1 /*
2 * Copyright (c) 1998, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/assembler.inline.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "code/compiledIC.hpp"
29 #include "code/debugInfo.hpp"
30 #include "code/debugInfoRec.hpp"
31 #include "compiler/compileBroker.hpp"
32 #include "compiler/compilerDirectives.hpp"
33 #include "compiler/oopMap.hpp"
34 #include "memory/allocation.inline.hpp"
35 #include "opto/ad.hpp"
36 #include "opto/callnode.hpp"
37 #include "opto/cfgnode.hpp"
38 #include "opto/locknode.hpp"
39 #include "opto/machnode.hpp"
40 #include "opto/optoreg.hpp"
41 #include "opto/output.hpp"
42 #include "opto/regalloc.hpp"
43 #include "opto/runtime.hpp"
44 #include "opto/subnode.hpp"
45 #include "opto/type.hpp"
46 #include "runtime/handles.inline.hpp"
47 #include "utilities/xmlstream.hpp"
48
49 #ifndef PRODUCT
50 #define DEBUG_ARG(x) , x
51 #else
52 #define DEBUG_ARG(x)
53 #endif
54
55 // Convert Nodes to instruction bits and pass off to the VM
56 void Compile::Output() {
57 // RootNode goes
58 assert( _cfg->get_root_block()->number_of_nodes() == 0, "" );
59
60 // The number of new nodes (mostly MachNop) is proportional to
61 // the number of java calls and inner loops which are aligned.
62 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
63 C->inner_loops()*(OptoLoopAlignment-1)),
64 "out of nodes before code generation" ) ) {
65 return;
66 }
67 // Make sure I can find the Start Node
68 Block *entry = _cfg->get_block(1);
69 Block *broot = _cfg->get_root_block();
70
71 const StartNode *start = entry->head()->as_Start();
72
73 // Replace StartNode with prolog
74 Label verified_entry;
75 MachPrologNode* prolog = new MachPrologNode(&verified_entry);
76 entry->map_node(prolog, 0);
77 _cfg->map_node_to_block(prolog, entry);
78 _cfg->unmap_node_from_block(start); // start is no longer in any block
79
80 // Virtual methods need an unverified entry point
81 if (is_osr_compilation()) {
82 if (PoisonOSREntry) {
83 // TODO: Should use a ShouldNotReachHereNode...
84 _cfg->insert( broot, 0, new MachBreakpointNode() );
85 }
86 } else {
87 if (_method) {
88 if (_method->has_scalarized_args()) {
89 // Add entry point to unpack all value type arguments
90 _cfg->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ false));
91 if (!_method->is_static()) {
92 // Add verified/unverified entry points to only unpack value type receiver at interface calls
93 _cfg->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ false));
94 _cfg->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ true));
95 _cfg->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ true));
96 }
97 } else if (!_method->is_static()) {
98 // Insert unvalidated entry point
99 _cfg->insert(broot, 0, new MachUEPNode());
100 }
101 }
102 }
103
104 // Break before main entry point
105 if ((_method && C->directive()->BreakAtExecuteOption) ||
106 (OptoBreakpoint && is_method_compilation()) ||
107 (OptoBreakpointOSR && is_osr_compilation()) ||
108 (OptoBreakpointC2R && !_method) ) {
109 // checking for _method means that OptoBreakpoint does not apply to
110 // runtime stubs or frame converters
111 _cfg->insert( entry, 1, new MachBreakpointNode() );
112 }
113
114 // Insert epilogs before every return
115 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
116 Block* block = _cfg->get_block(i);
117 if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point?
118 Node* m = block->end();
119 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
120 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
121 block->add_inst(epilog);
122 _cfg->map_node_to_block(epilog, block);
123 }
124 }
125 }
126
127 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1);
128 blk_starts[0] = 0;
129
130 // Initialize code buffer and process short branches.
131 CodeBuffer* cb = init_buffer(blk_starts);
132
133 if (cb == NULL || failing()) {
134 return;
135 }
136
137 if (_method && _method->has_scalarized_args()) {
138 // Compute the offsets of the entry points required by the value type calling convention
139 if (!_method->is_static()) {
140 // We have entries at the beginning of the method, implemented by the first 4 nodes.
141 // Entry (unverified) @ offset 0
142 // Verified_Value_Entry_RO
143 // Value_Entry (unverified)
144 // Verified_Value_Entry
145 uint offset = 0;
146 _code_offsets.set_value(CodeOffsets::Entry, offset);
147
148 offset += ((MachVEPNode*)broot->get_node(0))->size(_regalloc);
149 _code_offsets.set_value(CodeOffsets::Verified_Value_Entry_RO, offset);
150
151 offset += ((MachVEPNode*)broot->get_node(1))->size(_regalloc);
152 _code_offsets.set_value(CodeOffsets::Value_Entry, offset);
153
154 offset += ((MachVEPNode*)broot->get_node(2))->size(_regalloc);
155 _code_offsets.set_value(CodeOffsets::Verified_Value_Entry, offset);
156 } else {
157 _code_offsets.set_value(CodeOffsets::Entry, -1); // will be patched later
158 _code_offsets.set_value(CodeOffsets::Verified_Value_Entry, 0);
159 }
160 }
161
162 ScheduleAndBundle();
163
164 #ifndef PRODUCT
165 if (trace_opto_output()) {
166 tty->print("\n---- After ScheduleAndBundle ----\n");
167 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
168 tty->print("\nBB#%03d:\n", i);
169 Block* block = _cfg->get_block(i);
170 for (uint j = 0; j < block->number_of_nodes(); j++) {
171 Node* n = block->get_node(j);
172 OptoReg::Name reg = _regalloc->get_reg_first(n);
173 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
174 n->dump();
175 }
176 }
177 }
178 #endif
179
180 if (failing()) {
181 return;
182 }
183
184 BuildOopMaps();
185
186 if (failing()) {
187 return;
188 }
189
190 fill_buffer(cb, blk_starts);
191 }
192
193 bool Compile::need_stack_bang(int frame_size_in_bytes) const {
194 // Determine if we need to generate a stack overflow check.
195 // Do it if the method is not a stub function and
196 // has java calls or has frame size > vm_page_size/8.
197 // The debug VM checks that deoptimization doesn't trigger an
198 // unexpected stack overflow (compiled method stack banging should
199 // guarantee it doesn't happen) so we always need the stack bang in
200 // a debug VM.
201 return (UseStackBanging && stub_function() == NULL &&
202 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3
203 DEBUG_ONLY(|| true)));
204 }
205
206 bool Compile::need_register_stack_bang() const {
207 // Determine if we need to generate a register stack overflow check.
208 // This is only used on architectures which have split register
209 // and memory stacks (ie. IA64).
210 // Bang if the method is not a stub function and has java calls
211 return (stub_function() == NULL && has_java_calls());
212 }
213
214
215 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
216 // of a loop. When aligning a loop we need to provide enough instructions
217 // in cpu's fetch buffer to feed decoders. The loop alignment could be
218 // avoided if we have enough instructions in fetch buffer at the head of a loop.
219 // By default, the size is set to 999999 by Block's constructor so that
220 // a loop will be aligned if the size is not reset here.
221 //
222 // Note: Mach instructions could contain several HW instructions
223 // so the size is estimated only.
224 //
225 void Compile::compute_loop_first_inst_sizes() {
226 // The next condition is used to gate the loop alignment optimization.
227 // Don't aligned a loop if there are enough instructions at the head of a loop
228 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
229 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
230 // equal to 11 bytes which is the largest address NOP instruction.
231 if (MaxLoopPad < OptoLoopAlignment - 1) {
232 uint last_block = _cfg->number_of_blocks() - 1;
233 for (uint i = 1; i <= last_block; i++) {
234 Block* block = _cfg->get_block(i);
235 // Check the first loop's block which requires an alignment.
236 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
237 uint sum_size = 0;
238 uint inst_cnt = NumberOfLoopInstrToAlign;
239 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
240
241 // Check subsequent fallthrough blocks if the loop's first
242 // block(s) does not have enough instructions.
243 Block *nb = block;
244 while(inst_cnt > 0 &&
245 i < last_block &&
246 !_cfg->get_block(i + 1)->has_loop_alignment() &&
247 !nb->has_successor(block)) {
248 i++;
249 nb = _cfg->get_block(i);
250 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
251 } // while( inst_cnt > 0 && i < last_block )
252
253 block->set_first_inst_size(sum_size);
254 } // f( b->head()->is_Loop() )
255 } // for( i <= last_block )
256 } // if( MaxLoopPad < OptoLoopAlignment-1 )
257 }
258
259 // The architecture description provides short branch variants for some long
260 // branch instructions. Replace eligible long branches with short branches.
261 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) {
262 // Compute size of each block, method size, and relocation information size
263 uint nblocks = _cfg->number_of_blocks();
264
265 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
266 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
267 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
268
269 // Collect worst case block paddings
270 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
271 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
272
273 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
274 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
275
276 bool has_short_branch_candidate = false;
277
278 // Initialize the sizes to 0
279 code_size = 0; // Size in bytes of generated code
280 stub_size = 0; // Size in bytes of all stub entries
281 // Size in bytes of all relocation entries, including those in local stubs.
282 // Start with 2-bytes of reloc info for the unvalidated entry point
283 reloc_size = 1; // Number of relocation entries
284
285 // Make three passes. The first computes pessimistic blk_starts,
286 // relative jmp_offset and reloc_size information. The second performs
287 // short branch substitution using the pessimistic sizing. The
288 // third inserts nops where needed.
289
290 // Step one, perform a pessimistic sizing pass.
291 uint last_call_adr = max_juint;
292 uint last_avoid_back_to_back_adr = max_juint;
293 uint nop_size = (new MachNopNode())->size(_regalloc);
294 for (uint i = 0; i < nblocks; i++) { // For all blocks
295 Block* block = _cfg->get_block(i);
296
297 // During short branch replacement, we store the relative (to blk_starts)
298 // offset of jump in jmp_offset, rather than the absolute offset of jump.
299 // This is so that we do not need to recompute sizes of all nodes when
300 // we compute correct blk_starts in our next sizing pass.
301 jmp_offset[i] = 0;
302 jmp_size[i] = 0;
303 jmp_nidx[i] = -1;
304 DEBUG_ONLY( jmp_target[i] = 0; )
305 DEBUG_ONLY( jmp_rule[i] = 0; )
306
307 // Sum all instruction sizes to compute block size
308 uint last_inst = block->number_of_nodes();
309 uint blk_size = 0;
310 for (uint j = 0; j < last_inst; j++) {
311 Node* nj = block->get_node(j);
312 // Handle machine instruction nodes
313 if (nj->is_Mach()) {
314 MachNode *mach = nj->as_Mach();
315 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
316 reloc_size += mach->reloc();
317 if (mach->is_MachCall()) {
318 // add size information for trampoline stub
319 // class CallStubImpl is platform-specific and defined in the *.ad files.
320 stub_size += CallStubImpl::size_call_trampoline();
321 reloc_size += CallStubImpl::reloc_call_trampoline();
322
323 MachCallNode *mcall = mach->as_MachCall();
324 // This destination address is NOT PC-relative
325
326 if (mcall->entry_point() != NULL) {
327 mcall->method_set((intptr_t)mcall->entry_point());
328 }
329
330 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
331 stub_size += CompiledStaticCall::to_interp_stub_size();
332 reloc_size += CompiledStaticCall::reloc_to_interp_stub();
333 #if INCLUDE_AOT
334 stub_size += CompiledStaticCall::to_aot_stub_size();
335 reloc_size += CompiledStaticCall::reloc_to_aot_stub();
336 #endif
337 }
338 } else if (mach->is_MachSafePoint()) {
339 // If call/safepoint are adjacent, account for possible
340 // nop to disambiguate the two safepoints.
341 // ScheduleAndBundle() can rearrange nodes in a block,
342 // check for all offsets inside this block.
343 if (last_call_adr >= blk_starts[i]) {
344 blk_size += nop_size;
345 }
346 }
347 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
348 // Nop is inserted between "avoid back to back" instructions.
349 // ScheduleAndBundle() can rearrange nodes in a block,
350 // check for all offsets inside this block.
