Wed, 18 Sep 2013 14:34:56 -0700
8024342: PPC64 (part 111): Support for C calling conventions that require 64-bit ints.
Summary: Some platforms, as ppc and s390x/zArch require that 32-bit ints are passed as 64-bit values to C functions. This change adds support to adapt the signature and to issue proper casts to c2-compiled stubs. The functions are used in generate_native_wrapper(). Adapt signature used by the compiler as in PhaseIdealLoop::intrinsify_fill().
Reviewed-by: kvn
1 /*
2 * Copyright (c) 1998, 2012, 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 */
25 #include "precompiled.hpp"
26 #include "asm/assembler.inline.hpp"
27 #include "code/compiledIC.hpp"
28 #include "code/debugInfo.hpp"
29 #include "code/debugInfoRec.hpp"
30 #include "compiler/compileBroker.hpp"
31 #include "compiler/oopMap.hpp"
32 #include "memory/allocation.inline.hpp"
33 #include "opto/callnode.hpp"
34 #include "opto/cfgnode.hpp"
35 #include "opto/locknode.hpp"
36 #include "opto/machnode.hpp"
37 #include "opto/output.hpp"
38 #include "opto/regalloc.hpp"
39 #include "opto/runtime.hpp"
40 #include "opto/subnode.hpp"
41 #include "opto/type.hpp"
42 #include "runtime/handles.inline.hpp"
43 #include "utilities/xmlstream.hpp"
45 extern uint size_exception_handler();
46 extern uint size_deopt_handler();
48 #ifndef PRODUCT
49 #define DEBUG_ARG(x) , x
50 #else
51 #define DEBUG_ARG(x)
52 #endif
54 extern int emit_exception_handler(CodeBuffer &cbuf);
55 extern int emit_deopt_handler(CodeBuffer &cbuf);
57 // Convert Nodes to instruction bits and pass off to the VM
58 void Compile::Output() {
59 // RootNode goes
60 assert( _cfg->get_root_block()->_nodes.size() == 0, "" );
62 // The number of new nodes (mostly MachNop) is proportional to
63 // the number of java calls and inner loops which are aligned.
64 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
65 C->inner_loops()*(OptoLoopAlignment-1)),
66 "out of nodes before code generation" ) ) {
67 return;
68 }
69 // Make sure I can find the Start Node
70 Block *entry = _cfg->get_block(1);
71 Block *broot = _cfg->get_root_block();
73 const StartNode *start = entry->_nodes[0]->as_Start();
75 // Replace StartNode with prolog
76 MachPrologNode *prolog = new (this) MachPrologNode();
77 entry->_nodes.map( 0, prolog );
78 _cfg->map_node_to_block(prolog, entry);
79 _cfg->unmap_node_from_block(start); // start is no longer in any block
81 // Virtual methods need an unverified entry point
83 if( is_osr_compilation() ) {
84 if( PoisonOSREntry ) {
85 // TODO: Should use a ShouldNotReachHereNode...
86 _cfg->insert( broot, 0, new (this) MachBreakpointNode() );
87 }
88 } else {
89 if( _method && !_method->flags().is_static() ) {
90 // Insert unvalidated entry point
91 _cfg->insert( broot, 0, new (this) MachUEPNode() );
92 }
94 }
97 // Break before main entry point
98 if( (_method && _method->break_at_execute())
99 #ifndef PRODUCT
100 ||(OptoBreakpoint && is_method_compilation())
101 ||(OptoBreakpointOSR && is_osr_compilation())
102 ||(OptoBreakpointC2R && !_method)
103 #endif
104 ) {
105 // checking for _method means that OptoBreakpoint does not apply to
106 // runtime stubs or frame converters
107 _cfg->insert( entry, 1, new (this) MachBreakpointNode() );
108 }
110 // Insert epilogs before every return
111 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
112 Block* block = _cfg->get_block(i);
113 if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point?
114 Node* m = block->end();
115 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
116 MachEpilogNode* epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
117 block->add_inst(epilog);
118 _cfg->map_node_to_block(epilog, block);
119 }
120 }
121 }
123 # ifdef ENABLE_ZAP_DEAD_LOCALS
124 if (ZapDeadCompiledLocals) {
125 Insert_zap_nodes();
126 }
127 # endif
129 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1);
130 blk_starts[0] = 0;
132 // Initialize code buffer and process short branches.
133 CodeBuffer* cb = init_buffer(blk_starts);
135 if (cb == NULL || failing()) {
136 return;
137 }
139 ScheduleAndBundle();
141 #ifndef PRODUCT
142 if (trace_opto_output()) {
143 tty->print("\n---- After ScheduleAndBundle ----\n");
144 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
145 tty->print("\nBB#%03d:\n", i);
146 Block* block = _cfg->get_block(i);
147 for (uint j = 0; j < block->_nodes.size(); j++) {
148 Node* n = block->_nodes[j];
149 OptoReg::Name reg = _regalloc->get_reg_first(n);
150 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
151 n->dump();
152 }
153 }
154 }
155 #endif
157 if (failing()) {
158 return;
159 }
161 BuildOopMaps();
163 if (failing()) {
164 return;
165 }
167 fill_buffer(cb, blk_starts);
168 }
170 bool Compile::need_stack_bang(int frame_size_in_bytes) const {
171 // Determine if we need to generate a stack overflow check.
172 // Do it if the method is not a stub function and
173 // has java calls or has frame size > vm_page_size/8.
174 return (UseStackBanging && stub_function() == NULL &&
175 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3));
176 }
178 bool Compile::need_register_stack_bang() const {
179 // Determine if we need to generate a register stack overflow check.
180 // This is only used on architectures which have split register
181 // and memory stacks (ie. IA64).
182 // Bang if the method is not a stub function and has java calls
183 return (stub_function() == NULL && has_java_calls());
184 }
186 # ifdef ENABLE_ZAP_DEAD_LOCALS
189 // In order to catch compiler oop-map bugs, we have implemented
190 // a debugging mode called ZapDeadCompilerLocals.
191 // This mode causes the compiler to insert a call to a runtime routine,
192 // "zap_dead_locals", right before each place in compiled code
193 // that could potentially be a gc-point (i.e., a safepoint or oop map point).
194 // The runtime routine checks that locations mapped as oops are really
195 // oops, that locations mapped as values do not look like oops,
196 // and that locations mapped as dead are not used later
197 // (by zapping them to an invalid address).
199 int Compile::_CompiledZap_count = 0;
201 void Compile::Insert_zap_nodes() {
202 bool skip = false;
205 // Dink with static counts because code code without the extra
206 // runtime calls is MUCH faster for debugging purposes
208 if ( CompileZapFirst == 0 ) ; // nothing special
209 else if ( CompileZapFirst > CompiledZap_count() ) skip = true;
210 else if ( CompileZapFirst == CompiledZap_count() )
211 warning("starting zap compilation after skipping");
213 if ( CompileZapLast == -1 ) ; // nothing special
214 else if ( CompileZapLast < CompiledZap_count() ) skip = true;
215 else if ( CompileZapLast == CompiledZap_count() )
216 warning("about to compile last zap");
218 ++_CompiledZap_count; // counts skipped zaps, too
220 if ( skip ) return;
223 if ( _method == NULL )
224 return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care
226 // Insert call to zap runtime stub before every node with an oop map
227 for( uint i=0; i<_cfg->number_of_blocks(); i++ ) {
228 Block *b = _cfg->get_block(i);
229 for ( uint j = 0; j < b->_nodes.size(); ++j ) {
230 Node *n = b->_nodes[j];
232 // Determining if we should insert a zap-a-lot node in output.
233 // We do that for all nodes that has oopmap info, except for calls
234 // to allocation. Calls to allocation passes in the old top-of-eden pointer
235 // and expect the C code to reset it. Hence, there can be no safepoints between
236 // the inlined-allocation and the call to new_Java, etc.
237 // We also cannot zap monitor calls, as they must hold the microlock
238 // during the call to Zap, which also wants to grab the microlock.
239 bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL);
240 if ( insert ) { // it is MachSafePoint
241 if ( !n->is_MachCall() ) {
242 insert = false;
243 } else if ( n->is_MachCall() ) {
244 MachCallNode* call = n->as_MachCall();
245 if (call->entry_point() == OptoRuntime::new_instance_Java() ||
246 call->entry_point() == OptoRuntime::new_array_Java() ||
247 call->entry_point() == OptoRuntime::multianewarray2_Java() ||
248 call->entry_point() == OptoRuntime::multianewarray3_Java() ||
249 call->entry_point() == OptoRuntime::multianewarray4_Java() ||
250 call->entry_point() == OptoRuntime::multianewarray5_Java() ||
251 call->entry_point() == OptoRuntime::slow_arraycopy_Java() ||
252 call->entry_point() == OptoRuntime::complete_monitor_locking_Java()
253 ) {
254 insert = false;
255 }
256 }
257 if (insert) {
258 Node *zap = call_zap_node(n->as_MachSafePoint(), i);
259 b->_nodes.insert( j, zap );
260 _cfg->map_node_to_block(zap, b);
261 ++j;
262 }
263 }
264 }
265 }
266 }
269 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) {
270 const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type();
271 CallStaticJavaNode* ideal_node =
272 new (this) CallStaticJavaNode( tf,
273 OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()),
274 "call zap dead locals stub", 0, TypePtr::BOTTOM);
275 // We need to copy the OopMap from the site we're zapping at.
