
2096 }
2097 }
2098
2099 // No delay slot, add a nop to the usage
2100 if (!_unconditional_delay_slot) {
2101 // See if adding an instruction in the delay slot will overflow
2102 // the bundle.
2103 if (!NodeFitsInBundle(_nop)) {
2104 #ifndef PRODUCT
2105 if (_cfg->C->trace_opto_output())
2106 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2107 #endif
2108 step(1);
2109 }
2110
2111 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2112 _next_node = _nop;
2113 ++_bundle_instr_count;
2114 }
2115
2116 // See if the instruction in the delay slot requires a
2117 // step of the bundles
2118 if (!NodeFitsInBundle(n)) {
2119 #ifndef PRODUCT
2120 if (_cfg->C->trace_opto_output())
2121 tty->print("# *** STEP(branch won't fit) ***\n");
2122 #endif
2123 // Update the state information
2124 _bundle_instr_count = 0;
2125 _bundle_cycle_number += 1;
2126 _bundle_use.step(1);
2127 }
2128 }
2129
2130 // Get the number of instructions
2131 uint instruction_count = node_pipeline->instructionCount();
2132 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2133 instruction_count = 0;
2134
2135 // Compute the latency information
2136 uint delay = 0;
2137
2138 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2139 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2140 if (relative_latency < 0)
2141 relative_latency = 0;
2142
2143 delay = _bundle_use.full_latency(relative_latency, node_usage);
2144
2145 // Does not fit in this bundle, start a new one
2146 if (delay > 0) {
2147 step(delay);
2148
2149 #ifndef PRODUCT
2150 if (_cfg->C->trace_opto_output())
2151 tty->print("# *** STEP(%d) ***\n", delay);
2152 #endif
2153 }
2154 }
2155
2156 // If this was placed in the delay slot, ignore it
2157 if (n != _unconditional_delay_slot) {
2158
2159 if (delay == 0) {
2160 if (node_pipeline->hasMultipleBundles()) {
2161 #ifndef PRODUCT
2162 if (_cfg->C->trace_opto_output())
2163 tty->print("# *** STEP(multiple instructions) ***\n");
2164 #endif
2165 step(1);
2166 }
2167
2168 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2169 #ifndef PRODUCT
2170 if (_cfg->C->trace_opto_output())
2171 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2172 instruction_count + _bundle_instr_count,
2173 Pipeline::_max_instrs_per_cycle);
2174 #endif
2175 step(1);
2176 }
2177 }
2178
2179 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2180 _bundle_instr_count++;
2181
2182 // Set the node's latency
2183 _current_latency[n->_idx] = _bundle_cycle_number;
2184
2185 // Now merge the functional unit information
2186 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2187 _bundle_use.add_usage(node_usage);
2188
2189 // Increment the number of instructions in this bundle
2190 _bundle_instr_count += instruction_count;
2191
2192 // Remember this node for later
2193 if (n->is_Mach())
2194 _next_node = n;
2195 }
2196
2197 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2198 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2199 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2200 // into the block. All other scheduled nodes get put in the schedule here.
2201 int op = n->Opcode();
2202 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2203 (op != Op_Node && // Not an unused antidepedence node and
2204 // not an unallocated boxlock
2205 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2206
2207 // Push any trailing projections
2208 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2209 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2210 Node *foi = n->fast_out(i);
2211 if( foi->is_Proj() )
2212 _scheduled.push(foi);
2213 }
2214 }
2215
2216 // Put the instruction in the schedule list
2217 _scheduled.push(n);
2218 }
2219
2220 #ifndef PRODUCT
2221 if (_cfg->C->trace_opto_output())
2222 dump_available();
2223 #endif
2224
2225 // Walk all the definitions, decrementing use counts, and
2226 // if a definition has a 0 use count, place it in the available list.