351 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
352 blk_size += nop_size;
353 }
354 }
355 if (mach->may_be_short_branch()) {
356 if (!nj->is_MachBranch()) {
357 #ifndef PRODUCT
358 nj->dump(3);
359 #endif
360 Unimplemented();
361 }
362 assert(jmp_nidx[i] == -1, "block should have only one branch");
363 jmp_offset[i] = blk_size;
364 jmp_size[i] = nj->size(_regalloc);
365 jmp_nidx[i] = j;
366 has_short_branch_candidate = true;
367 }
368 }
369 blk_size += nj->size(_regalloc);
370 // Remember end of call offset
371 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
372 last_call_adr = blk_starts[i]+blk_size;
373 }
374 // Remember end of avoid_back_to_back offset
375 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
376 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
377 }
378 }
379
380 // When the next block starts a loop, we may insert pad NOP
381 // instructions. Since we cannot know our future alignment,
382 // assume the worst.
383 if (i < nblocks - 1) {
384 Block* nb = _cfg->get_block(i + 1);
385 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
386 if (max_loop_pad > 0) {
387 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
388 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
389 // If either is the last instruction in this block, bump by
390 // max_loop_pad in lock-step with blk_size, so sizing
391 // calculations in subsequent blocks still can conservatively
392 // detect that it may the last instruction in this block.
393 if (last_call_adr == blk_starts[i]+blk_size) {
394 last_call_adr += max_loop_pad;
395 }
396 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
397 last_avoid_back_to_back_adr += max_loop_pad;
398 }
399 blk_size += max_loop_pad;
400 block_worst_case_pad[i + 1] = max_loop_pad;
401 }
402 }
403
404 // Save block size; update total method size
405 blk_starts[i+1] = blk_starts[i]+blk_size;
406 }
407
408 // Step two, replace eligible long jumps.
409 bool progress = true;
410 uint last_may_be_short_branch_adr = max_juint;
411 while (has_short_branch_candidate && progress) {
412 progress = false;
413 has_short_branch_candidate = false;
414 int adjust_block_start = 0;
415 for (uint i = 0; i < nblocks; i++) {
416 Block* block = _cfg->get_block(i);
417 int idx = jmp_nidx[i];
418 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach();
419 if (mach != NULL && mach->may_be_short_branch()) {
420 #ifdef ASSERT
421 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
422 int j;
423 // Find the branch; ignore trailing NOPs.
424 for (j = block->number_of_nodes()-1; j>=0; j--) {
425 Node* n = block->get_node(j);
426 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
427 break;
428 }
429 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
430 #endif
431 int br_size = jmp_size[i];
432 int br_offs = blk_starts[i] + jmp_offset[i];
433
434 // This requires the TRUE branch target be in succs[0]
435 uint bnum = block->non_connector_successor(0)->_pre_order;
436 int offset = blk_starts[bnum] - br_offs;
437 if (bnum > i) { // adjust following block's offset
438 offset -= adjust_block_start;
439 }
440
441 // This block can be a loop header, account for the padding
442 // in the previous block.
443 int block_padding = block_worst_case_pad[i];
444 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
445 // In the following code a nop could be inserted before
446 // the branch which will increase the backward distance.
447 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
448 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
449
450 if (needs_padding && offset <= 0)
451 offset -= nop_size;
452
453 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
454 // We've got a winner. Replace this branch.
455 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
456
457 // Update the jmp_size.
458 int new_size = replacement->size(_regalloc);
459 int diff = br_size - new_size;
460 assert(diff >= (int)nop_size, "short_branch size should be smaller");
461 // Conservatively take into account padding between
462 // avoid_back_to_back branches. Previous branch could be
463 // converted into avoid_back_to_back branch during next
464 // rounds.
465 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
466 jmp_offset[i] += nop_size;
467 diff -= nop_size;
468 }
469 adjust_block_start += diff;
470 block->map_node(replacement, idx);
471 mach->subsume_by(replacement, C);
472 mach = replacement;
473 progress = true;
474
475 jmp_size[i] = new_size;
476 DEBUG_ONLY( jmp_target[i] = bnum; );
477 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
478 } else {
479 // The jump distance is not short, try again during next iteration.
480 has_short_branch_candidate = true;
481 }
482 } // (mach->may_be_short_branch())
483 if (mach != NULL && (mach->may_be_short_branch() ||
484 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
485 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
486 }
487 blk_starts[i+1] -= adjust_block_start;
488 }
489 }
490
491 #ifdef ASSERT
492 for (uint i = 0; i < nblocks; i++) { // For all blocks
493 if (jmp_target[i] != 0) {
494 int br_size = jmp_size[i];
495 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
496 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
497 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]);
498 }
499 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
500 }
501 }
502 #endif
503
504 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
505 // after ScheduleAndBundle().
506
507 // ------------------
508 // Compute size for code buffer
509 code_size = blk_starts[nblocks];
510
511 // Relocation records
512 reloc_size += 1; // Relo entry for exception handler
513
514 // Adjust reloc_size to number of record of relocation info
515 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
516 // a relocation index.
517 // The CodeBuffer will expand the locs array if this estimate is too low.
518 reloc_size *= 10 / sizeof(relocInfo);
519 }
520
521 //------------------------------FillLocArray-----------------------------------
522 // Create a bit of debug info and append it to the array. The mapping is from
523 // Java local or expression stack to constant, register or stack-slot. For
524 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
525 // entry has been taken care of and caller should skip it).
526 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
527 // This should never have accepted Bad before
528 assert(OptoReg::is_valid(regnum), "location must be valid");
529 return (OptoReg::is_reg(regnum))
530 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
531 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
532 }
533
534
535 ObjectValue*
536 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
537 for (int i = 0; i < objs->length(); i++) {
538 assert(objs->at(i)->is_object(), "corrupt object cache");
539 ObjectValue* sv = (ObjectValue*) objs->at(i);
540 if (sv->id() == id) {
541 return sv;
542 }
543 }
544 // Otherwise..
545 return NULL;
546 }
547
548 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
549 ObjectValue* sv ) {
550 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
551 objs->append(sv);
552 }
553
554
555 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
556 GrowableArray<ScopeValue*> *array,
557 GrowableArray<ScopeValue*> *objs ) {
558 assert( local, "use _top instead of null" );
559 if (array->length() != idx) {
560 assert(array->length() == idx + 1, "Unexpected array count");
561 // Old functionality:
562 // return
563 // New functionality:
564 // Assert if the local is not top. In product mode let the new node
565 // override the old entry.
566 assert(local == top(), "LocArray collision");
567 if (local == top()) {
568 return;
569 }
570 array->pop();
571 }
572 const Type *t = local->bottom_type();
573
574 // Is it a safepoint scalar object node?
575 if (local->is_SafePointScalarObject()) {
576 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
577
578 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
579 if (sv == NULL) {
580 ciKlass* cik = t->is_oopptr()->klass();
581 assert(cik->is_instance_klass() ||
582 cik->is_array_klass(), "Not supported allocation.");
583 sv = new ObjectValue(spobj->_idx,
584 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
585 Compile::set_sv_for_object_node(objs, sv);
586
587 uint first_ind = spobj->first_index(sfpt->jvms());
588 for (uint i = 0; i < spobj->n_fields(); i++) {
589 Node* fld_node = sfpt->in(first_ind+i);
590 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
591 }
592 }
593 array->append(sv);
594 return;
595 }
596
597 // Grab the register number for the local
598 OptoReg::Name regnum = _regalloc->get_reg_first(local);
599 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
600 // Record the double as two float registers.
601 // The register mask for such a value always specifies two adjacent
602 // float registers, with the lower register number even.
603 // Normally, the allocation of high and low words to these registers
604 // is irrelevant, because nearly all operations on register pairs
605 // (e.g., StoreD) treat them as a single unit.
606 // Here, we assume in addition that the words in these two registers
607 // stored "naturally" (by operations like StoreD and double stores
608 // within the interpreter) such that the lower-numbered register
609 // is written to the lower memory address. This may seem like
610 // a machine dependency, but it is not--it is a requirement on
611 // the author of the <arch>.ad file to ensure that, for every
612 // even/odd double-register pair to which a double may be allocated,
613 // the word in the even single-register is stored to the first
614 // memory word. (Note that register numbers are completely
615 // arbitrary, and are not tied to any machine-level encodings.)
616 #ifdef _LP64
617 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
618 array->append(new ConstantIntValue((jint)0));
619 array->append(new_loc_value( _regalloc, regnum, Location::dbl ));
620 } else if ( t->base() == Type::Long ) {
621 array->append(new ConstantIntValue((jint)0));
622 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
623 } else if ( t->base() == Type::RawPtr ) {
624 // jsr/ret return address which must be restored into a the full
625 // width 64-bit stack slot.
626 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
627 }
628 #else //_LP64
629 #ifdef SPARC
630 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {
631 // For SPARC we have to swap high and low words for
632 // long values stored in a single-register (g0-g7).
633 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
634 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
635 } else
636 #endif //SPARC
637 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
638 // Repack the double/long as two jints.
639 // The convention the interpreter uses is that the second local
640 // holds the first raw word of the native double representation.
641 // This is actually reasonable, since locals and stack arrays
642 // grow downwards in all implementations.
643 // (If, on some machine, the interpreter's Java locals or stack
644 // were to grow upwards, the embedded doubles would be word-swapped.)
645 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
646 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
647 }
648 #endif //_LP64
649 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
650 OptoReg::is_reg(regnum) ) {
651 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()
652 ? Location::float_in_dbl : Location::normal ));
653 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
654 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long
655 ? Location::int_in_long : Location::normal ));
656 } else if( t->base() == Type::NarrowOop ) {
657 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));
658 } else {
659 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));
660 }
661 return;
662 }
663
664 // No register. It must be constant data.
665 switch (t->base()) {
666 case Type::Half: // Second half of a double
667 ShouldNotReachHere(); // Caller should skip 2nd halves
668 break;
669 case Type::AnyPtr:
670 array->append(new ConstantOopWriteValue(NULL));
671 break;
672 case Type::AryPtr:
673 case Type::InstPtr: // fall through
674 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
675 break;
676 case Type::NarrowOop:
677 if (t == TypeNarrowOop::NULL_PTR) {
678 array->append(new ConstantOopWriteValue(NULL));
679 } else {
680 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
681 }
682 break;
683 case Type::Int:
684 array->append(new ConstantIntValue(t->is_int()->get_con()));
685 break;
686 case Type::RawPtr:
687 // A return address (T_ADDRESS).
688 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
689 #ifdef _LP64
690 // Must be restored to the full-width 64-bit stack slot.
691 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
692 #else
693 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
694 #endif
695 break;
696 case Type::FloatCon: {
697 float f = t->is_float_constant()->getf();
698 array->append(new ConstantIntValue(jint_cast(f)));
699 break;
700 }
701 case Type::DoubleCon: {
702 jdouble d = t->is_double_constant()->getd();
703 #ifdef _LP64
704 array->append(new ConstantIntValue((jint)0));
705 array->append(new ConstantDoubleValue(d));
706 #else
707 // Repack the double as two jints.
708 // The convention the interpreter uses is that the second local
709 // holds the first raw word of the native double representation.
710 // This is actually reasonable, since locals and stack arrays
711 // grow downwards in all implementations.
712 // (If, on some machine, the interpreter's Java locals or stack
713 // were to grow upwards, the embedded doubles would be word-swapped.)
714 jlong_accessor acc;
715 acc.long_value = jlong_cast(d);
716 array->append(new ConstantIntValue(acc.words[1]));
717 array->append(new ConstantIntValue(acc.words[0]));
718 #endif
719 break;
720 }
721 case Type::Long: {
722 jlong d = t->is_long()->get_con();
723 #ifdef _LP64
724 array->append(new ConstantIntValue((jint)0));
725 array->append(new ConstantLongValue(d));
726 #else
727 // Repack the long as two jints.
728 // The convention the interpreter uses is that the second local
729 // holds the first raw word of the native double representation.
730 // This is actually reasonable, since locals and stack arrays
731 // grow downwards in all implementations.
732 // (If, on some machine, the interpreter's Java locals or stack
733 // were to grow upwards, the embedded doubles would be word-swapped.)