276 // We have to make a copy, because the zap site might not be
277 // a call site, and zap_dead is a call site.
278 OopMap* clone = node_to_check->oop_map()->deep_copy();
280 // Add the cloned OopMap to the zap node
281 ideal_node->set_oop_map(clone);
282 return _matcher->match_sfpt(ideal_node);
283 }
285 bool Compile::is_node_getting_a_safepoint( Node* n) {
286 // This code duplicates the logic prior to the call of add_safepoint
287 // below in this file.
288 if( n->is_MachSafePoint() ) return true;
289 return false;
290 }
292 # endif // ENABLE_ZAP_DEAD_LOCALS
294 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
295 // of a loop. When aligning a loop we need to provide enough instructions
296 // in cpu's fetch buffer to feed decoders. The loop alignment could be
297 // avoided if we have enough instructions in fetch buffer at the head of a loop.
298 // By default, the size is set to 999999 by Block's constructor so that
299 // a loop will be aligned if the size is not reset here.
300 //
301 // Note: Mach instructions could contain several HW instructions
302 // so the size is estimated only.
303 //
304 void Compile::compute_loop_first_inst_sizes() {
305 // The next condition is used to gate the loop alignment optimization.
306 // Don't aligned a loop if there are enough instructions at the head of a loop
307 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
308 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
309 // equal to 11 bytes which is the largest address NOP instruction.
310 if (MaxLoopPad < OptoLoopAlignment - 1) {
311 uint last_block = _cfg->number_of_blocks() - 1;
312 for (uint i = 1; i <= last_block; i++) {
313 Block* block = _cfg->get_block(i);
314 // Check the first loop's block which requires an alignment.
315 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
316 uint sum_size = 0;
317 uint inst_cnt = NumberOfLoopInstrToAlign;
318 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
320 // Check subsequent fallthrough blocks if the loop's first
321 // block(s) does not have enough instructions.
322 Block *nb = block;
323 while(inst_cnt > 0 &&
324 i < last_block &&
325 !_cfg->get_block(i + 1)->has_loop_alignment() &&
326 !nb->has_successor(block)) {
327 i++;
328 nb = _cfg->get_block(i);
329 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
330 } // while( inst_cnt > 0 && i < last_block )
332 block->set_first_inst_size(sum_size);
333 } // f( b->head()->is_Loop() )
334 } // for( i <= last_block )
335 } // if( MaxLoopPad < OptoLoopAlignment-1 )
336 }
338 // The architecture description provides short branch variants for some long
339 // branch instructions. Replace eligible long branches with short branches.
340 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) {
341 // Compute size of each block, method size, and relocation information size
342 uint nblocks = _cfg->number_of_blocks();
344 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
345 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
346 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
347 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
348 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
350 bool has_short_branch_candidate = false;
352 // Initialize the sizes to 0
353 code_size = 0; // Size in bytes of generated code
354 stub_size = 0; // Size in bytes of all stub entries
355 // Size in bytes of all relocation entries, including those in local stubs.
356 // Start with 2-bytes of reloc info for the unvalidated entry point
357 reloc_size = 1; // Number of relocation entries
359 // Make three passes. The first computes pessimistic blk_starts,
360 // relative jmp_offset and reloc_size information. The second performs
361 // short branch substitution using the pessimistic sizing. The
362 // third inserts nops where needed.
364 // Step one, perform a pessimistic sizing pass.
365 uint last_call_adr = max_uint;
366 uint last_avoid_back_to_back_adr = max_uint;
367 uint nop_size = (new (this) MachNopNode())->size(_regalloc);
368 for (uint i = 0; i < nblocks; i++) { // For all blocks
369 Block* block = _cfg->get_block(i);
371 // During short branch replacement, we store the relative (to blk_starts)
372 // offset of jump in jmp_offset, rather than the absolute offset of jump.
373 // This is so that we do not need to recompute sizes of all nodes when
374 // we compute correct blk_starts in our next sizing pass.
375 jmp_offset[i] = 0;
376 jmp_size[i] = 0;
377 jmp_nidx[i] = -1;
378 DEBUG_ONLY( jmp_target[i] = 0; )
379 DEBUG_ONLY( jmp_rule[i] = 0; )
381 // Sum all instruction sizes to compute block size
382 uint last_inst = block->_nodes.size();
383 uint blk_size = 0;
384 for (uint j = 0; j < last_inst; j++) {
385 Node* nj = block->_nodes[j];
386 // Handle machine instruction nodes
387 if (nj->is_Mach()) {
388 MachNode *mach = nj->as_Mach();
389 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
390 reloc_size += mach->reloc();
391 if (mach->is_MachCall()) {
392 MachCallNode *mcall = mach->as_MachCall();
393 // This destination address is NOT PC-relative
395 mcall->method_set((intptr_t)mcall->entry_point());
397 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
398 stub_size += CompiledStaticCall::to_interp_stub_size();
399 reloc_size += CompiledStaticCall::reloc_to_interp_stub();
400 }
401 } else if (mach->is_MachSafePoint()) {
402 // If call/safepoint are adjacent, account for possible
403 // nop to disambiguate the two safepoints.
404 // ScheduleAndBundle() can rearrange nodes in a block,
405 // check for all offsets inside this block.
406 if (last_call_adr >= blk_starts[i]) {
407 blk_size += nop_size;
408 }
409 }
410 if (mach->avoid_back_to_back()) {
411 // Nop is inserted between "avoid back to back" instructions.
412 // ScheduleAndBundle() can rearrange nodes in a block,
413 // check for all offsets inside this block.
414 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
415 blk_size += nop_size;
416 }
417 }
418 if (mach->may_be_short_branch()) {
419 if (!nj->is_MachBranch()) {
420 #ifndef PRODUCT
421 nj->dump(3);
422 #endif
423 Unimplemented();
424 }
425 assert(jmp_nidx[i] == -1, "block should have only one branch");
426 jmp_offset[i] = blk_size;
427 jmp_size[i] = nj->size(_regalloc);
428 jmp_nidx[i] = j;
429 has_short_branch_candidate = true;
430 }
431 }
432 blk_size += nj->size(_regalloc);
433 // Remember end of call offset
434 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
435 last_call_adr = blk_starts[i]+blk_size;
436 }
437 // Remember end of avoid_back_to_back offset
438 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back()) {
439 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
440 }
441 }
443 // When the next block starts a loop, we may insert pad NOP
444 // instructions. Since we cannot know our future alignment,
445 // assume the worst.
446 if (i < nblocks - 1) {
447 Block* nb = _cfg->get_block(i + 1);
448 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
449 if (max_loop_pad > 0) {
450 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
451 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
452 // If either is the last instruction in this block, bump by
453 // max_loop_pad in lock-step with blk_size, so sizing
454 // calculations in subsequent blocks still can conservatively
455 // detect that it may the last instruction in this block.
456 if (last_call_adr == blk_starts[i]+blk_size) {
457 last_call_adr += max_loop_pad;
458 }
459 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
460 last_avoid_back_to_back_adr += max_loop_pad;
461 }
462 blk_size += max_loop_pad;
463 }
464 }
466 // Save block size; update total method size
467 blk_starts[i+1] = blk_starts[i]+blk_size;
468 }
470 // Step two, replace eligible long jumps.
471 bool progress = true;
472 uint last_may_be_short_branch_adr = max_uint;
473 while (has_short_branch_candidate && progress) {
474 progress = false;
475 has_short_branch_candidate = false;
476 int adjust_block_start = 0;
477 for (uint i = 0; i < nblocks; i++) {
478 Block* block = _cfg->get_block(i);
479 int idx = jmp_nidx[i];
480 MachNode* mach = (idx == -1) ? NULL: block->_nodes[idx]->as_Mach();
481 if (mach != NULL && mach->may_be_short_branch()) {
482 #ifdef ASSERT
483 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
484 int j;
485 // Find the branch; ignore trailing NOPs.
486 for (j = block->_nodes.size()-1; j>=0; j--) {
487 Node* n = block->_nodes[j];
488 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
489 break;
490 }
491 assert(j >= 0 && j == idx && block->_nodes[j] == (Node*)mach, "sanity");
492 #endif
493 int br_size = jmp_size[i];
494 int br_offs = blk_starts[i] + jmp_offset[i];
496 // This requires the TRUE branch target be in succs[0]
497 uint bnum = block->non_connector_successor(0)->_pre_order;
498 int offset = blk_starts[bnum] - br_offs;
499 if (bnum > i) { // adjust following block's offset
500 offset -= adjust_block_start;
501 }
502 // In the following code a nop could be inserted before
503 // the branch which will increase the backward distance.
504 bool needs_padding = ((uint)br_offs == last_may_be_short_branch_adr);
505 if (needs_padding && offset <= 0)
506 offset -= nop_size;
508 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
509 // We've got a winner. Replace this branch.
510 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
512 // Update the jmp_size.
513 int new_size = replacement->size(_regalloc);
514 int diff = br_size - new_size;
515 assert(diff >= (int)nop_size, "short_branch size should be smaller");
516 // Conservatively take into accound padding between
517 // avoid_back_to_back branches. Previous branch could be
518 // converted into avoid_back_to_back branch during next
519 // rounds.