2227 DecrementUseCounts(n,bb);
2228 }
2229
2230 // This method sets the use count within a basic block. We will ignore all
2231 // uses outside the current basic block. As we are doing a backwards walk,
2232 // any node we reach that has a use count of 0 may be scheduled. This also
2233 // avoids the problem of cyclic references from phi nodes, as long as phi
2234 // nodes are at the front of the basic block. This method also initializes
2235 // the available list to the set of instructions that have no uses within this
2236 // basic block.
2237 void Scheduling::ComputeUseCount(const Block *bb) {
2238 #ifndef PRODUCT
2239 if (_cfg->C->trace_opto_output())
2240 tty->print("# -> ComputeUseCount\n");
2241 #endif
2242
2243 // Clear the list of available and scheduled instructions, just in case
2244 _available.clear();
2245 _scheduled.clear();
2246
2247 // No delay slot specified
2248 _unconditional_delay_slot = NULL;
2249
2250 #ifdef ASSERT
2251 for( uint i=0; i < bb->number_of_nodes(); i++ )
2252 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2253 #endif
2254
2255 // Force the _uses count to never go to zero for unscheduable pieces
2256 // of the block
2257 for( uint k = 0; k < _bb_start; k++ )
2258 _uses[bb->get_node(k)->_idx] = 1;
2259 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2260 _uses[bb->get_node(l)->_idx] = 1;
2261
2262 // Iterate backwards over the instructions in the block. Don't count the
2263 // branch projections at end or the block header instructions.
2264 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2265 Node *n = bb->get_node(j);
2266 if( n->is_Proj() ) continue; // Projections handled another way
2267
2268 // Account for all uses
2269 for ( uint k = 0; k < n->len(); k++ ) {
2270 Node *inp = n->in(k);
2271 if (!inp) continue;
2272 assert(inp != n, "no cycles allowed" );
2273 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2274 if (inp->is_Proj()) { // Skip through Proj's
2275 inp = inp->in(0);
2276 }
2277 ++_uses[inp->_idx]; // Count 1 block-local use
2278 }
2279 }
2280
2281 // If this instruction has a 0 use count, then it is available
2282 if (!_uses[n->_idx]) {
2283 _current_latency[n->_idx] = _bundle_cycle_number;
2284 AddNodeToAvailableList(n);
2285 }
2286
2287 #ifndef PRODUCT
2288 if (_cfg->C->trace_opto_output()) {
2289 tty->print("# uses: %3d: ", _uses[n->_idx]);
2290 n->dump();
2291 }
2292 #endif
2293 }
2294
2295 #ifndef PRODUCT
2296 if (_cfg->C->trace_opto_output())
2297 tty->print("# <- ComputeUseCount\n");
2298 #endif
2299 }
2300
2301 // This routine performs scheduling on each basic block in reverse order,
2302 // using instruction latencies and taking into account function unit
2303 // availability.
2304 void Scheduling::DoScheduling() {
2305 #ifndef PRODUCT
2306 if (_cfg->C->trace_opto_output())
2307 tty->print("# -> DoScheduling\n");
2308 #endif
2309
2310 Block *succ_bb = NULL;
2311 Block *bb;
2312
2313 // Walk over all the basic blocks in reverse order
2314 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2315 bb = _cfg->get_block(i);
2316
2317 #ifndef PRODUCT
2318 if (_cfg->C->trace_opto_output()) {
2319 tty->print("# Schedule BB#%03d (initial)\n", i);
2320 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2321 bb->get_node(j)->dump();
2322 }
2323 }
2324 #endif
2325
2326 // On the head node, skip processing
2327 if (bb == _cfg->get_root_block()) {
2328 continue;
2329 }
2330
2331 // Skip empty, connector blocks
2332 if (bb->is_connector())
2333 continue;
2334
2335 // If the following block is not the sole successor of
2336 // this one, then reset the pipeline information
2337 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2338 #ifndef PRODUCT
2339 if (_cfg->C->trace_opto_output()) {
2340 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2341 _next_node->_idx, _bundle_instr_count);
2342 }
2343 #endif
2344 step_and_clear();
2345 }
2346
2347 // Leave untouched the starting instruction, any Phis, a CreateEx node
2348 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2349 _bb_end = bb->number_of_nodes()-1;
2350 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2351 Node *n = bb->get_node(_bb_start);
2352 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2353 // Also, MachIdealNodes do not get scheduled
2354 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2355 MachNode *mach = n->as_Mach();
2356 int iop = mach->ideal_Opcode();
2357 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2358 if( iop == Op_Con ) continue; // Do not schedule Top
2359 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2360 mach->pipeline() == MachNode::pipeline_class() &&
2361 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
2362 continue;
2363 break; // Funny loop structure to be sure...