734 jlong_accessor acc;
735 acc.long_value = d;
736 array->append(new ConstantIntValue(acc.words[1]));
737 array->append(new ConstantIntValue(acc.words[0]));
738 #endif
739 break;
740 }
741 case Type::Top: // Add an illegal value here
742 array->append(new LocationValue(Location()));
743 break;
744 default:
745 ShouldNotReachHere();
746 break;
747 }
748 }
749
750 // Determine if this node starts a bundle
751 bool Compile::starts_bundle(const Node *n) const {
752 return (_node_bundling_limit > n->_idx &&
753 _node_bundling_base[n->_idx].starts_bundle());
754 }
755
756 //--------------------------Process_OopMap_Node--------------------------------
757 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
758
759 // Handle special safepoint nodes for synchronization
760 MachSafePointNode *sfn = mach->as_MachSafePoint();
761 MachCallNode *mcall;
762
763 int safepoint_pc_offset = current_offset;
764 bool is_method_handle_invoke = false;
765 bool return_oop = false;
766 bool return_vt = false;
767
768 // Add the safepoint in the DebugInfoRecorder
769 if( !mach->is_MachCall() ) {
770 mcall = NULL;
771 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
772 } else {
773 mcall = mach->as_MachCall();
774
775 // Is the call a MethodHandle call?
776 if (mcall->is_MachCallJava()) {
777 if (mcall->as_MachCallJava()->_method_handle_invoke) {
778 assert(has_method_handle_invokes(), "must have been set during call generation");
779 is_method_handle_invoke = true;
780 }
781 }
782
783 // Check if a call returns an object.
784 if (mcall->returns_pointer() || mcall->returns_vt()) {
785 return_oop = true;
786 }
787 if (mcall->returns_vt()) {
788 return_vt = true;
789 }
790 safepoint_pc_offset += mcall->ret_addr_offset();
791 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
792 }
793
794 // Loop over the JVMState list to add scope information
795 // Do not skip safepoints with a NULL method, they need monitor info
796 JVMState* youngest_jvms = sfn->jvms();
797 int max_depth = youngest_jvms->depth();
798
799 // Allocate the object pool for scalar-replaced objects -- the map from
800 // small-integer keys (which can be recorded in the local and ostack
801 // arrays) to descriptions of the object state.
802 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
803
804 // Visit scopes from oldest to youngest.
805 for (int depth = 1; depth <= max_depth; depth++) {
806 JVMState* jvms = youngest_jvms->of_depth(depth);
807 int idx;
808 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
809 // Safepoints that do not have method() set only provide oop-map and monitor info
810 // to support GC; these do not support deoptimization.
811 int num_locs = (method == NULL) ? 0 : jvms->loc_size();
812 int num_exps = (method == NULL) ? 0 : jvms->stk_size();
813 int num_mon = jvms->nof_monitors();
814 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
815 "JVMS local count must match that of the method");
816
817 // Add Local and Expression Stack Information
818
819 // Insert locals into the locarray
820 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
821 for( idx = 0; idx < num_locs; idx++ ) {
822 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
823 }
824
825 // Insert expression stack entries into the exparray
826 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
827 for( idx = 0; idx < num_exps; idx++ ) {
828 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
829 }
830
831 // Add in mappings of the monitors
832 assert( !method ||
833 !method->is_synchronized() ||
834 method->is_native() ||
835 num_mon > 0 ||
836 !GenerateSynchronizationCode,
837 "monitors must always exist for synchronized methods");
838
839 // Build the growable array of ScopeValues for exp stack
840 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
841
842 // Loop over monitors and insert into array
843 for (idx = 0; idx < num_mon; idx++) {
844 // Grab the node that defines this monitor
845 Node* box_node = sfn->monitor_box(jvms, idx);
846 Node* obj_node = sfn->monitor_obj(jvms, idx);
847
848 // Create ScopeValue for object
849 ScopeValue *scval = NULL;
850
851 if (obj_node->is_SafePointScalarObject()) {
852 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
853 scval = Compile::sv_for_node_id(objs, spobj->_idx);
854 if (scval == NULL) {
855 const Type *t = spobj->bottom_type();
856 ciKlass* cik = t->is_oopptr()->klass();
857 assert(cik->is_instance_klass() ||
858 cik->is_array_klass(), "Not supported allocation.");
859 ObjectValue* sv = new ObjectValue(spobj->_idx,
860 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
861 Compile::set_sv_for_object_node(objs, sv);
862
863 uint first_ind = spobj->first_index(youngest_jvms);
864 for (uint i = 0; i < spobj->n_fields(); i++) {
865 Node* fld_node = sfn->in(first_ind+i);
866 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
867 }
868 scval = sv;
869 }
870 } else if (!obj_node->is_Con()) {
871 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);
872 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
873 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );
874 } else {
875 scval = new_loc_value( _regalloc, obj_reg, Location::oop );
876 }
877 } else {
878 const TypePtr *tp = obj_node->get_ptr_type();
879 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
880 }
881
882 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
883 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
884 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
885 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
886 }
887
888 // We dump the object pool first, since deoptimization reads it in first.
889 debug_info()->dump_object_pool(objs);
890
891 // Build first class objects to pass to scope
892 DebugToken *locvals = debug_info()->create_scope_values(locarray);
893 DebugToken *expvals = debug_info()->create_scope_values(exparray);
894 DebugToken *monvals = debug_info()->create_monitor_values(monarray);
895
896 // Make method available for all Safepoints
897 ciMethod* scope_method = method ? method : _method;
898 // Describe the scope here
899 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
900 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
901 // Now we can describe the scope.
902 methodHandle null_mh;
903 bool rethrow_exception = false;
904 debug_info()->describe_scope(safepoint_pc_offset, null_mh, scope_method, jvms->bci(), jvms->should_reexecute(), rethrow_exception, is_method_handle_invoke, return_oop, return_vt, locvals, expvals, monvals);
905 } // End jvms loop
906
907 // Mark the end of the scope set.
908 debug_info()->end_safepoint(safepoint_pc_offset);
909 }
910
911
912
913 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
914 class NonSafepointEmitter {
915 Compile* C;
916 JVMState* _pending_jvms;
917 int _pending_offset;
918
919 void emit_non_safepoint();
920
921 public:
922 NonSafepointEmitter(Compile* compile) {
923 this->C = compile;
924 _pending_jvms = NULL;
925 _pending_offset = 0;
926 }
927
928 void observe_instruction(Node* n, int pc_offset) {
929 if (!C->debug_info()->recording_non_safepoints()) return;
930
931 Node_Notes* nn = C->node_notes_at(n->_idx);
932 if (nn == NULL || nn->jvms() == NULL) return;
933 if (_pending_jvms != NULL &&
934 _pending_jvms->same_calls_as(nn->jvms())) {
935 // Repeated JVMS? Stretch it up here.
936 _pending_offset = pc_offset;
937 } else {
938 if (_pending_jvms != NULL &&
939 _pending_offset < pc_offset) {
940 emit_non_safepoint();
941 }
942 _pending_jvms = NULL;
943 if (pc_offset > C->debug_info()->last_pc_offset()) {
944 // This is the only way _pending_jvms can become non-NULL:
945 _pending_jvms = nn->jvms();
946 _pending_offset = pc_offset;
947 }
948 }
949 }
950
951 // Stay out of the way of real safepoints:
952 void observe_safepoint(JVMState* jvms, int pc_offset) {
953 if (_pending_jvms != NULL &&
954 !_pending_jvms->same_calls_as(jvms) &&
955 _pending_offset < pc_offset) {
956 emit_non_safepoint();
957 }
958 _pending_jvms = NULL;
959 }
960
961 void flush_at_end() {
962 if (_pending_jvms != NULL) {
963 emit_non_safepoint();
964 }
965 _pending_jvms = NULL;
966 }
967 };
968
969 void NonSafepointEmitter::emit_non_safepoint() {
970 JVMState* youngest_jvms = _pending_jvms;
971 int pc_offset = _pending_offset;
972
973 // Clear it now:
974 _pending_jvms = NULL;
975
976 DebugInformationRecorder* debug_info = C->debug_info();
977 assert(debug_info->recording_non_safepoints(), "sanity");
978
979 debug_info->add_non_safepoint(pc_offset);
980 int max_depth = youngest_jvms->depth();
981
982 // Visit scopes from oldest to youngest.
983 for (int depth = 1; depth <= max_depth; depth++) {
984 JVMState* jvms = youngest_jvms->of_depth(depth);
985 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
986 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
987 methodHandle null_mh;
988 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
989 }
990
991 // Mark the end of the scope set.
992 debug_info->end_non_safepoint(pc_offset);
993 }
994
995 //------------------------------init_buffer------------------------------------
996 CodeBuffer* Compile::init_buffer(uint* blk_starts) {
997
998 // Set the initially allocated size
999 int code_req = initial_code_capacity;
1000 int locs_req = initial_locs_capacity;
1001 int stub_req = initial_stub_capacity;
1002 int const_req = initial_const_capacity;
1003
1004 int pad_req = NativeCall::instruction_size;
1005 // The extra spacing after the code is necessary on some platforms.
1006 // Sometimes we need to patch in a jump after the last instruction,
1007 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1008
1009 // Compute the byte offset where we can store the deopt pc.
1010 if (fixed_slots() != 0) {
1011 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1012 }
1013 if (C->needs_stack_repair()) {
1014 // Compute the byte offset of the stack increment value
1015 _sp_inc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_sp_inc_slot));
1016 }
1017
1018 // Compute prolog code size
1019 _method_size = 0;
1020 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;
1021 #if defined(IA64) && !defined(AIX)
1022 if (save_argument_registers()) {
1023 // 4815101: this is a stub with implicit and unknown precision fp args.
1024 // The usual spill mechanism can only generate stfd's in this case, which
1025 // doesn't work if the fp reg to spill contains a single-precision denorm.
1026 // Instead, we hack around the normal spill mechanism using stfspill's and
1027 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate
1028 // space here for the fp arg regs (f8-f15) we're going to thusly spill.
1029 //
1030 // If we ever implement 16-byte 'registers' == stack slots, we can
1031 // get rid of this hack and have SpillCopy generate stfspill/ldffill
1032 // instead of stfd/stfs/ldfd/ldfs.
1033 _frame_slots += 8*(16/BytesPerInt);
1034 }
1035 #endif
1036 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1037
1038 if (has_mach_constant_base_node()) {
1039 uint add_size = 0;
1040 // Fill the constant table.
1041 // Note: This must happen before shorten_branches.
1042 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1043 Block* b = _cfg->get_block(i);
1044
1045 for (uint j = 0; j < b->number_of_nodes(); j++) {
1046 Node* n = b->get_node(j);
1047
1048 // If the node is a MachConstantNode evaluate the constant
1049 // value section.
1050 if (n->is_MachConstant()) {
1051 MachConstantNode* machcon = n->as_MachConstant();
1052 machcon->eval_constant(C);
1053 } else if (n->is_Mach()) {
1054 // On Power there are more nodes that issue constants.
1055 add_size += (n->as_Mach()->ins_num_consts() * 8);
1056 }
1057 }
1058 }
1059
1060 // Calculate the offsets of the constants and the size of the
1061 // constant table (including the padding to the next section).
1062 constant_table().calculate_offsets_and_size();
1063 const_req = constant_table().size() + add_size;
1064 }
1065
1066 // Initialize the space for the BufferBlob used to find and verify
1067 // instruction size in MachNode::emit_size()
1068 init_scratch_buffer_blob(const_req);
1069 if (failing()) return NULL; // Out of memory
1070
1071 // Pre-compute the length of blocks and replace
1072 // long branches with short if machine supports it.
1073 shorten_branches(blk_starts, code_req, locs_req, stub_req);
1074
1075 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1076 // class HandlerImpl is platform-specific and defined in the *.ad files.
1077 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1078 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1079 stub_req += MAX_stubs_size; // ensure per-stub margin
1080 code_req += MAX_inst_size; // ensure per-instruction margin
1081
1082 if (StressCodeBuffers)
1083 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1084
1085 int total_req =
1086 const_req +
1087 code_req +
1088 pad_req +
1089 stub_req +
1090 exception_handler_req +
1091 deopt_handler_req; // deopt handler
1092
1093 if (has_method_handle_invokes())
1094 total_req += deopt_handler_req; // deopt MH handler
1095
1096 CodeBuffer* cb = code_buffer();
1097 cb->initialize(total_req, locs_req);
1098
1099 // Have we run out of code space?
1100 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1101 C->record_failure("CodeCache is full");
1102 return NULL;
1103 }
1104 // Configure the code buffer.