520 if (needs_padding && replacement->avoid_back_to_back()) {
521 jmp_offset[i] += nop_size;
522 diff -= nop_size;
523 }
524 adjust_block_start += diff;
525 block->_nodes.map(idx, replacement);
526 mach->subsume_by(replacement, C);
527 mach = replacement;
528 progress = true;
530 jmp_size[i] = new_size;
531 DEBUG_ONLY( jmp_target[i] = bnum; );
532 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
533 } else {
534 // The jump distance is not short, try again during next iteration.
535 has_short_branch_candidate = true;
536 }
537 } // (mach->may_be_short_branch())
538 if (mach != NULL && (mach->may_be_short_branch() ||
539 mach->avoid_back_to_back())) {
540 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
541 }
542 blk_starts[i+1] -= adjust_block_start;
543 }
544 }
546 #ifdef ASSERT
547 for (uint i = 0; i < nblocks; i++) { // For all blocks
548 if (jmp_target[i] != 0) {
549 int br_size = jmp_size[i];
550 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
551 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
552 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]);
553 }
554 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
555 }
556 }
557 #endif
559 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
560 // after ScheduleAndBundle().
562 // ------------------
563 // Compute size for code buffer
564 code_size = blk_starts[nblocks];
566 // Relocation records
567 reloc_size += 1; // Relo entry for exception handler
569 // Adjust reloc_size to number of record of relocation info
570 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
571 // a relocation index.
572 // The CodeBuffer will expand the locs array if this estimate is too low.
573 reloc_size *= 10 / sizeof(relocInfo);
574 }
576 //------------------------------FillLocArray-----------------------------------
577 // Create a bit of debug info and append it to the array. The mapping is from
578 // Java local or expression stack to constant, register or stack-slot. For
579 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
580 // entry has been taken care of and caller should skip it).
581 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
582 // This should never have accepted Bad before
583 assert(OptoReg::is_valid(regnum), "location must be valid");
584 return (OptoReg::is_reg(regnum))
585 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
586 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
587 }
590 ObjectValue*
591 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
592 for (int i = 0; i < objs->length(); i++) {
593 assert(objs->at(i)->is_object(), "corrupt object cache");
594 ObjectValue* sv = (ObjectValue*) objs->at(i);
595 if (sv->id() == id) {
596 return sv;
597 }
598 }
599 // Otherwise..
600 return NULL;
601 }
603 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
604 ObjectValue* sv ) {
605 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
606 objs->append(sv);
607 }
610 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
611 GrowableArray<ScopeValue*> *array,
612 GrowableArray<ScopeValue*> *objs ) {
613 assert( local, "use _top instead of null" );
614 if (array->length() != idx) {
615 assert(array->length() == idx + 1, "Unexpected array count");
616 // Old functionality:
617 // return
618 // New functionality:
619 // Assert if the local is not top. In product mode let the new node
620 // override the old entry.
621 assert(local == top(), "LocArray collision");
622 if (local == top()) {
623 return;
624 }
625 array->pop();
626 }
627 const Type *t = local->bottom_type();
629 // Is it a safepoint scalar object node?
630 if (local->is_SafePointScalarObject()) {
631 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
633 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
634 if (sv == NULL) {
635 ciKlass* cik = t->is_oopptr()->klass();
636 assert(cik->is_instance_klass() ||
637 cik->is_array_klass(), "Not supported allocation.");
638 sv = new ObjectValue(spobj->_idx,
639 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
640 Compile::set_sv_for_object_node(objs, sv);
642 uint first_ind = spobj->first_index();
643 for (uint i = 0; i < spobj->n_fields(); i++) {
644 Node* fld_node = sfpt->in(first_ind+i);
645 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
646 }
647 }
648 array->append(sv);
649 return;
650 }
652 // Grab the register number for the local
653 OptoReg::Name regnum = _regalloc->get_reg_first(local);
654 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
655 // Record the double as two float registers.
656 // The register mask for such a value always specifies two adjacent
657 // float registers, with the lower register number even.
658 // Normally, the allocation of high and low words to these registers
659 // is irrelevant, because nearly all operations on register pairs
660 // (e.g., StoreD) treat them as a single unit.
661 // Here, we assume in addition that the words in these two registers
662 // stored "naturally" (by operations like StoreD and double stores
663 // within the interpreter) such that the lower-numbered register
664 // is written to the lower memory address. This may seem like
665 // a machine dependency, but it is not--it is a requirement on
666 // the author of the <arch>.ad file to ensure that, for every
667 // even/odd double-register pair to which a double may be allocated,
668 // the word in the even single-register is stored to the first
669 // memory word. (Note that register numbers are completely
670 // arbitrary, and are not tied to any machine-level encodings.)
671 #ifdef _LP64
672 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
673 array->append(new ConstantIntValue(0));
674 array->append(new_loc_value( _regalloc, regnum, Location::dbl ));
675 } else if ( t->base() == Type::Long ) {
676 array->append(new ConstantIntValue(0));
677 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
678 } else if ( t->base() == Type::RawPtr ) {
679 // jsr/ret return address which must be restored into a the full
680 // width 64-bit stack slot.
681 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
682 }
683 #else //_LP64
684 #ifdef SPARC
685 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {
686 // For SPARC we have to swap high and low words for
687 // long values stored in a single-register (g0-g7).
688 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
689 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
690 } else
691 #endif //SPARC
692 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
693 // Repack the double/long as two jints.
694 // The convention the interpreter uses is that the second local
695 // holds the first raw word of the native double representation.
696 // This is actually reasonable, since locals and stack arrays
697 // grow downwards in all implementations.
698 // (If, on some machine, the interpreter's Java locals or stack
699 // were to grow upwards, the embedded doubles would be word-swapped.)
700 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
701 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
702 }
703 #endif //_LP64
704 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
705 OptoReg::is_reg(regnum) ) {
706 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()
707 ? Location::float_in_dbl : Location::normal ));
708 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
709 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long
710 ? Location::int_in_long : Location::normal ));
711 } else if( t->base() == Type::NarrowOop ) {
712 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));
713 } else {
714 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));
715 }
716 return;
717 }
719 // No register. It must be constant data.
720 switch (t->base()) {
721 case Type::Half: // Second half of a double
722 ShouldNotReachHere(); // Caller should skip 2nd halves
723 break;
724 case Type::AnyPtr:
725 array->append(new ConstantOopWriteValue(NULL));
726 break;
727 case Type::AryPtr:
728 case Type::InstPtr: // fall through
729 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
730 break;
731 case Type::NarrowOop:
732 if (t == TypeNarrowOop::NULL_PTR) {
733 array->append(new ConstantOopWriteValue(NULL));
734 } else {
735 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
736 }
737 break;
738 case Type::Int:
739 array->append(new ConstantIntValue(t->is_int()->get_con()));
740 break;
741 case Type::RawPtr:
742 // A return address (T_ADDRESS).
743 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
744 #ifdef _LP64
745 // Must be restored to the full-width 64-bit stack slot.
746 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
747 #else
748 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
749 #endif
750 break;
751 case Type::FloatCon: {
752 float f = t->is_float_constant()->getf();
753 array->append(new ConstantIntValue(jint_cast(f)));
754 break;
755 }
756 case Type::DoubleCon: {
757 jdouble d = t->is_double_constant()->getd();
758 #ifdef _LP64
759 array->append(new ConstantIntValue(0));
760 array->append(new ConstantDoubleValue(d));
761 #else
762 // Repack the double as two jints.
763 // The convention the interpreter uses is that the second local
764 // holds the first raw word of the native double representation.
765 // This is actually reasonable, since locals and stack arrays
766 // grow downwards in all implementations.
767 // (If, on some machine, the interpreter's Java locals or stack
768 // were to grow upwards, the embedded doubles would be word-swapped.)
769 jint *dp = (jint*)&d;
770 array->append(new ConstantIntValue(dp[1]));
771 array->append(new ConstantIntValue(dp[0]));
772 #endif
773 break;
774 }
775 case Type::Long: {
776 jlong d = t->is_long()->get_con();
777 #ifdef _LP64
778 array->append(new ConstantIntValue(0));
779 array->append(new ConstantLongValue(d));
780 #else
781 // Repack the long as two jints.
782 // The convention the interpreter uses is that the second local
783 // holds the first raw word of the native double representation.
784 // This is actually reasonable, since locals and stack arrays
785 // grow downwards in all implementations.
786 // (If, on some machine, the interpreter's Java locals or stack
787 // were to grow upwards, the embedded doubles would be word-swapped.)
788 jint *dp = (jint*)&d;
789 array->append(new ConstantIntValue(dp[1]));
790 array->append(new ConstantIntValue(dp[0]));
791 #endif
792 break;
793 }
794 case Type::Top: // Add an illegal value here
795 array->append(new LocationValue(Location()));
796 break;
797 default:
798 ShouldNotReachHere();
799 break;
800 }
801 }
803 // Determine if this node starts a bundle
804 bool Compile::starts_bundle(const Node *n) const {
805 return (_node_bundling_limit > n->_idx &&
806 _node_bundling_base[n->_idx].starts_bundle());
807 }
809 //--------------------------Process_OopMap_Node--------------------------------
810 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
812 // Handle special safepoint nodes for synchronization
813 MachSafePointNode *sfn = mach->as_MachSafePoint();
814 MachCallNode *mcall;
816 #ifdef ENABLE_ZAP_DEAD_LOCALS
817 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative");
818 #endif
820 int safepoint_pc_offset = current_offset;
821 bool is_method_handle_invoke = false;
822 bool return_oop = false;
824 // Add the safepoint in the DebugInfoRecorder
825 if( !mach->is_MachCall() ) {
826 mcall = NULL;
827 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
828 } else {
829 mcall = mach->as_MachCall();
831 // Is the call a MethodHandle call?