2364 }
2365 // Compute last "interesting" instruction in block - last instruction we
2366 // might schedule. _bb_end points just after last schedulable inst. We
2367 // normally schedule conditional branches (despite them being forced last
2368 // in the block), because they have delay slots we can fill. Calls all
2369 // have their delay slots filled in the template expansions, so we don't
2370 // bother scheduling them.
2371 Node *last = bb->get_node(_bb_end);
2372 // Ignore trailing NOPs.
2373 while (_bb_end > 0 && last->is_Mach() &&
2374 last->as_Mach()->ideal_Opcode() == Op_Con) {
2375 last = bb->get_node(--_bb_end);
2376 }
2377 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2378 if( last->is_Catch() ||
2379 // Exclude unreachable path case when Halt node is in a separate block.
2380 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2381 // There must be a prior call. Skip it.
2382 while( !bb->get_node(--_bb_end)->is_MachCall() ) {
2383 assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" );
2384 }
2385 } else if( last->is_MachNullCheck() ) {
2386 // Backup so the last null-checked memory instruction is
2387 // outside the schedulable range. Skip over the nullcheck,
2388 // projection, and the memory nodes.
2389 Node *mem = last->in(1);
2390 do {
2391 _bb_end--;
2392 } while (mem != bb->get_node(_bb_end));
2393 } else {
2394 // Set _bb_end to point after last schedulable inst.
2395 _bb_end++;
2396 }
2397
2398 assert( _bb_start <= _bb_end, "inverted block ends" );
2399
2400 // Compute the register antidependencies for the basic block
2401 ComputeRegisterAntidependencies(bb);
2402 if (_cfg->C->failing()) return; // too many D-U pinch points
2403
2404 // Compute intra-bb latencies for the nodes
2405 ComputeLocalLatenciesForward(bb);
2406
2407 // Compute the usage within the block, and set the list of all nodes
2408 // in the block that have no uses within the block.
2409 ComputeUseCount(bb);
2410
2411 // Schedule the remaining instructions in the block
2412 while ( _available.size() > 0 ) {
2413 Node *n = ChooseNodeToBundle();
2414 guarantee(n != NULL, "no nodes available");
2415 AddNodeToBundle(n,bb);
2416 }
2417
2418 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2419 #ifdef ASSERT
2420 for( uint l = _bb_start; l < _bb_end; l++ ) {
2421 Node *n = bb->get_node(l);
2422 uint m;
2423 for( m = 0; m < _bb_end-_bb_start; m++ )
2424 if( _scheduled[m] == n )
2425 break;
2426 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2427 }
2428 #endif
2429
2430 // Now copy the instructions (in reverse order) back to the block
2431 for ( uint k = _bb_start; k < _bb_end; k++ )
2432 bb->map_node(_scheduled[_bb_end-k-1], k);
2433
2434 #ifndef PRODUCT
2435 if (_cfg->C->trace_opto_output()) {
2436 tty->print("# Schedule BB#%03d (final)\n", i);
2437 uint current = 0;
2438 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2439 Node *n = bb->get_node(j);
2440 if( valid_bundle_info(n) ) {
2441 Bundle *bundle = node_bundling(n);
2442 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2443 tty->print("*** Bundle: ");
2444 bundle->dump();
2445 }
2446 n->dump();
2447 }
2448 }
2449 }
2450 #endif
2451 #ifdef ASSERT
2452 verify_good_schedule(bb,"after block local scheduling");
2453 #endif
2454 }
2455
2456 #ifndef PRODUCT
2457 if (_cfg->C->trace_opto_output())
2458 tty->print("# <- DoScheduling\n");
2459 #endif
2460
2461 // Record final node-bundling array location
2462 _regalloc->C->set_node_bundling_base(_node_bundling_base);
2463
2464 } // end DoScheduling
2465
2466 // Verify that no live-range used in the block is killed in the block by a
2467 // wrong DEF. This doesn't verify live-ranges that span blocks.