1105 cb->initialize_consts_size(const_req);
1106 cb->initialize_stubs_size(stub_req);
1107 cb->initialize_oop_recorder(env()->oop_recorder());
1108
1109 // fill in the nop array for bundling computations
1110 MachNode *_nop_list[Bundle::_nop_count];
1111 Bundle::initialize_nops(_nop_list);
1112
1113 return cb;
1114 }
1115
1116 //------------------------------fill_buffer------------------------------------
1117 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1118 // blk_starts[] contains offsets calculated during short branches processing,
1119 // offsets should not be increased during following steps.
1120
1121 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1122 // of a loop. It is used to determine the padding for loop alignment.
1123 compute_loop_first_inst_sizes();
1124
1125 // Create oopmap set.
1126 _oop_map_set = new OopMapSet();
1127
1128 // !!!!! This preserves old handling of oopmaps for now
1129 debug_info()->set_oopmaps(_oop_map_set);
1130
1131 uint nblocks = _cfg->number_of_blocks();
1132 // Count and start of implicit null check instructions
1133 uint inct_cnt = 0;
1134 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1135
1136 // Count and start of calls
1137 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1138
1139 uint return_offset = 0;
1140 int nop_size = (new MachNopNode())->size(_regalloc);
1141
1142 int previous_offset = 0;
1143 int current_offset = 0;
1144 int last_call_offset = -1;
1145 int last_avoid_back_to_back_offset = -1;
1146 #ifdef ASSERT
1147 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1148 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1149 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1150 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1151 #endif
1152
1153 // Create an array of unused labels, one for each basic block, if printing is enabled
1154 #ifndef PRODUCT
1155 int *node_offsets = NULL;
1156 uint node_offset_limit = unique();
1157
1158 if (print_assembly())
1159 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1160 #endif
1161
1162 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily
1163
1164 // Emit the constant table.
1165 if (has_mach_constant_base_node()) {
1166 constant_table().emit(*cb);
1167 }
1168
1169 // Create an array of labels, one for each basic block
1170 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1171 for (uint i=0; i <= nblocks; i++) {
1172 blk_labels[i].init();
1173 }
1174
1175 // ------------------
1176 // Now fill in the code buffer
1177 Node *delay_slot = NULL;
1178
1179 for (uint i = 0; i < nblocks; i++) {
1180 Block* block = _cfg->get_block(i);
1181 Node* head = block->head();
1182
1183 // If this block needs to start aligned (i.e, can be reached other
1184 // than by falling-thru from the previous block), then force the
1185 // start of a new bundle.
1186 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1187 cb->flush_bundle(true);
1188 }
1189
1190 #ifdef ASSERT
1191 if (!block->is_connector()) {
1192 stringStream st;
1193 block->dump_head(_cfg, &st);
1194 MacroAssembler(cb).block_comment(st.as_string());
1195 }
1196 jmp_target[i] = 0;
1197 jmp_offset[i] = 0;
1198 jmp_size[i] = 0;
1199 jmp_rule[i] = 0;
1200 #endif
1201 int blk_offset = current_offset;
1202
1203 // Define the label at the beginning of the basic block
1204 MacroAssembler(cb).bind(blk_labels[block->_pre_order]);
1205
1206 uint last_inst = block->number_of_nodes();
1207
1208 // Emit block normally, except for last instruction.
1209 // Emit means "dump code bits into code buffer".
1210 for (uint j = 0; j<last_inst; j++) {
1211
1212 // Get the node
1213 Node* n = block->get_node(j);
1214
1215 // See if delay slots are supported
1216 if (valid_bundle_info(n) &&
1217 node_bundling(n)->used_in_unconditional_delay()) {
1218 assert(delay_slot == NULL, "no use of delay slot node");
1219 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1220
1221 delay_slot = n;
1222 continue;
1223 }
1224
1225 // If this starts a new instruction group, then flush the current one
1226 // (but allow split bundles)
1227 if (Pipeline::requires_bundling() && starts_bundle(n))
1228 cb->flush_bundle(false);
1229
1230 // Special handling for SafePoint/Call Nodes
1231 bool is_mcall = false;
1232 if (n->is_Mach()) {
1233 MachNode *mach = n->as_Mach();
1234 is_mcall = n->is_MachCall();
1235 bool is_sfn = n->is_MachSafePoint();
1236
1237 // If this requires all previous instructions be flushed, then do so
1238 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1239 cb->flush_bundle(true);
1240 current_offset = cb->insts_size();
1241 }
1242
1243 // A padding may be needed again since a previous instruction
1244 // could be moved to delay slot.
1245
1246 // align the instruction if necessary
1247 int padding = mach->compute_padding(current_offset);
1248 // Make sure safepoint node for polling is distinct from a call's
1249 // return by adding a nop if needed.
1250 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1251 padding = nop_size;
1252 }
1253 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1254 current_offset == last_avoid_back_to_back_offset) {
1255 // Avoid back to back some instructions.
1256 padding = nop_size;
1257 }
1258
1259 if (padding > 0) {
1260 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1261 int nops_cnt = padding / nop_size;
1262 MachNode *nop = new MachNopNode(nops_cnt);
1263 block->insert_node(nop, j++);
1264 last_inst++;
1265 _cfg->map_node_to_block(nop, block);
1266 // Ensure enough space.
1267 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1268 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1269 C->record_failure("CodeCache is full");
1270 return;
1271 }
1272 nop->emit(*cb, _regalloc);
1273 cb->flush_bundle(true);
1274 current_offset = cb->insts_size();
1275 }
1276
1277 // Remember the start of the last call in a basic block
1278 if (is_mcall) {
1279 MachCallNode *mcall = mach->as_MachCall();
1280
1281 if (mcall->entry_point() != NULL) {
1282 // This destination address is NOT PC-relative
1283 mcall->method_set((intptr_t)mcall->entry_point());
1284 }
1285
1286 // Save the return address
1287 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1288
1289 if (mcall->is_MachCallLeaf()) {
1290 is_mcall = false;
1291 is_sfn = false;
1292 }
1293 }
1294
1295 // sfn will be valid whenever mcall is valid now because of inheritance
1296 if (is_sfn || is_mcall) {
1297
1298 // Handle special safepoint nodes for synchronization
1299 if (!is_mcall) {
1300 MachSafePointNode *sfn = mach->as_MachSafePoint();
1301 // !!!!! Stubs only need an oopmap right now, so bail out
1302 if (sfn->jvms()->method() == NULL) {
1303 // Write the oopmap directly to the code blob??!!
1304 continue;
1305 }
1306 } // End synchronization
1307
1308 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1309 current_offset);
1310 Process_OopMap_Node(mach, current_offset);
1311 } // End if safepoint
1312
1313 // If this is a null check, then add the start of the previous instruction to the list
1314 else if( mach->is_MachNullCheck() ) {
1315 inct_starts[inct_cnt++] = previous_offset;
1316 }
1317
1318 // If this is a branch, then fill in the label with the target BB's label
1319 else if (mach->is_MachBranch()) {
1320 // This requires the TRUE branch target be in succs[0]
1321 uint block_num = block->non_connector_successor(0)->_pre_order;
1322
1323 // Try to replace long branch if delay slot is not used,
1324 // it is mostly for back branches since forward branch's
1325 // distance is not updated yet.
1326 bool delay_slot_is_used = valid_bundle_info(n) &&
1327 node_bundling(n)->use_unconditional_delay();
1328 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1329 assert(delay_slot == NULL, "not expecting delay slot node");
1330 int br_size = n->size(_regalloc);
1331 int offset = blk_starts[block_num] - current_offset;
1332 if (block_num >= i) {
1333 // Current and following block's offset are not
1334 // finalized yet, adjust distance by the difference
1335 // between calculated and final offsets of current block.
1336 offset -= (blk_starts[i] - blk_offset);
1337 }
1338 // In the following code a nop could be inserted before
1339 // the branch which will increase the backward distance.
1340 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1341 if (needs_padding && offset <= 0)
1342 offset -= nop_size;
1343
1344 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
1345 // We've got a winner. Replace this branch.
1346 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1347
1348 // Update the jmp_size.
1349 int new_size = replacement->size(_regalloc);
1350 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1351 // Insert padding between avoid_back_to_back branches.
1352 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1353 MachNode *nop = new MachNopNode();
1354 block->insert_node(nop, j++);
1355 _cfg->map_node_to_block(nop, block);
1356 last_inst++;
1357 nop->emit(*cb, _regalloc);
1358 cb->flush_bundle(true);
1359 current_offset = cb->insts_size();
1360 }
1361 #ifdef ASSERT
1362 jmp_target[i] = block_num;
1363 jmp_offset[i] = current_offset - blk_offset;
1364 jmp_size[i] = new_size;
1365 jmp_rule[i] = mach->rule();
1366 #endif
1367 block->map_node(replacement, j);
1368 mach->subsume_by(replacement, C);
1369 n = replacement;
1370 mach = replacement;
1371 }
1372 }
1373 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1374 } else if (mach->ideal_Opcode() == Op_Jump) {
1375 for (uint h = 0; h < block->_num_succs; h++) {
1376 Block* succs_block = block->_succs[h];
1377 for (uint j = 1; j < succs_block->num_preds(); j++) {
1378 Node* jpn = succs_block->pred(j);
1379 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1380 uint block_num = succs_block->non_connector()->_pre_order;
1381 Label *blkLabel = &blk_labels[block_num];
1382 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1383 }
1384 }
1385 }
1386 }
1387 #ifdef ASSERT
1388 // Check that oop-store precedes the card-mark
1389 else if (mach->ideal_Opcode() == Op_StoreCM) {
1390 uint storeCM_idx = j;
1391 int count = 0;
1392 for (uint prec = mach->req(); prec < mach->len(); prec++) {
1393 Node *oop_store = mach->in(prec); // Precedence edge
1394 if (oop_store == NULL) continue;
1395 count++;
1396 uint i4;
1397 for (i4 = 0; i4 < last_inst; ++i4) {
1398 if (block->get_node(i4) == oop_store) {
1399 break;
1400 }
1401 }
1402 // Note: This test can provide a false failure if other precedence
1403 // edges have been added to the storeCMNode.
1404 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1405 }
1406 assert(count > 0, "storeCM expects at least one precedence edge");
1407 }
1408 #endif
1409 else if (!n->is_Proj()) {
1410 // Remember the beginning of the previous instruction, in case
1411 // it's followed by a flag-kill and a null-check. Happens on
1412 // Intel all the time, with add-to-memory kind of opcodes.
1413 previous_offset = current_offset;
1414 }
1415
1416 // Not an else-if!
1417 // If this is a trap based cmp then add its offset to the list.
1418 if (mach->is_TrapBasedCheckNode()) {
1419 inct_starts[inct_cnt++] = current_offset;
1420 }
1421 }
1422
1423 // Verify that there is sufficient space remaining
1424 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1425 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1426 C->record_failure("CodeCache is full");
1427 return;
1428 }
1429
1430 // Save the offset for the listing
1431 #ifndef PRODUCT
1432 if (node_offsets && n->_idx < node_offset_limit)
1433 node_offsets[n->_idx] = cb->insts_size();
1434 #endif
1435
1436 // "Normal" instruction case
1437 DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
1438 n->emit(*cb, _regalloc);
1439 current_offset = cb->insts_size();
1440
1441 // Above we only verified that there is enough space in the instruction section.
1442 // However, the instruction may emit stubs that cause code buffer expansion.
1443 // Bail out here if expansion failed due to a lack of code cache space.
1444 if (failing()) {
1445 return;
1446 }
1447
1448 #ifdef ASSERT
1449 if (n->size(_regalloc) < (current_offset-instr_offset)) {
1450 n->dump();
1451 assert(false, "wrong size of mach node");
1452 }
1453 #endif
1454 non_safepoints.observe_instruction(n, current_offset);
1455
1456 // mcall is last "call" that can be a safepoint
1457 // record it so we can see if a poll will directly follow it
1458 // in which case we'll need a pad to make the PcDesc sites unique
1459 // see 5010568. This can be slightly inaccurate but conservative
1460 // in the case that return address is not actually at current_offset.
1461 // This is a small price to pay.
1462
1463 if (is_mcall) {
1464 last_call_offset = current_offset;
1465 }
1466
1467 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1468 // Avoid back to back some instructions.