832 if (mcall->is_MachCallJava()) {
833 if (mcall->as_MachCallJava()->_method_handle_invoke) {
834 assert(has_method_handle_invokes(), "must have been set during call generation");
835 is_method_handle_invoke = true;
836 }
837 }
839 // Check if a call returns an object.
840 if (mcall->return_value_is_used() &&
841 mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) {
842 return_oop = true;
843 }
844 safepoint_pc_offset += mcall->ret_addr_offset();
845 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
846 }
848 // Loop over the JVMState list to add scope information
849 // Do not skip safepoints with a NULL method, they need monitor info
850 JVMState* youngest_jvms = sfn->jvms();
851 int max_depth = youngest_jvms->depth();
853 // Allocate the object pool for scalar-replaced objects -- the map from
854 // small-integer keys (which can be recorded in the local and ostack
855 // arrays) to descriptions of the object state.
856 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
858 // Visit scopes from oldest to youngest.
859 for (int depth = 1; depth <= max_depth; depth++) {
860 JVMState* jvms = youngest_jvms->of_depth(depth);
861 int idx;
862 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
863 // Safepoints that do not have method() set only provide oop-map and monitor info
864 // to support GC; these do not support deoptimization.
865 int num_locs = (method == NULL) ? 0 : jvms->loc_size();
866 int num_exps = (method == NULL) ? 0 : jvms->stk_size();
867 int num_mon = jvms->nof_monitors();
868 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
869 "JVMS local count must match that of the method");
871 // Add Local and Expression Stack Information
873 // Insert locals into the locarray
874 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
875 for( idx = 0; idx < num_locs; idx++ ) {
876 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
877 }
879 // Insert expression stack entries into the exparray
880 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
881 for( idx = 0; idx < num_exps; idx++ ) {
882 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
883 }
885 // Add in mappings of the monitors
886 assert( !method ||
887 !method->is_synchronized() ||
888 method->is_native() ||
889 num_mon > 0 ||
890 !GenerateSynchronizationCode,
891 "monitors must always exist for synchronized methods");
893 // Build the growable array of ScopeValues for exp stack
894 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
896 // Loop over monitors and insert into array
897 for(idx = 0; idx < num_mon; idx++) {
898 // Grab the node that defines this monitor
899 Node* box_node = sfn->monitor_box(jvms, idx);
900 Node* obj_node = sfn->monitor_obj(jvms, idx);
902 // Create ScopeValue for object
903 ScopeValue *scval = NULL;
905 if( obj_node->is_SafePointScalarObject() ) {
906 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
907 scval = Compile::sv_for_node_id(objs, spobj->_idx);
908 if (scval == NULL) {
909 const Type *t = obj_node->bottom_type();
910 ciKlass* cik = t->is_oopptr()->klass();
911 assert(cik->is_instance_klass() ||
912 cik->is_array_klass(), "Not supported allocation.");
913 ObjectValue* sv = new ObjectValue(spobj->_idx,
914 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
915 Compile::set_sv_for_object_node(objs, sv);
917 uint first_ind = spobj->first_index();
918 for (uint i = 0; i < spobj->n_fields(); i++) {
919 Node* fld_node = sfn->in(first_ind+i);
920 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
921 }
922 scval = sv;
923 }
924 } else if( !obj_node->is_Con() ) {
925 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);
926 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
927 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );
928 } else {
929 scval = new_loc_value( _regalloc, obj_reg, Location::oop );
930 }
931 } else {
932 const TypePtr *tp = obj_node->get_ptr_type();
933 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
934 }
936 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
937 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
938 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
939 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
940 }
942 // We dump the object pool first, since deoptimization reads it in first.
943 debug_info()->dump_object_pool(objs);
945 // Build first class objects to pass to scope
946 DebugToken *locvals = debug_info()->create_scope_values(locarray);
947 DebugToken *expvals = debug_info()->create_scope_values(exparray);
948 DebugToken *monvals = debug_info()->create_monitor_values(monarray);
950 // Make method available for all Safepoints
951 ciMethod* scope_method = method ? method : _method;
952 // Describe the scope here
953 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
954 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
955 // Now we can describe the scope.
956 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals);
957 } // End jvms loop
959 // Mark the end of the scope set.
960 debug_info()->end_safepoint(safepoint_pc_offset);
961 }
965 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
966 class NonSafepointEmitter {
967 Compile* C;
968 JVMState* _pending_jvms;
969 int _pending_offset;
971 void emit_non_safepoint();
973 public:
974 NonSafepointEmitter(Compile* compile) {
975 this->C = compile;
976 _pending_jvms = NULL;
977 _pending_offset = 0;
978 }
980 void observe_instruction(Node* n, int pc_offset) {
981 if (!C->debug_info()->recording_non_safepoints()) return;
983 Node_Notes* nn = C->node_notes_at(n->_idx);
984 if (nn == NULL || nn->jvms() == NULL) return;
985 if (_pending_jvms != NULL &&
986 _pending_jvms->same_calls_as(nn->jvms())) {
987 // Repeated JVMS? Stretch it up here.
988 _pending_offset = pc_offset;
989 } else {
990 if (_pending_jvms != NULL &&
991 _pending_offset < pc_offset) {
992 emit_non_safepoint();
993 }
994 _pending_jvms = NULL;
995 if (pc_offset > C->debug_info()->last_pc_offset()) {
996 // This is the only way _pending_jvms can become non-NULL:
997 _pending_jvms = nn->jvms();
998 _pending_offset = pc_offset;
999 }
1000 }
1001 }
1003 // Stay out of the way of real safepoints:
1004 void observe_safepoint(JVMState* jvms, int pc_offset) {
1005 if (_pending_jvms != NULL &&
1006 !_pending_jvms->same_calls_as(jvms) &&
1007 _pending_offset < pc_offset) {
1008 emit_non_safepoint();
1009 }
1010 _pending_jvms = NULL;
1011 }
1013 void flush_at_end() {
1014 if (_pending_jvms != NULL) {
1015 emit_non_safepoint();
1016 }
1017 _pending_jvms = NULL;
1018 }
1019 };
1021 void NonSafepointEmitter::emit_non_safepoint() {
1022 JVMState* youngest_jvms = _pending_jvms;
1023 int pc_offset = _pending_offset;
1025 // Clear it now:
1026 _pending_jvms = NULL;
1028 DebugInformationRecorder* debug_info = C->debug_info();
1029 assert(debug_info->recording_non_safepoints(), "sanity");
1031 debug_info->add_non_safepoint(pc_offset);
1032 int max_depth = youngest_jvms->depth();
1034 // Visit scopes from oldest to youngest.
1035 for (int depth = 1; depth <= max_depth; depth++) {
1036 JVMState* jvms = youngest_jvms->of_depth(depth);
1037 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
1038 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1039 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute());
1040 }
1042 // Mark the end of the scope set.
1043 debug_info->end_non_safepoint(pc_offset);
1044 }
1046 //------------------------------init_buffer------------------------------------
1047 CodeBuffer* Compile::init_buffer(uint* blk_starts) {
1049 // Set the initially allocated size
1050 int code_req = initial_code_capacity;
1051 int locs_req = initial_locs_capacity;
1052 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity;
1053 int const_req = initial_const_capacity;
1055 int pad_req = NativeCall::instruction_size;
1056 // The extra spacing after the code is necessary on some platforms.
1057 // Sometimes we need to patch in a jump after the last instruction,
1058 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1060 // Compute the byte offset where we can store the deopt pc.
1061 if (fixed_slots() != 0) {
1062 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1063 }
1065 // Compute prolog code size
1066 _method_size = 0;
1067 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;
1068 #if defined(IA64) && !defined(AIX)
1069 if (save_argument_registers()) {
1070 // 4815101: this is a stub with implicit and unknown precision fp args.
1071 // The usual spill mechanism can only generate stfd's in this case, which
1072 // doesn't work if the fp reg to spill contains a single-precision denorm.
1073 // Instead, we hack around the normal spill mechanism using stfspill's and
1074 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate
1075 // space here for the fp arg regs (f8-f15) we're going to thusly spill.
1076 //
1077 // If we ever implement 16-byte 'registers' == stack slots, we can
1078 // get rid of this hack and have SpillCopy generate stfspill/ldffill
1079 // instead of stfd/stfs/ldfd/ldfs.
1080 _frame_slots += 8*(16/BytesPerInt);
1081 }
1082 #endif
1083 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1085 if (has_mach_constant_base_node()) {
1086 // Fill the constant table.
1087 // Note: This must happen before shorten_branches.
1088 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1089 Block* b = _cfg->get_block(i);
1091 for (uint j = 0; j < b->_nodes.size(); j++) {
1092 Node* n = b->_nodes[j];
1094 // If the node is a MachConstantNode evaluate the constant
1095 // value section.
1096 if (n->is_MachConstant()) {
1097 MachConstantNode* machcon = n->as_MachConstant();
1098 machcon->eval_constant(C);
1099 }
1100 }
1101 }
1103 // Calculate the offsets of the constants and the size of the
1104 // constant table (including the padding to the next section).