2468
2469 // Check for edge existence. Used to avoid adding redundant precedence edges.
2470 static bool edge_from_to( Node *from, Node *to ) {
2471 for( uint i=0; i<from->len(); i++ )
2472 if( from->in(i) == to )
2473 return true;
2474 return false;
2475 }
2476
2477 #ifdef ASSERT
2478 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2479 // Check for bad kills
2480 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2481 Node *prior_use = _reg_node[def];
2482 if( prior_use && !edge_from_to(prior_use,n) ) {
2483 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2484 n->dump();
2485 tty->print_cr("...");
2486 prior_use->dump();
2487 assert(edge_from_to(prior_use,n),msg);
2488 }
2489 _reg_node.map(def,NULL); // Kill live USEs
2490 }
2491 }
2492
2493 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2494
2495 // Zap to something reasonable for the verify code
2496 _reg_node.clear();
2497
2498 // Walk over the block backwards. Check to make sure each DEF doesn't
2499 // kill a live value (other than the one it's supposed to). Add each
2500 // USE to the live set.
2501 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2502 Node *n = b->get_node(i);
2503 int n_op = n->Opcode();
2504 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2505 // Fat-proj kills a slew of registers
2506 RegMask rm = n->out_RegMask();// Make local copy
2507 while( rm.is_NotEmpty() ) {
2508 OptoReg::Name kill = rm.find_first_elem();
2509 rm.Remove(kill);
2510 verify_do_def( n, kill, msg );
2511 }
2512 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2513 // Get DEF'd registers the normal way
2514 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2515 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2516 }
2517
2518 // Now make all USEs live
2519 for( uint i=1; i<n->req(); i++ ) {
2520 Node *def = n->in(i);
2521 assert(def != 0, "input edge required");
2522 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2523 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2524 if( OptoReg::is_valid(reg_lo) ) {
2525 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg);
2526 _reg_node.map(reg_lo,n);
2527 }
2528 if( OptoReg::is_valid(reg_hi) ) {
2529 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg);
2530 _reg_node.map(reg_hi,n);
2531 }
2532 }
2533
2534 }
2535
2536 // Zap to something reasonable for the Antidependence code
2537 _reg_node.clear();
2538 }
2539 #endif
2540
2541 // Conditionally add precedence edges. Avoid putting edges on Projs.
2542 static void add_prec_edge_from_to( Node *from, Node *to ) {
2543 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2544 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2545 from = from->in(0);
2546 }
2547 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2548 !edge_from_to( from, to ) ) // Avoid duplicate edge
2549 from->add_prec(to);
2550 }
2551
2552 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2553 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2554 return;
2555
2556 Node *pinch = _reg_node[def_reg]; // Get pinch point
2557 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2558 is_def ) { // Check for a true def (not a kill)
2559 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2560 return;
2561 }
2562
2563 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2564 debug_only( def = (Node*)0xdeadbeef; )
2565
2566 // After some number of kills there _may_ be a later def
2567 Node *later_def = NULL;
2568
2569 // Finding a kill requires a real pinch-point.
2570 // Check for not already having a pinch-point.