1469 last_avoid_back_to_back_offset = current_offset;
1470 }
1471
1472 // See if this instruction has a delay slot
1473 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1474 guarantee(delay_slot != NULL, "expecting delay slot node");
1475
1476 // Back up 1 instruction
1477 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1478
1479 // Save the offset for the listing
1480 #ifndef PRODUCT
1481 if (node_offsets && delay_slot->_idx < node_offset_limit)
1482 node_offsets[delay_slot->_idx] = cb->insts_size();
1483 #endif
1484
1485 // Support a SafePoint in the delay slot
1486 if (delay_slot->is_MachSafePoint()) {
1487 MachNode *mach = delay_slot->as_Mach();
1488 // !!!!! Stubs only need an oopmap right now, so bail out
1489 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
1490 // Write the oopmap directly to the code blob??!!
1491 delay_slot = NULL;
1492 continue;
1493 }
1494
1495 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1496 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1497 adjusted_offset);
1498 // Generate an OopMap entry
1499 Process_OopMap_Node(mach, adjusted_offset);
1500 }
1501
1502 // Insert the delay slot instruction
1503 delay_slot->emit(*cb, _regalloc);
1504
1505 // Don't reuse it
1506 delay_slot = NULL;
1507 }
1508
1509 } // End for all instructions in block
1510
1511 // If the next block is the top of a loop, pad this block out to align
1512 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1513 if (i < nblocks-1) {
1514 Block *nb = _cfg->get_block(i + 1);
1515 int padding = nb->alignment_padding(current_offset);
1516 if( padding > 0 ) {
1517 MachNode *nop = new MachNopNode(padding / nop_size);
1518 block->insert_node(nop, block->number_of_nodes());
1519 _cfg->map_node_to_block(nop, block);
1520 nop->emit(*cb, _regalloc);
1521 current_offset = cb->insts_size();
1522 }
1523 }
1524 // Verify that the distance for generated before forward
1525 // short branches is still valid.
1526 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1527
1528 // Save new block start offset
1529 blk_starts[i] = blk_offset;
1530 } // End of for all blocks
1531 blk_starts[nblocks] = current_offset;
1532
1533 non_safepoints.flush_at_end();
1534
1535 // Offset too large?
1536 if (failing()) return;
1537
1538 // Define a pseudo-label at the end of the code
1539 MacroAssembler(cb).bind( blk_labels[nblocks] );
1540
1541 // Compute the size of the first block
1542 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1543
1544 #ifdef ASSERT
1545 for (uint i = 0; i < nblocks; i++) { // For all blocks
1546 if (jmp_target[i] != 0) {
1547 int br_size = jmp_size[i];
1548 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1549 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1550 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]);
1551 assert(false, "Displacement too large for short jmp");
1552 }
1553 }
1554 }
1555 #endif
1556
1557 #ifndef PRODUCT
1558 // Information on the size of the method, without the extraneous code
1559 Scheduling::increment_method_size(cb->insts_size());
1560 #endif
1561
1562 // ------------------
1563 // Fill in exception table entries.
1564 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1565
1566 // Only java methods have exception handlers and deopt handlers
1567 // class HandlerImpl is platform-specific and defined in the *.ad files.
1568 if (_method) {
1569 // Emit the exception handler code.
1570 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb));
1571 if (failing()) {
1572 return; // CodeBuffer::expand failed
1573 }
1574 // Emit the deopt handler code.
1575 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb));
1576
1577 // Emit the MethodHandle deopt handler code (if required).
1578 if (has_method_handle_invokes() && !failing()) {
1579 // We can use the same code as for the normal deopt handler, we
1580 // just need a different entry point address.
1581 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb));
1582 }
1583 }
1584
1585 // One last check for failed CodeBuffer::expand:
1586 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1587 C->record_failure("CodeCache is full");
1588 return;
1589 }
1590
1591 #ifndef PRODUCT
1592 // Dump the assembly code, including basic-block numbers
1593 if (print_assembly()) {
1594 ttyLocker ttyl; // keep the following output all in one block
1595 if (!VMThread::should_terminate()) { // test this under the tty lock
1596 // This output goes directly to the tty, not the compiler log.
1597 // To enable tools to match it up with the compilation activity,
1598 // be sure to tag this tty output with the compile ID.
1599 if (xtty != NULL) {
1600 xtty->head("opto_assembly compile_id='%d'%s", compile_id(),
1601 is_osr_compilation() ? " compile_kind='osr'" :
1602 "");
1603 }
1604 if (method() != NULL) {
1605 method()->print_metadata();
1606 } else if (stub_name() != NULL) {
1607 tty->print_cr("Generating RuntimeStub - %s", stub_name());
1608 }
1609 dump_asm(node_offsets, node_offset_limit);
1610 if (xtty != NULL) {
1611 // print_metadata and dump_asm above may safepoint which makes us loose the ttylock.
1612 // Retake lock too make sure the end tag is coherent, and that xmlStream->pop_tag is done
1613 // thread safe
1614 ttyLocker ttyl2;
1615 xtty->tail("opto_assembly");
1616 }
1617 }
1618 }
1619 #endif
1620
1621 }
1622
1623 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1624 _inc_table.set_size(cnt);
1625
1626 uint inct_cnt = 0;
1627 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1628 Block* block = _cfg->get_block(i);
1629 Node *n = NULL;
1630 int j;
1631
1632 // Find the branch; ignore trailing NOPs.
1633 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1634 n = block->get_node(j);
1635 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1636 break;
1637 }
1638 }
1639
1640 // If we didn't find anything, continue
1641 if (j < 0) {
1642 continue;
1643 }
1644
1645 // Compute ExceptionHandlerTable subtable entry and add it
1646 // (skip empty blocks)
1647 if (n->is_Catch()) {
1648
1649 // Get the offset of the return from the call
1650 uint call_return = call_returns[block->_pre_order];
1651 #ifdef ASSERT
1652 assert( call_return > 0, "no call seen for this basic block" );
1653 while (block->get_node(--j)->is_MachProj()) ;
1654 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1655 #endif
1656 // last instruction is a CatchNode, find it's CatchProjNodes
1657 int nof_succs = block->_num_succs;
1658 // allocate space
1659 GrowableArray<intptr_t> handler_bcis(nof_succs);
1660 GrowableArray<intptr_t> handler_pcos(nof_succs);
1661 // iterate through all successors
1662 for (int j = 0; j < nof_succs; j++) {
1663 Block* s = block->_succs[j];
1664 bool found_p = false;
1665 for (uint k = 1; k < s->num_preds(); k++) {
1666 Node* pk = s->pred(k);
1667 if (pk->is_CatchProj() && pk->in(0) == n) {
1668 const CatchProjNode* p = pk->as_CatchProj();
1669 found_p = true;
1670 // add the corresponding handler bci & pco information
1671 if (p->_con != CatchProjNode::fall_through_index) {
1672 // p leads to an exception handler (and is not fall through)
1673 assert(s == _cfg->get_block(s->_pre_order), "bad numbering");
1674 // no duplicates, please
1675 if (!handler_bcis.contains(p->handler_bci())) {
1676 uint block_num = s->non_connector()->_pre_order;
1677 handler_bcis.append(p->handler_bci());
1678 handler_pcos.append(blk_labels[block_num].loc_pos());
1679 }
1680 }
1681 }
1682 }
1683 assert(found_p, "no matching predecessor found");
1684 // Note: Due to empty block removal, one block may have
1685 // several CatchProj inputs, from the same Catch.
1686 }
1687
1688 // Set the offset of the return from the call
1689 assert(handler_bcis.find(-1) != -1, "must have default handler");
1690 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
1691 continue;
1692 }
1693
1694 // Handle implicit null exception table updates
1695 if (n->is_MachNullCheck()) {
1696 uint block_num = block->non_connector_successor(0)->_pre_order;
1697 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1698 continue;
1699 }
1700 // Handle implicit exception table updates: trap instructions.
1701 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
1702 uint block_num = block->non_connector_successor(0)->_pre_order;
1703 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1704 continue;
1705 }
1706 } // End of for all blocks fill in exception table entries
1707 }
1708
1709 // Static Variables
1710 #ifndef PRODUCT
1711 uint Scheduling::_total_nop_size = 0;
1712 uint Scheduling::_total_method_size = 0;
1713 uint Scheduling::_total_branches = 0;
1714 uint Scheduling::_total_unconditional_delays = 0;
1715 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1716 #endif
1717
1718 // Initializer for class Scheduling
1719
1720 Scheduling::Scheduling(Arena *arena, Compile &compile)
1721 : _arena(arena),
1722 _cfg(compile.cfg()),
1723 _regalloc(compile.regalloc()),
1724 _scheduled(arena),
1725 _available(arena),
1726 _reg_node(arena),
1727 _pinch_free_list(arena),
1728 _next_node(NULL),
1729 _bundle_instr_count(0),
1730 _bundle_cycle_number(0),
1731 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
1732 #ifndef PRODUCT
1733 , _branches(0)
1734 , _unconditional_delays(0)
1735 #endif
1736 {
1737 // Create a MachNopNode
1738 _nop = new MachNopNode();
1739
1740 // Now that the nops are in the array, save the count
1741 // (but allow entries for the nops)
1742 _node_bundling_limit = compile.unique();
1743 uint node_max = _regalloc->node_regs_max_index();
1744
1745 compile.set_node_bundling_limit(_node_bundling_limit);
1746
1747 // This one is persistent within the Compile class
1748 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1749
1750 // Allocate space for fixed-size arrays
1751 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1752 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
1753 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1754
1755 // Clear the arrays
1756 for (uint i = 0; i < node_max; i++) {
1757 ::new (&_node_bundling_base[i]) Bundle();
1758 }
1759 memset(_node_latency, 0, node_max * sizeof(unsigned short));
1760 memset(_uses, 0, node_max * sizeof(short));
1761 memset(_current_latency, 0, node_max * sizeof(unsigned short));
1762
1763 // Clear the bundling information
1764 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
1765
1766 // Get the last node
1767 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
1768
1769 _next_node = block->get_node(block->number_of_nodes() - 1);
1770 }
1771
1772 #ifndef PRODUCT
1773 // Scheduling destructor
1774 Scheduling::~Scheduling() {
1775 _total_branches += _branches;
1776 _total_unconditional_delays += _unconditional_delays;
1777 }
1778 #endif
1779
1780 // Step ahead "i" cycles
1781 void Scheduling::step(uint i) {
1782
1783 Bundle *bundle = node_bundling(_next_node);
1784 bundle->set_starts_bundle();
1785
1786 // Update the bundle record, but leave the flags information alone
1787 if (_bundle_instr_count > 0) {
1788 bundle->set_instr_count(_bundle_instr_count);
1789 bundle->set_resources_used(_bundle_use.resourcesUsed());
1790 }
1791
1792 // Update the state information
1793 _bundle_instr_count = 0;
1794 _bundle_cycle_number += i;
1795 _bundle_use.step(i);
1796 }
1797
1798 void Scheduling::step_and_clear() {
1799 Bundle *bundle = node_bundling(_next_node);
1800 bundle->set_starts_bundle();
1801
1802 // Update the bundle record
1803 if (_bundle_instr_count > 0) {
1804 bundle->set_instr_count(_bundle_instr_count);
1805 bundle->set_resources_used(_bundle_use.resourcesUsed());
1806
1807 _bundle_cycle_number += 1;
1808 }
1809
1810 // Clear the bundling information
1811 _bundle_instr_count = 0;
1812 _bundle_use.reset();
1813
1814 memcpy(_bundle_use_elements,
1815 Pipeline_Use::elaborated_elements,
1816 sizeof(Pipeline_Use::elaborated_elements));
1817 }
1818
1819 // Perform instruction scheduling and bundling over the sequence of
1820 // instructions in backwards order.
1821 void Compile::ScheduleAndBundle() {
1822
1823 // Don't optimize this if it isn't a method
1824 if (!_method)
1825 return;
1826
1827 // Don't optimize this if scheduling is disabled
1828 if (!do_scheduling())
1829 return;
1830
1831 // Scheduling code works only with pairs (16 bytes) maximum.
1832 if (max_vector_size() > 16)
1833 return;
1834
1835 TracePhase tp("isched", &timers[_t_instrSched]);
1836
1837 // Create a data structure for all the scheduling information
1838 Scheduling scheduling(Thread::current()->resource_area(), *this);
1839
1840 // Walk backwards over each basic block, computing the needed alignment
1841 // Walk over all the basic blocks
1842 scheduling.DoScheduling();
1843 }
1844
1845 // Compute the latency of all the instructions. This is fairly simple,
1846 // because we already have a legal ordering. Walk over the instructions
1847 // from first to last, and compute the latency of the instruction based
1848 // on the latency of the preceding instruction(s).