1105 constant_table().calculate_offsets_and_size();
1106 const_req = constant_table().size();
1107 }
1109 // Initialize the space for the BufferBlob used to find and verify
1110 // instruction size in MachNode::emit_size()
1111 init_scratch_buffer_blob(const_req);
1112 if (failing()) return NULL; // Out of memory
1114 // Pre-compute the length of blocks and replace
1115 // long branches with short if machine supports it.
1116 shorten_branches(blk_starts, code_req, locs_req, stub_req);
1118 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1119 int exception_handler_req = size_exception_handler();
1120 int deopt_handler_req = size_deopt_handler();
1121 exception_handler_req += MAX_stubs_size; // add marginal slop for handler
1122 deopt_handler_req += MAX_stubs_size; // add marginal slop for handler
1123 stub_req += MAX_stubs_size; // ensure per-stub margin
1124 code_req += MAX_inst_size; // ensure per-instruction margin
1126 if (StressCodeBuffers)
1127 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1129 int total_req =
1130 const_req +
1131 code_req +
1132 pad_req +
1133 stub_req +
1134 exception_handler_req +
1135 deopt_handler_req; // deopt handler
1137 if (has_method_handle_invokes())
1138 total_req += deopt_handler_req; // deopt MH handler
1140 CodeBuffer* cb = code_buffer();
1141 cb->initialize(total_req, locs_req);
1143 // Have we run out of code space?
1144 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1145 C->record_failure("CodeCache is full");
1146 return NULL;
1147 }
1148 // Configure the code buffer.
1149 cb->initialize_consts_size(const_req);
1150 cb->initialize_stubs_size(stub_req);
1151 cb->initialize_oop_recorder(env()->oop_recorder());
1153 // fill in the nop array for bundling computations
1154 MachNode *_nop_list[Bundle::_nop_count];
1155 Bundle::initialize_nops(_nop_list, this);
1157 return cb;
1158 }
1160 //------------------------------fill_buffer------------------------------------
1161 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1162 // blk_starts[] contains offsets calculated during short branches processing,
1163 // offsets should not be increased during following steps.
1165 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1166 // of a loop. It is used to determine the padding for loop alignment.
1167 compute_loop_first_inst_sizes();
1169 // Create oopmap set.
1170 _oop_map_set = new OopMapSet();
1172 // !!!!! This preserves old handling of oopmaps for now
1173 debug_info()->set_oopmaps(_oop_map_set);
1175 uint nblocks = _cfg->number_of_blocks();
1176 // Count and start of implicit null check instructions
1177 uint inct_cnt = 0;
1178 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1180 // Count and start of calls
1181 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1183 uint return_offset = 0;
1184 int nop_size = (new (this) MachNopNode())->size(_regalloc);
1186 int previous_offset = 0;
1187 int current_offset = 0;
1188 int last_call_offset = -1;
1189 int last_avoid_back_to_back_offset = -1;
1190 #ifdef ASSERT
1191 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1192 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1193 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1194 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1195 #endif
1197 // Create an array of unused labels, one for each basic block, if printing is enabled
1198 #ifndef PRODUCT
1199 int *node_offsets = NULL;
1200 uint node_offset_limit = unique();
1202 if (print_assembly())
1203 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1204 #endif
1206 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily
1208 // Emit the constant table.
1209 if (has_mach_constant_base_node()) {
1210 constant_table().emit(*cb);
1211 }
1213 // Create an array of labels, one for each basic block
1214 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1215 for (uint i=0; i <= nblocks; i++) {
1216 blk_labels[i].init();
1217 }
1219 // ------------------
1220 // Now fill in the code buffer
1221 Node *delay_slot = NULL;
1223 for (uint i = 0; i < nblocks; i++) {
1224 Block* block = _cfg->get_block(i);
1225 Node* head = block->head();
1227 // If this block needs to start aligned (i.e, can be reached other
1228 // than by falling-thru from the previous block), then force the
1229 // start of a new bundle.
1230 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1231 cb->flush_bundle(true);
1232 }
1234 #ifdef ASSERT
1235 if (!block->is_connector()) {
1236 stringStream st;
1237 block->dump_head(_cfg, &st);
1238 MacroAssembler(cb).block_comment(st.as_string());
1239 }
1240 jmp_target[i] = 0;
1241 jmp_offset[i] = 0;
1242 jmp_size[i] = 0;
1243 jmp_rule[i] = 0;
1244 #endif
1245 int blk_offset = current_offset;
1247 // Define the label at the beginning of the basic block
1248 MacroAssembler(cb).bind(blk_labels[block->_pre_order]);
1250 uint last_inst = block->_nodes.size();
1252 // Emit block normally, except for last instruction.
1253 // Emit means "dump code bits into code buffer".
1254 for (uint j = 0; j<last_inst; j++) {
1256 // Get the node
1257 Node* n = block->_nodes[j];
1259 // See if delay slots are supported
1260 if (valid_bundle_info(n) &&
1261 node_bundling(n)->used_in_unconditional_delay()) {
1262 assert(delay_slot == NULL, "no use of delay slot node");
1263 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1265 delay_slot = n;
1266 continue;
1267 }
1269 // If this starts a new instruction group, then flush the current one
1270 // (but allow split bundles)
1271 if (Pipeline::requires_bundling() && starts_bundle(n))
1272 cb->flush_bundle(false);
1274 // The following logic is duplicated in the code ifdeffed for
1275 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It
1276 // should be factored out. Or maybe dispersed to the nodes?
1278 // Special handling for SafePoint/Call Nodes
1279 bool is_mcall = false;
1280 if (n->is_Mach()) {
1281 MachNode *mach = n->as_Mach();
1282 is_mcall = n->is_MachCall();
1283 bool is_sfn = n->is_MachSafePoint();
1285 // If this requires all previous instructions be flushed, then do so
1286 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1287 cb->flush_bundle(true);
1288 current_offset = cb->insts_size();
1289 }
1291 // A padding may be needed again since a previous instruction
1292 // could be moved to delay slot.
1294 // align the instruction if necessary
1295 int padding = mach->compute_padding(current_offset);
1296 // Make sure safepoint node for polling is distinct from a call's
1297 // return by adding a nop if needed.
1298 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1299 padding = nop_size;
1300 }
1301 if (padding == 0 && mach->avoid_back_to_back() &&
1302 current_offset == last_avoid_back_to_back_offset) {
1303 // Avoid back to back some instructions.
1304 padding = nop_size;
1305 }
1307 if(padding > 0) {
1308 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1309 int nops_cnt = padding / nop_size;
1310 MachNode *nop = new (this) MachNopNode(nops_cnt);
1311 block->_nodes.insert(j++, nop);
1312 last_inst++;
1313 _cfg->map_node_to_block(nop, block);
1314 nop->emit(*cb, _regalloc);
1315 cb->flush_bundle(true);
1316 current_offset = cb->insts_size();
1317 }
1319 // Remember the start of the last call in a basic block
1320 if (is_mcall) {
1321 MachCallNode *mcall = mach->as_MachCall();
1323 // This destination address is NOT PC-relative
1324 mcall->method_set((intptr_t)mcall->entry_point());
1326 // Save the return address
1327 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1329 if (mcall->is_MachCallLeaf()) {
1330 is_mcall = false;
1331 is_sfn = false;
1332 }
1333 }
1335 // sfn will be valid whenever mcall is valid now because of inheritance
1336 if (is_sfn || is_mcall) {
1338 // Handle special safepoint nodes for synchronization
1339 if (!is_mcall) {
1340 MachSafePointNode *sfn = mach->as_MachSafePoint();
1341 // !!!!! Stubs only need an oopmap right now, so bail out
1342 if (sfn->jvms()->method() == NULL) {
1343 // Write the oopmap directly to the code blob??!!
1344 # ifdef ENABLE_ZAP_DEAD_LOCALS
1345 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive");
1346 # endif
1347 continue;
1348 }
1349 } // End synchronization
1351 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1352 current_offset);
1353 Process_OopMap_Node(mach, current_offset);
1354 } // End if safepoint
1356 // If this is a null check, then add the start of the previous instruction to the list
1357 else if( mach->is_MachNullCheck() ) {
1358 inct_starts[inct_cnt++] = previous_offset;
1359 }
1361 // If this is a branch, then fill in the label with the target BB's label
1362 else if (mach->is_MachBranch()) {
1363 // This requires the TRUE branch target be in succs[0]
1364 uint block_num = block->non_connector_successor(0)->_pre_order;
1366 // Try to replace long branch if delay slot is not used,
1367 // it is mostly for back branches since forward branch's
1368 // distance is not updated yet.
1369 bool delay_slot_is_used = valid_bundle_info(n) &&
1370 node_bundling(n)->use_unconditional_delay();
1371 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1372 assert(delay_slot == NULL, "not expecting delay slot node");
1373 int br_size = n->size(_regalloc);
1374 int offset = blk_starts[block_num] - current_offset;
1375 if (block_num >= i) {
1376 // Current and following block's offset are not
1377 // finilized yet, adjust distance by the difference
1378 // between calculated and final offsets of current block.
1379 offset -= (blk_starts[i] - blk_offset);
1380 }
1381 // In the following code a nop could be inserted before
1382 // the branch which will increase the backward distance.
1383 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1384 if (needs_padding && offset <= 0)
1385 offset -= nop_size;
1387 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
1388 // We've got a winner. Replace this branch.