2571 // Pinch points are Op_Node's.
2572 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2573 later_def = pinch; // Must be def/kill as optimistic pinch-point
2574 if ( _pinch_free_list.size() > 0) {
2575 pinch = _pinch_free_list.pop();
2576 } else {
2577 pinch = new Node(1); // Pinch point to-be
2578 }
2579 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2580 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2581 return;
2582 }
2583 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
2584 _reg_node.map(def_reg,pinch); // Record pinch-point
2585 //_regalloc->set_bad(pinch->_idx); // Already initialized this way.
2586 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2587 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call
2588 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2589 later_def = NULL; // and no later def
2590 }
2591 pinch->set_req(0,later_def); // Hook later def so we can find it
2592 } else { // Else have valid pinch point
2593 if( pinch->in(0) ) // If there is a later-def
2594 later_def = pinch->in(0); // Get it
2595 }
2596
2597 // Add output-dependence edge from later def to kill
2598 if( later_def ) // If there is some original def
2599 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2600
2601 // See if current kill is also a use, and so is forced to be the pinch-point.
2602 if( pinch->Opcode() == Op_Node ) {
2603 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2604 for( uint i=1; i<uses->req(); i++ ) {
2605 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2606 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2607 // Yes, found a use/kill pinch-point
2608 pinch->set_req(0,NULL); //
2609 pinch->replace_by(kill); // Move anti-dep edges up
2610 pinch = kill;
2611 _reg_node.map(def_reg,pinch);
2612 return;
2613 }
2614 }
2615 }
2616
2617 // Add edge from kill to pinch-point
2618 add_prec_edge_from_to(kill,pinch);
2619 }
2620
2621 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2622 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2623 return;
2624 Node *pinch = _reg_node[use_reg]; // Get pinch point
2625 // Check for no later def_reg/kill in block
2626 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b &&
2627 // Use has to be block-local as well
2628 _cfg->get_block_for_node(use) == b) {
2629 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2630 pinch->req() == 1 ) { // pinch not yet in block?
2631 pinch->del_req(0); // yank pointer to later-def, also set flag
2632 // Insert the pinch-point in the block just after the last use
2633 b->insert_node(pinch, b->find_node(use) + 1);
2634 _bb_end++; // Increase size scheduled region in block
2635 }
2636
2637 add_prec_edge_from_to(pinch,use);
2638 }
2639 }
2640
2641 // We insert antidependences between the reads and following write of
2642 // allocated registers to prevent illegal code motion. Hopefully, the
2643 // number of added references should be fairly small, especially as we
2644 // are only adding references within the current basic block.
2645 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2646
2647 #ifdef ASSERT
2648 verify_good_schedule(b,"before block local scheduling");
2649 #endif
2650
2651 // A valid schedule, for each register independently, is an endless cycle
2652 // of: a def, then some uses (connected to the def by true dependencies),
2653 // then some kills (defs with no uses), finally the cycle repeats with a new
2654 // def. The uses are allowed to float relative to each other, as are the
2655 // kills. No use is allowed to slide past a kill (or def). This requires
2656 // antidependencies between all uses of a single def and all kills that
2657 // follow, up to the next def. More edges are redundant, because later defs
2658 // & kills are already serialized with true or antidependencies. To keep
2659 // the edge count down, we add a 'pinch point' node if there's more than
2660 // one use or more than one kill/def.
2661
2662 // We add dependencies in one bottom-up pass.
2663
2664 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2665
2666 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2667 // register. If not, we record the DEF/KILL in _reg_node, the
2668 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
2669 // "pinch point", a new Node that's in the graph but not in the block.
2670 // We put edges from the prior and current DEF/KILLs to the pinch point.
2671 // We put the pinch point in _reg_node. If there's already a pinch point
2672 // we merely add an edge from the current DEF/KILL to the pinch point.
2673
2674 // After doing the DEF/KILLs, we handle USEs. For each used register, we
2675 // put an edge from the pinch point to the USE.