1849 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
1850 #ifndef PRODUCT
1851 if (_cfg->C->trace_opto_output())
1852 tty->print("# -> ComputeLocalLatenciesForward\n");
1853 #endif
1854
1855 // Walk over all the schedulable instructions
1856 for( uint j=_bb_start; j < _bb_end; j++ ) {
1857
1858 // This is a kludge, forcing all latency calculations to start at 1.
1859 // Used to allow latency 0 to force an instruction to the beginning
1860 // of the bb
1861 uint latency = 1;
1862 Node *use = bb->get_node(j);
1863 uint nlen = use->len();
1864
1865 // Walk over all the inputs
1866 for ( uint k=0; k < nlen; k++ ) {
1867 Node *def = use->in(k);
1868 if (!def)
1869 continue;
1870
1871 uint l = _node_latency[def->_idx] + use->latency(k);
1872 if (latency < l)
1873 latency = l;
1874 }
1875
1876 _node_latency[use->_idx] = latency;
1877
1878 #ifndef PRODUCT
1879 if (_cfg->C->trace_opto_output()) {
1880 tty->print("# latency %4d: ", latency);
1881 use->dump();
1882 }
1883 #endif
1884 }
1885
1886 #ifndef PRODUCT
1887 if (_cfg->C->trace_opto_output())
1888 tty->print("# <- ComputeLocalLatenciesForward\n");
1889 #endif
1890
1891 } // end ComputeLocalLatenciesForward
1892
1893 // See if this node fits into the present instruction bundle
1894 bool Scheduling::NodeFitsInBundle(Node *n) {
1895 uint n_idx = n->_idx;
1896
1897 // If this is the unconditional delay instruction, then it fits
1898 if (n == _unconditional_delay_slot) {
1899 #ifndef PRODUCT
1900 if (_cfg->C->trace_opto_output())
1901 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
1902 #endif
1903 return (true);
1904 }
1905
1906 // If the node cannot be scheduled this cycle, skip it
1907 if (_current_latency[n_idx] > _bundle_cycle_number) {
1908 #ifndef PRODUCT
1909 if (_cfg->C->trace_opto_output())
1910 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
1911 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
1912 #endif
1913 return (false);
1914 }
1915
1916 const Pipeline *node_pipeline = n->pipeline();
1917
1918 uint instruction_count = node_pipeline->instructionCount();
1919 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
1920 instruction_count = 0;
1921 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
1922 instruction_count++;
1923
1924 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
1925 #ifndef PRODUCT
1926 if (_cfg->C->trace_opto_output())
1927 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
1928 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
1929 #endif
1930 return (false);
1931 }
1932
1933 // Don't allow non-machine nodes to be handled this way
1934 if (!n->is_Mach() && instruction_count == 0)
1935 return (false);
1936
1937 // See if there is any overlap
1938 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
1939
1940 if (delay > 0) {
1941 #ifndef PRODUCT
1942 if (_cfg->C->trace_opto_output())
1943 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
1944 #endif
1945 return false;
1946 }
1947
1948 #ifndef PRODUCT
1949 if (_cfg->C->trace_opto_output())
1950 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
1951 #endif
1952
1953 return true;
1954 }
1955
1956 Node * Scheduling::ChooseNodeToBundle() {
1957 uint siz = _available.size();
1958
1959 if (siz == 0) {
1960
1961 #ifndef PRODUCT
1962 if (_cfg->C->trace_opto_output())
1963 tty->print("# ChooseNodeToBundle: NULL\n");
1964 #endif
1965 return (NULL);
1966 }
1967
1968 // Fast path, if only 1 instruction in the bundle
1969 if (siz == 1) {
1970 #ifndef PRODUCT
1971 if (_cfg->C->trace_opto_output()) {
1972 tty->print("# ChooseNodeToBundle (only 1): ");
1973 _available[0]->dump();
1974 }
1975 #endif
1976 return (_available[0]);
1977 }
1978
1979 // Don't bother, if the bundle is already full
1980 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
1981 for ( uint i = 0; i < siz; i++ ) {
1982 Node *n = _available[i];
1983
1984 // Skip projections, we'll handle them another way
1985 if (n->is_Proj())
1986 continue;
1987
1988 // This presupposed that instructions are inserted into the
1989 // available list in a legality order; i.e. instructions that
1990 // must be inserted first are at the head of the list
1991 if (NodeFitsInBundle(n)) {
1992 #ifndef PRODUCT
1993 if (_cfg->C->trace_opto_output()) {
1994 tty->print("# ChooseNodeToBundle: ");
1995 n->dump();
1996 }
1997 #endif
1998 return (n);
1999 }
2000 }
2001 }
2002
2003 // Nothing fits in this bundle, choose the highest priority
2004 #ifndef PRODUCT
2005 if (_cfg->C->trace_opto_output()) {
2006 tty->print("# ChooseNodeToBundle: ");
2007 _available[0]->dump();
2008 }
2009 #endif
2010
2011 return _available[0];
2012 }
2013
2014 void Scheduling::AddNodeToAvailableList(Node *n) {
2015 assert( !n->is_Proj(), "projections never directly made available" );
2016 #ifndef PRODUCT
2017 if (_cfg->C->trace_opto_output()) {
2018 tty->print("# AddNodeToAvailableList: ");
2019 n->dump();
2020 }
2021 #endif
2022
2023 int latency = _current_latency[n->_idx];
2024
2025 // Insert in latency order (insertion sort)
2026 uint i;
2027 for ( i=0; i < _available.size(); i++ )
2028 if (_current_latency[_available[i]->_idx] > latency)
2029 break;
2030
2031 // Special Check for compares following branches
2032 if( n->is_Mach() && _scheduled.size() > 0 ) {
2033 int op = n->as_Mach()->ideal_Opcode();
2034 Node *last = _scheduled[0];
2035 if( last->is_MachIf() && last->in(1) == n &&
2036 ( op == Op_CmpI ||
2037 op == Op_CmpU ||
2038 op == Op_CmpUL ||
2039 op == Op_CmpP ||
2040 op == Op_CmpF ||
2041 op == Op_CmpD ||
2042 op == Op_CmpL ) ) {
2043
2044 // Recalculate position, moving to front of same latency
2045 for ( i=0 ; i < _available.size(); i++ )
2046 if (_current_latency[_available[i]->_idx] >= latency)
2047 break;
2048 }
2049 }
2050
2051 // Insert the node in the available list
2052 _available.insert(i, n);
2053
2054 #ifndef PRODUCT
2055 if (_cfg->C->trace_opto_output())
2056 dump_available();
2057 #endif
2058 }
2059
2060 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2061 for ( uint i=0; i < n->len(); i++ ) {
2062 Node *def = n->in(i);
2063 if (!def) continue;
2064 if( def->is_Proj() ) // If this is a machine projection, then
2065 def = def->in(0); // propagate usage thru to the base instruction
2066
2067 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2068 continue;
2069 }
2070
2071 // Compute the latency
2072 uint l = _bundle_cycle_number + n->latency(i);
2073 if (_current_latency[def->_idx] < l)
2074 _current_latency[def->_idx] = l;
2075
2076 // If this does not have uses then schedule it
2077 if ((--_uses[def->_idx]) == 0)
2078 AddNodeToAvailableList(def);
2079 }
2080 }
2081
2082 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2083 #ifndef PRODUCT
2084 if (_cfg->C->trace_opto_output()) {
2085 tty->print("# AddNodeToBundle: ");
2086 n->dump();
2087 }
2088 #endif
2089
2090 // Remove this from the available list
2091 uint i;
2092 for (i = 0; i < _available.size(); i++)
2093 if (_available[i] == n)
2094 break;
2095 assert(i < _available.size(), "entry in _available list not found");
2096 _available.remove(i);
2097
2098 // See if this fits in the current bundle
2099 const Pipeline *node_pipeline = n->pipeline();
2100 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2101
2102 // Check for instructions to be placed in the delay slot. We
2103 // do this before we actually schedule the current instruction,
2104 // because the delay slot follows the current instruction.
2105 if (Pipeline::_branch_has_delay_slot &&
2106 node_pipeline->hasBranchDelay() &&
2107 !_unconditional_delay_slot) {
2108
2109 uint siz = _available.size();
2110
2111 // Conditional branches can support an instruction that
2112 // is unconditionally executed and not dependent by the
2113 // branch, OR a conditionally executed instruction if
2114 // the branch is taken. In practice, this means that
2115 // the first instruction at the branch target is
2116 // copied to the delay slot, and the branch goes to
2117 // the instruction after that at the branch target
2118 if ( n->is_MachBranch() ) {
2119
2120 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2121 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2122
2123 #ifndef PRODUCT
2124 _branches++;
2125 #endif
2126
2127 // At least 1 instruction is on the available list
2128 // that is not dependent on the branch
2129 for (uint i = 0; i < siz; i++) {
2130 Node *d = _available[i];
2131 const Pipeline *avail_pipeline = d->pipeline();
2132
2133 // Don't allow safepoints in the branch shadow, that will
2134 // cause a number of difficulties
2135 if ( avail_pipeline->instructionCount() == 1 &&
2136 !avail_pipeline->hasMultipleBundles() &&
2137 !avail_pipeline->hasBranchDelay() &&
2138 Pipeline::instr_has_unit_size() &&
2139 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2140 NodeFitsInBundle(d) &&
2141 !node_bundling(d)->used_in_delay()) {
2142
2143 if (d->is_Mach() && !d->is_MachSafePoint()) {
2144 // A node that fits in the delay slot was found, so we need to
2145 // set the appropriate bits in the bundle pipeline information so
2146 // that it correctly indicates resource usage. Later, when we
2147 // attempt to add this instruction to the bundle, we will skip
2148 // setting the resource usage.
2149 _unconditional_delay_slot = d;
2150 node_bundling(n)->set_use_unconditional_delay();
2151 node_bundling(d)->set_used_in_unconditional_delay();
2152 _bundle_use.add_usage(avail_pipeline->resourceUse());
2153 _current_latency[d->_idx] = _bundle_cycle_number;
2154 _next_node = d;
2155 ++_bundle_instr_count;
2156 #ifndef PRODUCT
2157 _unconditional_delays++;
2158 #endif
2159 break;
2160 }
2161 }
2162 }
2163 }
2164
2165 // No delay slot, add a nop to the usage
2166 if (!_unconditional_delay_slot) {
2167 // See if adding an instruction in the delay slot will overflow
2168 // the bundle.
2169 if (!NodeFitsInBundle(_nop)) {
2170 #ifndef PRODUCT
2171 if (_cfg->C->trace_opto_output())
2172 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2173 #endif
2174 step(1);
2175 }
2176
2177 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2178 _next_node = _nop;
2179 ++_bundle_instr_count;
2180 }
2181
2182 // See if the instruction in the delay slot requires a
2183 // step of the bundles
2184 if (!NodeFitsInBundle(n)) {
2185 #ifndef PRODUCT
2186 if (_cfg->C->trace_opto_output())
2187 tty->print("# *** STEP(branch won't fit) ***\n");
2188 #endif
2189 // Update the state information
2190 _bundle_instr_count = 0;
2191 _bundle_cycle_number += 1;
2192 _bundle_use.step(1);
2193 }
2194 }
2195
2196 // Get the number of instructions
2197 uint instruction_count = node_pipeline->instructionCount();
2198 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2199 instruction_count = 0;
2200
2201 // Compute the latency information
2202 uint delay = 0;
2203
2204 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2205 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2206 if (relative_latency < 0)
2207 relative_latency = 0;
2208
2209 delay = _bundle_use.full_latency(relative_latency, node_usage);
2210
2211 // Does not fit in this bundle, start a new one
2212 if (delay > 0) {
2213 step(delay);
2214
2215 #ifndef PRODUCT
2216 if (_cfg->C->trace_opto_output())
2217 tty->print("# *** STEP(%d) ***\n", delay);
2218 #endif
2219 }
2220 }
2221
2222 // If this was placed in the delay slot, ignore it
2223 if (n != _unconditional_delay_slot) {
2224
2225 if (delay == 0) {
2226 if (node_pipeline->hasMultipleBundles()) {
2227 #ifndef PRODUCT
2228 if (_cfg->C->trace_opto_output())
2229 tty->print("# *** STEP(multiple instructions) ***\n");
2230 #endif
2231 step(1);
2232 }
2233
2234 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2235 #ifndef PRODUCT
2236 if (_cfg->C->trace_opto_output())
2237 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2238 instruction_count + _bundle_instr_count,
2239 Pipeline::_max_instrs_per_cycle);
2240 #endif
2241 step(1);
2242 }
2243 }
2244
2245 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2246 _bundle_instr_count++;
2247
2248 // Set the node's latency
2249 _current_latency[n->_idx] = _bundle_cycle_number;
2250
2251 // Now merge the functional unit information
2252 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2253 _bundle_use.add_usage(node_usage);
2254
2255 // Increment the number of instructions in this bundle
2256 _bundle_instr_count += instruction_count;
2257
2258 // Remember this node for later
2259 if (n->is_Mach())
2260 _next_node = n;
2261 }
2262
2263 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2264 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2265 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2266 // into the block. All other scheduled nodes get put in the schedule here.