1389 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
1391 // Update the jmp_size.
1392 int new_size = replacement->size(_regalloc);
1393 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1394 // Insert padding between avoid_back_to_back branches.
1395 if (needs_padding && replacement->avoid_back_to_back()) {
1396 MachNode *nop = new (this) MachNopNode();
1397 block->_nodes.insert(j++, nop);
1398 _cfg->map_node_to_block(nop, block);
1399 last_inst++;
1400 nop->emit(*cb, _regalloc);
1401 cb->flush_bundle(true);
1402 current_offset = cb->insts_size();
1403 }
1404 #ifdef ASSERT
1405 jmp_target[i] = block_num;
1406 jmp_offset[i] = current_offset - blk_offset;
1407 jmp_size[i] = new_size;
1408 jmp_rule[i] = mach->rule();
1409 #endif
1410 block->_nodes.map(j, replacement);
1411 mach->subsume_by(replacement, C);
1412 n = replacement;
1413 mach = replacement;
1414 }
1415 }
1416 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1417 } else if (mach->ideal_Opcode() == Op_Jump) {
1418 for (uint h = 0; h < block->_num_succs; h++) {
1419 Block* succs_block = block->_succs[h];
1420 for (uint j = 1; j < succs_block->num_preds(); j++) {
1421 Node* jpn = succs_block->pred(j);
1422 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1423 uint block_num = succs_block->non_connector()->_pre_order;
1424 Label *blkLabel = &blk_labels[block_num];
1425 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1426 }
1427 }
1428 }
1429 }
1430 #ifdef ASSERT
1431 // Check that oop-store precedes the card-mark
1432 else if (mach->ideal_Opcode() == Op_StoreCM) {
1433 uint storeCM_idx = j;
1434 int count = 0;
1435 for (uint prec = mach->req(); prec < mach->len(); prec++) {
1436 Node *oop_store = mach->in(prec); // Precedence edge
1437 if (oop_store == NULL) continue;
1438 count++;
1439 uint i4;
1440 for (i4 = 0; i4 < last_inst; ++i4) {
1441 if (block->_nodes[i4] == oop_store) {
1442 break;
1443 }
1444 }
1445 // Note: This test can provide a false failure if other precedence
1446 // edges have been added to the storeCMNode.
1447 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1448 }
1449 assert(count > 0, "storeCM expects at least one precedence edge");
1450 }
1451 #endif
1452 else if (!n->is_Proj()) {
1453 // Remember the beginning of the previous instruction, in case
1454 // it's followed by a flag-kill and a null-check. Happens on
1455 // Intel all the time, with add-to-memory kind of opcodes.
1456 previous_offset = current_offset;
1457 }
1458 }
1460 // Verify that there is sufficient space remaining
1461 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1462 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1463 C->record_failure("CodeCache is full");
1464 return;
1465 }
1467 // Save the offset for the listing
1468 #ifndef PRODUCT
1469 if (node_offsets && n->_idx < node_offset_limit)
1470 node_offsets[n->_idx] = cb->insts_size();
1471 #endif
1473 // "Normal" instruction case
1474 DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
1475 n->emit(*cb, _regalloc);
1476 current_offset = cb->insts_size();
1478 #ifdef ASSERT
1479 if (n->size(_regalloc) < (current_offset-instr_offset)) {
1480 n->dump();
1481 assert(false, "wrong size of mach node");
1482 }
1483 #endif
1484 non_safepoints.observe_instruction(n, current_offset);
1486 // mcall is last "call" that can be a safepoint
1487 // record it so we can see if a poll will directly follow it
1488 // in which case we'll need a pad to make the PcDesc sites unique
1489 // see 5010568. This can be slightly inaccurate but conservative
1490 // in the case that return address is not actually at current_offset.
1491 // This is a small price to pay.
1493 if (is_mcall) {
1494 last_call_offset = current_offset;
1495 }
1497 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back()) {
1498 // Avoid back to back some instructions.
1499 last_avoid_back_to_back_offset = current_offset;
1500 }
1502 // See if this instruction has a delay slot
1503 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1504 assert(delay_slot != NULL, "expecting delay slot node");
1506 // Back up 1 instruction
1507 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1509 // Save the offset for the listing
1510 #ifndef PRODUCT
1511 if (node_offsets && delay_slot->_idx < node_offset_limit)
1512 node_offsets[delay_slot->_idx] = cb->insts_size();
1513 #endif
1515 // Support a SafePoint in the delay slot
1516 if (delay_slot->is_MachSafePoint()) {
1517 MachNode *mach = delay_slot->as_Mach();
1518 // !!!!! Stubs only need an oopmap right now, so bail out
1519 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
1520 // Write the oopmap directly to the code blob??!!
1521 # ifdef ENABLE_ZAP_DEAD_LOCALS
1522 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive");
1523 # endif
1524 delay_slot = NULL;
1525 continue;
1526 }
1528 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1529 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1530 adjusted_offset);
1531 // Generate an OopMap entry
1532 Process_OopMap_Node(mach, adjusted_offset);
1533 }
1535 // Insert the delay slot instruction
1536 delay_slot->emit(*cb, _regalloc);
1538 // Don't reuse it
1539 delay_slot = NULL;
1540 }
1542 } // End for all instructions in block
1544 // If the next block is the top of a loop, pad this block out to align
1545 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1546 if (i < nblocks-1) {
1547 Block *nb = _cfg->get_block(i + 1);
1548 int padding = nb->alignment_padding(current_offset);
1549 if( padding > 0 ) {
1550 MachNode *nop = new (this) MachNopNode(padding / nop_size);
1551 block->_nodes.insert(block->_nodes.size(), nop);
1552 _cfg->map_node_to_block(nop, block);
1553 nop->emit(*cb, _regalloc);
1554 current_offset = cb->insts_size();
1555 }
1556 }
1557 // Verify that the distance for generated before forward
1558 // short branches is still valid.
1559 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1561 // Save new block start offset
1562 blk_starts[i] = blk_offset;
1563 } // End of for all blocks
1564 blk_starts[nblocks] = current_offset;
1566 non_safepoints.flush_at_end();
1568 // Offset too large?
1569 if (failing()) return;
1571 // Define a pseudo-label at the end of the code
1572 MacroAssembler(cb).bind( blk_labels[nblocks] );
1574 // Compute the size of the first block
1575 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1577 assert(cb->insts_size() < 500000, "method is unreasonably large");
1579 #ifdef ASSERT
1580 for (uint i = 0; i < nblocks; i++) { // For all blocks
1581 if (jmp_target[i] != 0) {
1582 int br_size = jmp_size[i];
1583 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1584 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1585 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]);
1586 assert(false, "Displacement too large for short jmp");
1587 }
1588 }
1589 }
1590 #endif
1592 #ifndef PRODUCT
1593 // Information on the size of the method, without the extraneous code
1594 Scheduling::increment_method_size(cb->insts_size());
1595 #endif
1597 // ------------------
1598 // Fill in exception table entries.
1599 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1601 // Only java methods have exception handlers and deopt handlers
1602 if (_method) {
1603 // Emit the exception handler code.
1604 _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb));
1605 // Emit the deopt handler code.
1606 _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb));
1608 // Emit the MethodHandle deopt handler code (if required).
1609 if (has_method_handle_invokes()) {
1610 // We can use the same code as for the normal deopt handler, we
1611 // just need a different entry point address.
1612 _code_offsets.set_value(CodeOffsets::DeoptMH, emit_deopt_handler(*cb));
1613 }
1614 }
1616 // One last check for failed CodeBuffer::expand:
1617 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1618 C->record_failure("CodeCache is full");
1619 return;
1620 }
1622 #ifndef PRODUCT
1623 // Dump the assembly code, including basic-block numbers
1624 if (print_assembly()) {
1625 ttyLocker ttyl; // keep the following output all in one block
1626 if (!VMThread::should_terminate()) { // test this under the tty lock
1627 // This output goes directly to the tty, not the compiler log.
1628 // To enable tools to match it up with the compilation activity,
1629 // be sure to tag this tty output with the compile ID.
1630 if (xtty != NULL) {
1631 xtty->head("opto_assembly compile_id='%d'%s", compile_id(),
1632 is_osr_compilation() ? " compile_kind='osr'" :
1633 "");
1634 }
1635 if (method() != NULL) {
1636 method()->print_metadata();
1637 }
1638 dump_asm(node_offsets, node_offset_limit);
1639 if (xtty != NULL) {
1640 xtty->tail("opto_assembly");
1641 }
1642 }
1643 }
1644 #endif
1646 }
1648 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1649 _inc_table.set_size(cnt);
1651 uint inct_cnt = 0;
1652 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1653 Block* block = _cfg->get_block(i);
1654 Node *n = NULL;
1655 int j;
1657 // Find the branch; ignore trailing NOPs.