2676
2677 // To be expedient, the _reg_node array is pre-allocated for the whole
2678 // compilation. _reg_node is lazily initialized; it either contains a NULL,
2679 // or a valid def/kill/pinch-point, or a leftover node from some prior
2680 // block. Leftover node from some prior block is treated like a NULL (no
2681 // prior def, so no anti-dependence needed). Valid def is distinguished by
2682 // it being in the current block.
2683 bool fat_proj_seen = false;
2684 uint last_safept = _bb_end-1;
2685 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL;
2686 Node* last_safept_node = end_node;
2687 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2688 Node *n = b->get_node(i);
2689 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
2690 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
2691 // Fat-proj kills a slew of registers
2692 // This can add edges to 'n' and obscure whether or not it was a def,
2693 // hence the is_def flag.
2694 fat_proj_seen = true;
2695 RegMask rm = n->out_RegMask();// Make local copy
2696 while( rm.is_NotEmpty() ) {
2697 OptoReg::Name kill = rm.find_first_elem();
2698 rm.Remove(kill);
2699 anti_do_def( b, n, kill, is_def );
2700 }
2701 } else {
2702 // Get DEF'd registers the normal way
2703 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2704 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2705 }
2706
2707 // Kill projections on a branch should appear to occur on the
2708 // branch, not afterwards, so grab the masks from the projections
2709 // and process them.
2710 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) {
2711 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2712 Node* use = n->fast_out(i);
2713 if (use->is_Proj()) {
2714 RegMask rm = use->out_RegMask();// Make local copy
2715 while( rm.is_NotEmpty() ) {
2716 OptoReg::Name kill = rm.find_first_elem();
2717 rm.Remove(kill);
2718 anti_do_def( b, n, kill, false );
2719 }
2720 }
2721 }
2722 }
2723
2724 // Check each register used by this instruction for a following DEF/KILL
2725 // that must occur afterward and requires an anti-dependence edge.
2726 for( uint j=0; j<n->req(); j++ ) {
2727 Node *def = n->in(j);
2728 if( def ) {
2729 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
2730 anti_do_use( b, n, _regalloc->get_reg_first(def) );
2731 anti_do_use( b, n, _regalloc->get_reg_second(def) );
2732 }
2733 }
2734 // Do not allow defs of new derived values to float above GC
2735 // points unless the base is definitely available at the GC point.
2736
2737 Node *m = b->get_node(i);
2738
2739 // Add precedence edge from following safepoint to use of derived pointer
2740 if( last_safept_node != end_node &&
2741 m != last_safept_node) {
2742 for (uint k = 1; k < m->req(); k++) {
2743 const Type *t = m->in(k)->bottom_type();
2744 if( t->isa_oop_ptr() &&
2745 t->is_ptr()->offset() != 0 ) {
2746 last_safept_node->add_prec( m );
2747 break;
2748 }
2749 }
2750 }
2751
2752 if( n->jvms() ) { // Precedence edge from derived to safept
2753 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2754 if( b->get_node(last_safept) != last_safept_node ) {
2755 last_safept = b->find_node(last_safept_node);
2756 }
2757 for( uint j=last_safept; j > i; j-- ) {
2758 Node *mach = b->get_node(j);
2759 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2760 mach->add_prec( n );
2761 }
2762 last_safept = i;
2763 last_safept_node = m;
2764 }
2765 }
2766
2767 if (fat_proj_seen) {
2768 // Garbage collect pinch nodes that were not consumed.
2769 // They are usually created by a fat kill MachProj for a call.
2770 garbage_collect_pinch_nodes();
2771 }
2772 }
2773
2774 // Garbage collect pinch nodes for reuse by other blocks.