2267 int op = n->Opcode();
2268 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2269 (op != Op_Node && // Not an unused antidepedence node and
2270 // not an unallocated boxlock
2271 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2272
2273 // Push any trailing projections
2274 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2275 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2276 Node *foi = n->fast_out(i);
2277 if( foi->is_Proj() )
2278 _scheduled.push(foi);
2279 }
2280 }
2281
2282 // Put the instruction in the schedule list
2283 _scheduled.push(n);
2284 }
2285
2286 #ifndef PRODUCT
2287 if (_cfg->C->trace_opto_output())
2288 dump_available();
2289 #endif
2290
2291 // Walk all the definitions, decrementing use counts, and
2292 // if a definition has a 0 use count, place it in the available list.
2293 DecrementUseCounts(n,bb);
2294 }
2295
2296 // This method sets the use count within a basic block. We will ignore all
2297 // uses outside the current basic block. As we are doing a backwards walk,
2298 // any node we reach that has a use count of 0 may be scheduled. This also
2299 // avoids the problem of cyclic references from phi nodes, as long as phi
2300 // nodes are at the front of the basic block. This method also initializes
2301 // the available list to the set of instructions that have no uses within this
2302 // basic block.
2303 void Scheduling::ComputeUseCount(const Block *bb) {
2304 #ifndef PRODUCT
2305 if (_cfg->C->trace_opto_output())
2306 tty->print("# -> ComputeUseCount\n");
2307 #endif
2308
2309 // Clear the list of available and scheduled instructions, just in case
2310 _available.clear();
2311 _scheduled.clear();
2312
2313 // No delay slot specified
2314 _unconditional_delay_slot = NULL;
2315
2316 #ifdef ASSERT
2317 for( uint i=0; i < bb->number_of_nodes(); i++ )
2318 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2319 #endif
2320
2321 // Force the _uses count to never go to zero for unscheduable pieces
2322 // of the block
2323 for( uint k = 0; k < _bb_start; k++ )
2324 _uses[bb->get_node(k)->_idx] = 1;
2325 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2326 _uses[bb->get_node(l)->_idx] = 1;
2327
2328 // Iterate backwards over the instructions in the block. Don't count the
2329 // branch projections at end or the block header instructions.
2330 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2331 Node *n = bb->get_node(j);
2332 if( n->is_Proj() ) continue; // Projections handled another way
2333
2334 // Account for all uses
2335 for ( uint k = 0; k < n->len(); k++ ) {
2336 Node *inp = n->in(k);
2337 if (!inp) continue;
2338 assert(inp != n, "no cycles allowed" );
2339 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2340 if (inp->is_Proj()) { // Skip through Proj's
2341 inp = inp->in(0);
2342 }
2343 ++_uses[inp->_idx]; // Count 1 block-local use
2344 }
2345 }
2346
2347 // If this instruction has a 0 use count, then it is available
2348 if (!_uses[n->_idx]) {
2349 _current_latency[n->_idx] = _bundle_cycle_number;
2350 AddNodeToAvailableList(n);
2351 }
2352
2353 #ifndef PRODUCT
2354 if (_cfg->C->trace_opto_output()) {
2355 tty->print("# uses: %3d: ", _uses[n->_idx]);
2356 n->dump();
2357 }
2358 #endif
2359 }
2360
2361 #ifndef PRODUCT
2362 if (_cfg->C->trace_opto_output())
2363 tty->print("# <- ComputeUseCount\n");
2364 #endif
2365 }
2366
2367 // This routine performs scheduling on each basic block in reverse order,
2368 // using instruction latencies and taking into account function unit
2369 // availability.
2370 void Scheduling::DoScheduling() {
2371 #ifndef PRODUCT
2372 if (_cfg->C->trace_opto_output())
2373 tty->print("# -> DoScheduling\n");
2374 #endif
2375
2376 Block *succ_bb = NULL;
2377 Block *bb;
2378
2379 // Walk over all the basic blocks in reverse order
2380 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2381 bb = _cfg->get_block(i);
2382
2383 #ifndef PRODUCT
2384 if (_cfg->C->trace_opto_output()) {
2385 tty->print("# Schedule BB#%03d (initial)\n", i);
2386 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2387 bb->get_node(j)->dump();
2388 }
2389 }
2390 #endif
2391
2392 // On the head node, skip processing
2393 if (bb == _cfg->get_root_block()) {
2394 continue;
2395 }
2396
2397 // Skip empty, connector blocks
2398 if (bb->is_connector())
2399 continue;
2400
2401 // If the following block is not the sole successor of
2402 // this one, then reset the pipeline information
2403 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2404 #ifndef PRODUCT
2405 if (_cfg->C->trace_opto_output()) {
2406 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2407 _next_node->_idx, _bundle_instr_count);
2408 }
2409 #endif
2410 step_and_clear();
2411 }
2412
2413 // Leave untouched the starting instruction, any Phis, a CreateEx node
2414 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2415 _bb_end = bb->number_of_nodes()-1;
2416 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2417 Node *n = bb->get_node(_bb_start);
2418 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2419 // Also, MachIdealNodes do not get scheduled
2420 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2421 MachNode *mach = n->as_Mach();
2422 int iop = mach->ideal_Opcode();
2423 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2424 if( iop == Op_Con ) continue; // Do not schedule Top
2425 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2426 mach->pipeline() == MachNode::pipeline_class() &&
2427 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
2428 continue;
2429 break; // Funny loop structure to be sure...
2430 }
2431 // Compute last "interesting" instruction in block - last instruction we
2432 // might schedule. _bb_end points just after last schedulable inst. We
2433 // normally schedule conditional branches (despite them being forced last
2434 // in the block), because they have delay slots we can fill. Calls all
2435 // have their delay slots filled in the template expansions, so we don't
2436 // bother scheduling them.
2437 Node *last = bb->get_node(_bb_end);
2438 // Ignore trailing NOPs.
2439 while (_bb_end > 0 && last->is_Mach() &&
2440 last->as_Mach()->ideal_Opcode() == Op_Con) {
2441 last = bb->get_node(--_bb_end);
2442 }
2443 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2444 if( last->is_Catch() ||
2445 // Exclude unreachable path case when Halt node is in a separate block.
2446 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2447 // There must be a prior call. Skip it.
2448 while( !bb->get_node(--_bb_end)->is_MachCall() ) {
2449 assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" );
2450 }
2451 } else if( last->is_MachNullCheck() ) {
2452 // Backup so the last null-checked memory instruction is
2453 // outside the schedulable range. Skip over the nullcheck,
2454 // projection, and the memory nodes.
2455 Node *mem = last->in(1);
2456 do {
2457 _bb_end--;
2458 } while (mem != bb->get_node(_bb_end));
2459 } else {
2460 // Set _bb_end to point after last schedulable inst.
2461 _bb_end++;
2462 }
2463
2464 assert( _bb_start <= _bb_end, "inverted block ends" );
2465
2466 // Compute the register antidependencies for the basic block
2467 ComputeRegisterAntidependencies(bb);
2468 if (_cfg->C->failing()) return; // too many D-U pinch points
2469
2470 // Compute intra-bb latencies for the nodes
2471 ComputeLocalLatenciesForward(bb);
2472
2473 // Compute the usage within the block, and set the list of all nodes
2474 // in the block that have no uses within the block.
2475 ComputeUseCount(bb);
2476
2477 // Schedule the remaining instructions in the block
2478 while ( _available.size() > 0 ) {
2479 Node *n = ChooseNodeToBundle();
2480 guarantee(n != NULL, "no nodes available");
2481 AddNodeToBundle(n,bb);
2482 }
2483
2484 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2485 #ifdef ASSERT
2486 for( uint l = _bb_start; l < _bb_end; l++ ) {
2487 Node *n = bb->get_node(l);
2488 uint m;
2489 for( m = 0; m < _bb_end-_bb_start; m++ )
2490 if( _scheduled[m] == n )
2491 break;
2492 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2493 }
2494 #endif
2495
2496 // Now copy the instructions (in reverse order) back to the block
2497 for ( uint k = _bb_start; k < _bb_end; k++ )
2498 bb->map_node(_scheduled[_bb_end-k-1], k);
2499
2500 #ifndef PRODUCT
2501 if (_cfg->C->trace_opto_output()) {
2502 tty->print("# Schedule BB#%03d (final)\n", i);
2503 uint current = 0;
2504 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2505 Node *n = bb->get_node(j);
2506 if( valid_bundle_info(n) ) {
2507 Bundle *bundle = node_bundling(n);
2508 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2509 tty->print("*** Bundle: ");
2510 bundle->dump();
2511 }
2512 n->dump();
2513 }
2514 }
2515 }
2516 #endif
2517 #ifdef ASSERT
2518 verify_good_schedule(bb,"after block local scheduling");
2519 #endif
2520 }
2521
2522 #ifndef PRODUCT
2523 if (_cfg->C->trace_opto_output())
2524 tty->print("# <- DoScheduling\n");
2525 #endif
2526
2527 // Record final node-bundling array location
2528 _regalloc->C->set_node_bundling_base(_node_bundling_base);
2529
2530 } // end DoScheduling
2531
2532 // Verify that no live-range used in the block is killed in the block by a
2533 // wrong DEF. This doesn't verify live-ranges that span blocks.
2534
2535 // Check for edge existence. Used to avoid adding redundant precedence edges.
2536 static bool edge_from_to( Node *from, Node *to ) {
2537 for( uint i=0; i<from->len(); i++ )
2538 if( from->in(i) == to )
2539 return true;
2540 return false;
2541 }
2542
2543 #ifdef ASSERT
2544 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2545 // Check for bad kills
2546 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2547 Node *prior_use = _reg_node[def];
2548 if( prior_use && !edge_from_to(prior_use,n) ) {
2549 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2550 n->dump();
2551 tty->print_cr("...");
2552 prior_use->dump();
2553 assert(edge_from_to(prior_use,n), "%s", msg);
2554 }
2555 _reg_node.map(def,NULL); // Kill live USEs
2556 }
2557 }
2558
2559 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2560
2561 // Zap to something reasonable for the verify code
2562 _reg_node.clear();
2563
2564 // Walk over the block backwards. Check to make sure each DEF doesn't
2565 // kill a live value (other than the one it's supposed to). Add each
2566 // USE to the live set.
2567 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2568 Node *n = b->get_node(i);
2569 int n_op = n->Opcode();
2570 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2571 // Fat-proj kills a slew of registers
2572 RegMask rm = n->out_RegMask();// Make local copy
2573 while( rm.is_NotEmpty() ) {
2574 OptoReg::Name kill = rm.find_first_elem();
2575 rm.Remove(kill);
2576 verify_do_def( n, kill, msg );
2577 }
2578 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2579 // Get DEF'd registers the normal way
2580 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2581 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2582 }
2583
2584 // Now make all USEs live
2585 for( uint i=1; i<n->req(); i++ ) {
2586 Node *def = n->in(i);
2587 assert(def != 0, "input edge required");
2588 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2589 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2590 if( OptoReg::is_valid(reg_lo) ) {
2591 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
2592 _reg_node.map(reg_lo,n);
2593 }
2594 if( OptoReg::is_valid(reg_hi) ) {
2595 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
2596 _reg_node.map(reg_hi,n);
2597 }
2598 }
2599
2600 }
2601
2602 // Zap to something reasonable for the Antidependence code
2603 _reg_node.clear();
2604 }
2605 #endif
2606
2607 // Conditionally add precedence edges. Avoid putting edges on Projs.