1658 for (j = block->_nodes.size() - 1; j >= 0; j--) {
1659 n = block->_nodes[j];
1660 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1661 break;
1662 }
1663 }
1665 // If we didn't find anything, continue
1666 if (j < 0) {
1667 continue;
1668 }
1670 // Compute ExceptionHandlerTable subtable entry and add it
1671 // (skip empty blocks)
1672 if (n->is_Catch()) {
1674 // Get the offset of the return from the call
1675 uint call_return = call_returns[block->_pre_order];
1676 #ifdef ASSERT
1677 assert( call_return > 0, "no call seen for this basic block" );
1678 while (block->_nodes[--j]->is_MachProj()) ;
1679 assert(block->_nodes[j]->is_MachCall(), "CatchProj must follow call");
1680 #endif
1681 // last instruction is a CatchNode, find it's CatchProjNodes
1682 int nof_succs = block->_num_succs;
1683 // allocate space
1684 GrowableArray<intptr_t> handler_bcis(nof_succs);
1685 GrowableArray<intptr_t> handler_pcos(nof_succs);
1686 // iterate through all successors
1687 for (int j = 0; j < nof_succs; j++) {
1688 Block* s = block->_succs[j];
1689 bool found_p = false;
1690 for (uint k = 1; k < s->num_preds(); k++) {
1691 Node* pk = s->pred(k);
1692 if (pk->is_CatchProj() && pk->in(0) == n) {
1693 const CatchProjNode* p = pk->as_CatchProj();
1694 found_p = true;
1695 // add the corresponding handler bci & pco information
1696 if (p->_con != CatchProjNode::fall_through_index) {
1697 // p leads to an exception handler (and is not fall through)
1698 assert(s == _cfg->get_block(s->_pre_order), "bad numbering");
1699 // no duplicates, please
1700 if (!handler_bcis.contains(p->handler_bci())) {
1701 uint block_num = s->non_connector()->_pre_order;
1702 handler_bcis.append(p->handler_bci());
1703 handler_pcos.append(blk_labels[block_num].loc_pos());
1704 }
1705 }
1706 }
1707 }
1708 assert(found_p, "no matching predecessor found");
1709 // Note: Due to empty block removal, one block may have
1710 // several CatchProj inputs, from the same Catch.
1711 }
1713 // Set the offset of the return from the call
1714 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
1715 continue;
1716 }
1718 // Handle implicit null exception table updates
1719 if (n->is_MachNullCheck()) {
1720 uint block_num = block->non_connector_successor(0)->_pre_order;
1721 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1722 continue;
1723 }
1724 } // End of for all blocks fill in exception table entries
1725 }
1727 // Static Variables
1728 #ifndef PRODUCT
1729 uint Scheduling::_total_nop_size = 0;
1730 uint Scheduling::_total_method_size = 0;
1731 uint Scheduling::_total_branches = 0;
1732 uint Scheduling::_total_unconditional_delays = 0;
1733 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1734 #endif
1736 // Initializer for class Scheduling
1738 Scheduling::Scheduling(Arena *arena, Compile &compile)
1739 : _arena(arena),
1740 _cfg(compile.cfg()),
1741 _regalloc(compile.regalloc()),
1742 _reg_node(arena),
1743 _bundle_instr_count(0),
1744 _bundle_cycle_number(0),
1745 _scheduled(arena),
1746 _available(arena),
1747 _next_node(NULL),
1748 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]),
1749 _pinch_free_list(arena)
1750 #ifndef PRODUCT
1751 , _branches(0)
1752 , _unconditional_delays(0)
1753 #endif
1754 {
1755 // Create a MachNopNode
1756 _nop = new (&compile) MachNopNode();
1758 // Now that the nops are in the array, save the count
1759 // (but allow entries for the nops)
1760 _node_bundling_limit = compile.unique();
1761 uint node_max = _regalloc->node_regs_max_index();
1763 compile.set_node_bundling_limit(_node_bundling_limit);
1765 // This one is persistent within the Compile class
1766 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1768 // Allocate space for fixed-size arrays
1769 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1770 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
1771 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1773 // Clear the arrays
1774 memset(_node_bundling_base, 0, node_max * sizeof(Bundle));
1775 memset(_node_latency, 0, node_max * sizeof(unsigned short));
1776 memset(_uses, 0, node_max * sizeof(short));
1777 memset(_current_latency, 0, node_max * sizeof(unsigned short));
1779 // Clear the bundling information
1780 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
1782 // Get the last node
1783 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
1785 _next_node = block->_nodes[block->_nodes.size() - 1];
1786 }
1788 #ifndef PRODUCT
1789 // Scheduling destructor
1790 Scheduling::~Scheduling() {
1791 _total_branches += _branches;
1792 _total_unconditional_delays += _unconditional_delays;
1793 }
1794 #endif
1796 // Step ahead "i" cycles
1797 void Scheduling::step(uint i) {
1799 Bundle *bundle = node_bundling(_next_node);
1800 bundle->set_starts_bundle();
1802 // Update the bundle record, but leave the flags information alone
1803 if (_bundle_instr_count > 0) {
1804 bundle->set_instr_count(_bundle_instr_count);
1805 bundle->set_resources_used(_bundle_use.resourcesUsed());
1806 }
1808 // Update the state information
1809 _bundle_instr_count = 0;
1810 _bundle_cycle_number += i;
1811 _bundle_use.step(i);
1812 }
1814 void Scheduling::step_and_clear() {
1815 Bundle *bundle = node_bundling(_next_node);
1816 bundle->set_starts_bundle();
1818 // Update the bundle record
1819 if (_bundle_instr_count > 0) {
1820 bundle->set_instr_count(_bundle_instr_count);
1821 bundle->set_resources_used(_bundle_use.resourcesUsed());
1823 _bundle_cycle_number += 1;
1824 }
1826 // Clear the bundling information
1827 _bundle_instr_count = 0;
1828 _bundle_use.reset();
1830 memcpy(_bundle_use_elements,
1831 Pipeline_Use::elaborated_elements,
1832 sizeof(Pipeline_Use::elaborated_elements));
1833 }
1835 // Perform instruction scheduling and bundling over the sequence of
1836 // instructions in backwards order.
1837 void Compile::ScheduleAndBundle() {
1839 // Don't optimize this if it isn't a method
1840 if (!_method)
1841 return;
1843 // Don't optimize this if scheduling is disabled
1844 if (!do_scheduling())
1845 return;
1847 // Scheduling code works only with pairs (8 bytes) maximum.
1848 if (max_vector_size() > 8)
1849 return;
1851 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); )
1853 // Create a data structure for all the scheduling information
1854 Scheduling scheduling(Thread::current()->resource_area(), *this);
1856 // Walk backwards over each basic block, computing the needed alignment
1857 // Walk over all the basic blocks
1858 scheduling.DoScheduling();
1859 }
1861 // Compute the latency of all the instructions. This is fairly simple,
1862 // because we already have a legal ordering. Walk over the instructions
1863 // from first to last, and compute the latency of the instruction based
1864 // on the latency of the preceding instruction(s).
1865 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
1866 #ifndef PRODUCT
1867 if (_cfg->C->trace_opto_output())
1868 tty->print("# -> ComputeLocalLatenciesForward\n");
1869 #endif
1871 // Walk over all the schedulable instructions
1872 for( uint j=_bb_start; j < _bb_end; j++ ) {
1874 // This is a kludge, forcing all latency calculations to start at 1.