2775 //
2776 // The block scheduler's insertion of anti-dependence
2777 // edges creates many pinch nodes when the block contains
2778 // 2 or more Calls. A pinch node is used to prevent a
2779 // combinatorial explosion of edges. If a set of kills for a
2780 // register is anti-dependent on a set of uses (or defs), rather
2781 // than adding an edge in the graph between each pair of kill
2782 // and use (or def), a pinch is inserted between them:
2783 //
2784 // use1 use2 use3
2785 // \ | /
2786 // \ | /
2787 // pinch
2788 // / | \
2789 // / | \
2790 // kill1 kill2 kill3
2791 //
2792 // One pinch node is created per register killed when
2793 // the second call is encountered during a backwards pass
2794 // over the block. Most of these pinch nodes are never
2795 // wired into the graph because the register is never
2796 // used or def'ed in the block.
2797 //
2798 void Scheduling::garbage_collect_pinch_nodes() {
2799 #ifndef PRODUCT
2800 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
2801 #endif
2802 int trace_cnt = 0;
2803 for (uint k = 0; k < _reg_node.Size(); k++) {
2804 Node* pinch = _reg_node[k];
2805 if ((pinch != NULL) && pinch->Opcode() == Op_Node &&
2806 // no predecence input edges
2807 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
2808 cleanup_pinch(pinch);
2809 _pinch_free_list.push(pinch);
2810 _reg_node.map(k, NULL);
2811 #ifndef PRODUCT
2812 if (_cfg->C->trace_opto_output()) {
2813 trace_cnt++;
2814 if (trace_cnt > 40) {
2815 tty->print("\n");
2816 trace_cnt = 0;
2817 }
2818 tty->print(" %d", pinch->_idx);
2819 }
2820 #endif
2821 }
2822 }
2823 #ifndef PRODUCT
2824 if (_cfg->C->trace_opto_output()) tty->print("\n");
2825 #endif
2826 }
2827
2828 // Clean up a pinch node for reuse.
2829 void Scheduling::cleanup_pinch( Node *pinch ) {
2830 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
2831
2832 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
2833 Node* use = pinch->last_out(i);
2834 uint uses_found = 0;
2835 for (uint j = use->req(); j < use->len(); j++) {
2836 if (use->in(j) == pinch) {
2837 use->rm_prec(j);
2838 uses_found++;
2839 }
2840 }
2841 assert(uses_found > 0, "must be a precedence edge");
2842 i -= uses_found; // we deleted 1 or more copies of this edge
2843 }
2844 // May have a later_def entry
2845 pinch->set_req(0, NULL);
2846 }
2847
2848 #ifndef PRODUCT
2849
2850 void Scheduling::dump_available() const {
2851 tty->print("#Availist ");
2852 for (uint i = 0; i < _available.size(); i++)
2853 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
2854 tty->cr();
2855 }
2856
2857 // Print Scheduling Statistics
2858 void Scheduling::print_statistics() {
2859 // Print the size added by nops for bundling
2860 tty->print("Nops added %d bytes to total of %d bytes",
2861 _total_nop_size, _total_method_size);
2862 if (_total_method_size > 0)
2863 tty->print(", for %.2f%%",
2864 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
2865 tty->print("\n");
2866
2867 // Print the number of branch shadows filled
2868 if (Pipeline::_branch_has_delay_slot) {
2869 tty->print("Of %d branches, %d had unconditional delay slots filled",
2870 _total_branches, _total_unconditional_delays);
2871 if (_total_branches > 0)
2872 tty->print(", for %.2f%%",
2873 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
2874 tty->print("\n");
2875 }
2876
2877 uint total_instructions = 0, total_bundles = 0;
2878
2879 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
2880 uint bundle_count = _total_instructions_per_bundle[i];
2881 total_instructions += bundle_count * i;
2882 total_bundles += bundle_count;
2883 }
2884
2885 if (total_bundles > 0)
2886 tty->print("Average ILP (excluding nops) is %.2f\n",
2887 ((double)total_instructions) / ((double)total_bundles));
2888 }
2889 #endif