2608 static void add_prec_edge_from_to( Node *from, Node *to ) {
2609 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2610 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2611 from = from->in(0);
2612 }
2613 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2614 !edge_from_to( from, to ) ) // Avoid duplicate edge
2615 from->add_prec(to);
2616 }
2617
2618 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2619 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2620 return;
2621
2622 Node *pinch = _reg_node[def_reg]; // Get pinch point
2623 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2624 is_def ) { // Check for a true def (not a kill)
2625 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2626 return;
2627 }
2628
2629 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2630 debug_only( def = (Node*)((intptr_t)0xdeadbeef); )
2631
2632 // After some number of kills there _may_ be a later def
2633 Node *later_def = NULL;
2634
2635 // Finding a kill requires a real pinch-point.
2636 // Check for not already having a pinch-point.
2637 // Pinch points are Op_Node's.
2638 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2639 later_def = pinch; // Must be def/kill as optimistic pinch-point
2640 if ( _pinch_free_list.size() > 0) {
2641 pinch = _pinch_free_list.pop();
2642 } else {
2643 pinch = new Node(1); // Pinch point to-be
2644 }
2645 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2646 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2647 return;
2648 }
2649 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
2650 _reg_node.map(def_reg,pinch); // Record pinch-point
2651 //_regalloc->set_bad(pinch->_idx); // Already initialized this way.
2652 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2653 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call
2654 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2655 later_def = NULL; // and no later def
2656 }
2657 pinch->set_req(0,later_def); // Hook later def so we can find it
2658 } else { // Else have valid pinch point
2659 if( pinch->in(0) ) // If there is a later-def
2660 later_def = pinch->in(0); // Get it
2661 }
2662
2663 // Add output-dependence edge from later def to kill
2664 if( later_def ) // If there is some original def
2665 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2666
2667 // See if current kill is also a use, and so is forced to be the pinch-point.
2668 if( pinch->Opcode() == Op_Node ) {
2669 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2670 for( uint i=1; i<uses->req(); i++ ) {
2671 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2672 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2673 // Yes, found a use/kill pinch-point
2674 pinch->set_req(0,NULL); //
2675 pinch->replace_by(kill); // Move anti-dep edges up
2676 pinch = kill;
2677 _reg_node.map(def_reg,pinch);
2678 return;
2679 }
2680 }
2681 }
2682
2683 // Add edge from kill to pinch-point
2684 add_prec_edge_from_to(kill,pinch);
2685 }
2686
2687 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2688 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2689 return;
2690 Node *pinch = _reg_node[use_reg]; // Get pinch point
2691 // Check for no later def_reg/kill in block
2692 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b &&
2693 // Use has to be block-local as well
2694 _cfg->get_block_for_node(use) == b) {
2695 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2696 pinch->req() == 1 ) { // pinch not yet in block?
2697 pinch->del_req(0); // yank pointer to later-def, also set flag
2698 // Insert the pinch-point in the block just after the last use
2699 b->insert_node(pinch, b->find_node(use) + 1);
2700 _bb_end++; // Increase size scheduled region in block
2701 }
2702
2703 add_prec_edge_from_to(pinch,use);
2704 }
2705 }
2706
2707 // We insert antidependences between the reads and following write of
2708 // allocated registers to prevent illegal code motion. Hopefully, the
2709 // number of added references should be fairly small, especially as we
2710 // are only adding references within the current basic block.
2711 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2712
2713 #ifdef ASSERT
2714 verify_good_schedule(b,"before block local scheduling");
2715 #endif
2716
2717 // A valid schedule, for each register independently, is an endless cycle
2718 // of: a def, then some uses (connected to the def by true dependencies),
2719 // then some kills (defs with no uses), finally the cycle repeats with a new
2720 // def. The uses are allowed to float relative to each other, as are the
2721 // kills. No use is allowed to slide past a kill (or def). This requires
2722 // antidependencies between all uses of a single def and all kills that
2723 // follow, up to the next def. More edges are redundant, because later defs
2724 // & kills are already serialized with true or antidependencies. To keep
2725 // the edge count down, we add a 'pinch point' node if there's more than
2726 // one use or more than one kill/def.
2727
2728 // We add dependencies in one bottom-up pass.
2729
2730 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2731
2732 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2733 // register. If not, we record the DEF/KILL in _reg_node, the
2734 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
2735 // "pinch point", a new Node that's in the graph but not in the block.
2736 // We put edges from the prior and current DEF/KILLs to the pinch point.
2737 // We put the pinch point in _reg_node. If there's already a pinch point
2738 // we merely add an edge from the current DEF/KILL to the pinch point.
2739
2740 // After doing the DEF/KILLs, we handle USEs. For each used register, we
2741 // put an edge from the pinch point to the USE.
2742
2743 // To be expedient, the _reg_node array is pre-allocated for the whole
2744 // compilation. _reg_node is lazily initialized; it either contains a NULL,
2745 // or a valid def/kill/pinch-point, or a leftover node from some prior
2746 // block. Leftover node from some prior block is treated like a NULL (no
2747 // prior def, so no anti-dependence needed). Valid def is distinguished by
2748 // it being in the current block.
2749 bool fat_proj_seen = false;
2750 uint last_safept = _bb_end-1;
2751 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL;
2752 Node* last_safept_node = end_node;
2753 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2754 Node *n = b->get_node(i);
2755 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
2756 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
2757 // Fat-proj kills a slew of registers
2758 // This can add edges to 'n' and obscure whether or not it was a def,
2759 // hence the is_def flag.
2760 fat_proj_seen = true;
2761 RegMask rm = n->out_RegMask();// Make local copy
2762 while( rm.is_NotEmpty() ) {
2763 OptoReg::Name kill = rm.find_first_elem();
2764 rm.Remove(kill);
2765 anti_do_def( b, n, kill, is_def );
2766 }
2767 } else {
2768 // Get DEF'd registers the normal way
2769 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2770 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2771 }
2772
2773 // Kill projections on a branch should appear to occur on the
2774 // branch, not afterwards, so grab the masks from the projections
2775 // and process them.
2776 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
2777 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2778 Node* use = n->fast_out(i);
2779 if (use->is_Proj()) {
2780 RegMask rm = use->out_RegMask();// Make local copy
2781 while( rm.is_NotEmpty() ) {
2782 OptoReg::Name kill = rm.find_first_elem();
2783 rm.Remove(kill);
2784 anti_do_def( b, n, kill, false );
2785 }
2786 }
2787 }
2788 }
2789
2790 // Check each register used by this instruction for a following DEF/KILL
2791 // that must occur afterward and requires an anti-dependence edge.
2792 for( uint j=0; j<n->req(); j++ ) {
2793 Node *def = n->in(j);
2794 if( def ) {
2795 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
2796 anti_do_use( b, n, _regalloc->get_reg_first(def) );
2797 anti_do_use( b, n, _regalloc->get_reg_second(def) );
2798 }
2799 }
2800 // Do not allow defs of new derived values to float above GC
2801 // points unless the base is definitely available at the GC point.
2802
2803 Node *m = b->get_node(i);
2804
2805 // Add precedence edge from following safepoint to use of derived pointer
2806 if( last_safept_node != end_node &&
2807 m != last_safept_node) {
2808 for (uint k = 1; k < m->req(); k++) {
2809 const Type *t = m->in(k)->bottom_type();
2810 if( t->isa_oop_ptr() &&
2811 t->is_ptr()->offset() != 0 ) {
2812 last_safept_node->add_prec( m );
2813 break;
2814 }
2815 }
2816 }
2817
2818 if( n->jvms() ) { // Precedence edge from derived to safept
2819 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2820 if( b->get_node(last_safept) != last_safept_node ) {
2821 last_safept = b->find_node(last_safept_node);
2822 }
2823 for( uint j=last_safept; j > i; j-- ) {
2824 Node *mach = b->get_node(j);
2825 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2826 mach->add_prec( n );
2827 }
2828 last_safept = i;
2829 last_safept_node = m;
2830 }
2831 }
2832
2833 if (fat_proj_seen) {
2834 // Garbage collect pinch nodes that were not consumed.
2835 // They are usually created by a fat kill MachProj for a call.
2836 garbage_collect_pinch_nodes();
2837 }
2838 }
2839
2840 // Garbage collect pinch nodes for reuse by other blocks.
2841 //
2842 // The block scheduler's insertion of anti-dependence
2843 // edges creates many pinch nodes when the block contains
2844 // 2 or more Calls. A pinch node is used to prevent a
2845 // combinatorial explosion of edges. If a set of kills for a
2846 // register is anti-dependent on a set of uses (or defs), rather
2847 // than adding an edge in the graph between each pair of kill
2848 // and use (or def), a pinch is inserted between them:
2849 //
2850 // use1 use2 use3
2851 // \ | /
2852 // \ | /
2853 // pinch
2854 // / | \
2855 // / | \
2856 // kill1 kill2 kill3
2857 //
2858 // One pinch node is created per register killed when
2859 // the second call is encountered during a backwards pass
2860 // over the block. Most of these pinch nodes are never
2861 // wired into the graph because the register is never
2862 // used or def'ed in the block.
2863 //
2864 void Scheduling::garbage_collect_pinch_nodes() {
2865 #ifndef PRODUCT
2866 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
2867 #endif
2868 int trace_cnt = 0;
2869 for (uint k = 0; k < _reg_node.Size(); k++) {
2870 Node* pinch = _reg_node[k];
2871 if ((pinch != NULL) && pinch->Opcode() == Op_Node &&
2872 // no predecence input edges
2873 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
2874 cleanup_pinch(pinch);
2875 _pinch_free_list.push(pinch);
2876 _reg_node.map(k, NULL);
2877 #ifndef PRODUCT
2878 if (_cfg->C->trace_opto_output()) {
2879 trace_cnt++;
2880 if (trace_cnt > 40) {
2881 tty->print("\n");
2882 trace_cnt = 0;
2883 }
2884 tty->print(" %d", pinch->_idx);
2885 }
2886 #endif
2887 }
2888 }
2889 #ifndef PRODUCT
2890 if (_cfg->C->trace_opto_output()) tty->print("\n");
2891 #endif
2892 }
2893
2894 // Clean up a pinch node for reuse.
2895 void Scheduling::cleanup_pinch( Node *pinch ) {
2896 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
2897
2898 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
2899 Node* use = pinch->last_out(i);
2900 uint uses_found = 0;
2901 for (uint j = use->req(); j < use->len(); j++) {
2902 if (use->in(j) == pinch) {
2903 use->rm_prec(j);
2904 uses_found++;
2905 }
2906 }
2907 assert(uses_found > 0, "must be a precedence edge");
2908 i -= uses_found; // we deleted 1 or more copies of this edge
2909 }
2910 // May have a later_def entry
2911 pinch->set_req(0, NULL);
2912 }
2913
2914 #ifndef PRODUCT
2915
2916 void Scheduling::dump_available() const {
2917 tty->print("#Availist ");
2918 for (uint i = 0; i < _available.size(); i++)
2919 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
2920 tty->cr();
2921 }
2922
2923 // Print Scheduling Statistics
2924 void Scheduling::print_statistics() {
2925 // Print the size added by nops for bundling
2926 tty->print("Nops added %d bytes to total of %d bytes",
2927 _total_nop_size, _total_method_size);
2928 if (_total_method_size > 0)
2929 tty->print(", for %.2f%%",
2930 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
2931 tty->print("\n");
2932
2933 // Print the number of branch shadows filled
2934 if (Pipeline::_branch_has_delay_slot) {
2935 tty->print("Of %d branches, %d had unconditional delay slots filled",
2936 _total_branches, _total_unconditional_delays);
2937 if (_total_branches > 0)
2938 tty->print(", for %.2f%%",
2939 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
2940 tty->print("\n");
2941 }
2942
2943 uint total_instructions = 0, total_bundles = 0;
2944
2945 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
2946 uint bundle_count = _total_instructions_per_bundle[i];
2947 total_instructions += bundle_count * i;
2948 total_bundles += bundle_count;
2949 }
2950
2951 if (total_bundles > 0)
2952 tty->print("Average ILP (excluding nops) is %.2f\n",
2953 ((double)total_instructions) / ((double)total_bundles));
2954 }
2955 #endif