1875 // Used to allow latency 0 to force an instruction to the beginning
1876 // of the bb
1877 uint latency = 1;
1878 Node *use = bb->_nodes[j];
1879 uint nlen = use->len();
1881 // Walk over all the inputs
1882 for ( uint k=0; k < nlen; k++ ) {
1883 Node *def = use->in(k);
1884 if (!def)
1885 continue;
1887 uint l = _node_latency[def->_idx] + use->latency(k);
1888 if (latency < l)
1889 latency = l;
1890 }
1892 _node_latency[use->_idx] = latency;
1894 #ifndef PRODUCT
1895 if (_cfg->C->trace_opto_output()) {
1896 tty->print("# latency %4d: ", latency);
1897 use->dump();
1898 }
1899 #endif
1900 }
1902 #ifndef PRODUCT
1903 if (_cfg->C->trace_opto_output())
1904 tty->print("# <- ComputeLocalLatenciesForward\n");
1905 #endif
1907 } // end ComputeLocalLatenciesForward
1909 // See if this node fits into the present instruction bundle
1910 bool Scheduling::NodeFitsInBundle(Node *n) {
1911 uint n_idx = n->_idx;
1913 // If this is the unconditional delay instruction, then it fits
1914 if (n == _unconditional_delay_slot) {
1915 #ifndef PRODUCT
1916 if (_cfg->C->trace_opto_output())
1917 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
1918 #endif
1919 return (true);
1920 }
1922 // If the node cannot be scheduled this cycle, skip it
1923 if (_current_latency[n_idx] > _bundle_cycle_number) {
1924 #ifndef PRODUCT
1925 if (_cfg->C->trace_opto_output())
1926 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
1927 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
1928 #endif
1929 return (false);
1930 }
1932 const Pipeline *node_pipeline = n->pipeline();
1934 uint instruction_count = node_pipeline->instructionCount();
1935 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
1936 instruction_count = 0;
1937 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
1938 instruction_count++;
1940 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
1941 #ifndef PRODUCT
1942 if (_cfg->C->trace_opto_output())
1943 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
1944 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
1945 #endif
1946 return (false);
1947 }
1949 // Don't allow non-machine nodes to be handled this way
1950 if (!n->is_Mach() && instruction_count == 0)
1951 return (false);
1953 // See if there is any overlap
1954 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
1956 if (delay > 0) {
1957 #ifndef PRODUCT
1958 if (_cfg->C->trace_opto_output())
1959 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
1960 #endif
1961 return false;
1962 }
1964 #ifndef PRODUCT
1965 if (_cfg->C->trace_opto_output())
1966 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
1967 #endif
1969 return true;
1970 }
1972 Node * Scheduling::ChooseNodeToBundle() {
1973 uint siz = _available.size();
1975 if (siz == 0) {
1977 #ifndef PRODUCT
1978 if (_cfg->C->trace_opto_output())
1979 tty->print("# ChooseNodeToBundle: NULL\n");
1980 #endif
1981 return (NULL);
1982 }
1984 // Fast path, if only 1 instruction in the bundle
1985 if (siz == 1) {
1986 #ifndef PRODUCT
1987 if (_cfg->C->trace_opto_output()) {
1988 tty->print("# ChooseNodeToBundle (only 1): ");
1989 _available[0]->dump();
1990 }
1991 #endif
1992 return (_available[0]);
1993 }
1995 // Don't bother, if the bundle is already full
1996 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
1997 for ( uint i = 0; i < siz; i++ ) {
1998 Node *n = _available[i];
2000 // Skip projections, we'll handle them another way
2001 if (n->is_Proj())
2002 continue;
2004 // This presupposed that instructions are inserted into the
2005 // available list in a legality order; i.e. instructions that
2006 // must be inserted first are at the head of the list
2007 if (NodeFitsInBundle(n)) {
2008 #ifndef PRODUCT
2009 if (_cfg->C->trace_opto_output()) {
2010 tty->print("# ChooseNodeToBundle: ");
2011 n->dump();
2012 }
2013 #endif
2014 return (n);
2015 }
2016 }
2017 }
2019 // Nothing fits in this bundle, choose the highest priority
2020 #ifndef PRODUCT
2021 if (_cfg->C->trace_opto_output()) {
2022 tty->print("# ChooseNodeToBundle: ");
2023 _available[0]->dump();
2024 }
2025 #endif
2027 return _available[0];
2028 }
2030 void Scheduling::AddNodeToAvailableList(Node *n) {
2031 assert( !n->is_Proj(), "projections never directly made available" );
2032 #ifndef PRODUCT
2033 if (_cfg->C->trace_opto_output()) {
2034 tty->print("# AddNodeToAvailableList: ");
2035 n->dump();
2036 }
2037 #endif
2039 int latency = _current_latency[n->_idx];
2041 // Insert in latency order (insertion sort)
2042 uint i;
2043 for ( i=0; i < _available.size(); i++ )
2044 if (_current_latency[_available[i]->_idx] > latency)
2045 break;
2047 // Special Check for compares following branches
2048 if( n->is_Mach() && _scheduled.size() > 0 ) {
2049 int op = n->as_Mach()->ideal_Opcode();
2050 Node *last = _scheduled[0];
2051 if( last->is_MachIf() && last->in(1) == n &&
2052 ( op == Op_CmpI ||
2053 op == Op_CmpU ||
2054 op == Op_CmpP ||
2055 op == Op_CmpF ||
2056 op == Op_CmpD ||
2057 op == Op_CmpL ) ) {
2059 // Recalculate position, moving to front of same latency
2060 for ( i=0 ; i < _available.size(); i++ )
2061 if (_current_latency[_available[i]->_idx] >= latency)
2062 break;
2063 }
2064 }
2066 // Insert the node in the available list
2067 _available.insert(i, n);
2069 #ifndef PRODUCT
2070 if (_cfg->C->trace_opto_output())
2071 dump_available();
2072 #endif
2073 }
2075 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2076 for ( uint i=0; i < n->len(); i++ ) {
2077 Node *def = n->in(i);
2078 if (!def) continue;
2079 if( def->is_Proj() ) // If this is a machine projection, then
2080 def = def->in(0); // propagate usage thru to the base instruction
2082 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2083 continue;
2084 }
2086 // Compute the latency
2087 uint l = _bundle_cycle_number + n->latency(i);
2088 if (_current_latency[def->_idx] < l)
2089 _current_latency[def->_idx] = l;
2091 // If this does not have uses then schedule it
2092 if ((--_uses[def->_idx]) == 0)
2093 AddNodeToAvailableList(def);
2094 }
2095 }
2097 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2098 #ifndef PRODUCT
2099 if (_cfg->C->trace_opto_output()) {
2100 tty->print("# AddNodeToBundle: ");
2101 n->dump();
2102 }
2103 #endif
2105 // Remove this from the available list
2106 uint i;
2107 for (i = 0; i < _available.size(); i++)
2108 if (_available[i] == n)
2109 break;
2110 assert(i < _available.size(), "entry in _available list not found");
2111 _available.remove(i);
2113 // See if this fits in the current bundle
2114 const Pipeline *node_pipeline = n->pipeline();
2115 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2117 // Check for instructions to be placed in the delay slot. We
2118 // do this before we actually schedule the current instruction,
2119 // because the delay slot follows the current instruction.
2120 if (Pipeline::_branch_has_delay_slot &&
2121 node_pipeline->hasBranchDelay() &&
2122 !_unconditional_delay_slot) {
2124 uint siz = _available.size();
2126 // Conditional branches can support an instruction that
2127 // is unconditionally executed and not dependent by the
2128 // branch, OR a conditionally executed instruction if
2129 // the branch is taken. In practice, this means that
2130 // the first instruction at the branch target is
2131 // copied to the delay slot, and the branch goes to
2132 // the instruction after that at the branch target
2133 if ( n->is_MachBranch() ) {
2135 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2136 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2138 #ifndef PRODUCT
2139 _branches++;
2140 #endif
2142 // At least 1 instruction is on the available list
2143 // that is not dependent on the branch
2144 for (uint i = 0; i < siz; i++) {
2145 Node *d = _available[i];
2146 const Pipeline *avail_pipeline = d->pipeline();
2148 // Don't allow safepoints in the branch shadow, that will
2149 // cause a number of difficulties
2150 if ( avail_pipeline->instructionCount() == 1 &&
2151 !avail_pipeline->hasMultipleBundles() &&
2152 !avail_pipeline->hasBranchDelay() &&
2153 Pipeline::instr_has_unit_size() &&
2154 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2155 NodeFitsInBundle(d) &&
2156 !node_bundling(d)->used_in_delay()) {
2158 if (d->is_Mach() && !d->is_MachSafePoint()) {
2159 // A node that fits in the delay slot was found, so we need to
2160 // set the appropriate bits in the bundle pipeline information so
2161 // that it correctly indicates resource usage. Later, when we
2162 // attempt to add this instruction to the bundle, we will skip
2163 // setting the resource usage.
2164 _unconditional_delay_slot = d;
2165 node_bundling(n)->set_use_unconditional_delay();
2166 node_bundling(d)->set_used_in_unconditional_delay();
2167 _bundle_use.add_usage(avail_pipeline->resourceUse());
2168 _current_latency[d->_idx] = _bundle_cycle_number;
2169 _next_node = d;
2170 ++_bundle_instr_count;
2171 #ifndef PRODUCT
2172 _unconditional_delays++;
2173 #endif
2174 break;
2175 }
2176 }
2177 }
2178 }
2180 // No delay slot, add a nop to the usage
2181 if (!_unconditional_delay_slot) {
2182 // See if adding an instruction in the delay slot will overflow
2183 // the bundle.
2184 if (!NodeFitsInBundle(_nop)) {
2185 #ifndef PRODUCT
2186 if (_cfg->C->trace_opto_output())
2187 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2188 #endif
2189 step(1);
2190 }
2192 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2193 _next_node = _nop;
2194 ++_bundle_instr_count;
2195 }
2197 // See if the instruction in the delay slot requires a
2198 // step of the bundles
2199 if (!NodeFitsInBundle(n)) {
2200 #ifndef PRODUCT
2201 if (_cfg->C->trace_opto_output())
2202 tty->print("# *** STEP(branch won't fit) ***\n");
2203 #endif
2204 // Update the state information
2205 _bundle_instr_count = 0;
2206 _bundle_cycle_number += 1;
2207 _bundle_use.step(1);
2208 }
2209 }
2211 // Get the number of instructions
2212 uint instruction_count = node_pipeline->instructionCount();
2213 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2214 instruction_count = 0;
2216 // Compute the latency information
2217 uint delay = 0;
2219 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2220 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2221 if (relative_latency < 0)
2222 relative_latency = 0;
2224 delay = _bundle_use.full_latency(relative_latency, node_usage);
2226 // Does not fit in this bundle, start a new one
2227 if (delay > 0) {
2228 step(delay);
2230 #ifndef PRODUCT
2231 if (_cfg->C->trace_opto_output())
2232 tty->print("# *** STEP(%d) ***\n", delay);
2233 #endif
2234 }
2235 }
2237 // If this was placed in the delay slot, ignore it
2238 if (n != _unconditional_delay_slot) {
2240 if (delay == 0) {
2241 if (node_pipeline->hasMultipleBundles()) {
2242 #ifndef PRODUCT
2243 if (_cfg->C->trace_opto_output())
2244 tty->print("# *** STEP(multiple instructions) ***\n");
2245 #endif
2246 step(1);
2247 }
2249 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2250 #ifndef PRODUCT
2251 if (_cfg->C->trace_opto_output())
2252 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2253 instruction_count + _bundle_instr_count,
2254 Pipeline::_max_instrs_per_cycle);
2255 #endif
2256 step(1);
2257 }
2258 }
2260 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2261 _bundle_instr_count++;
2263 // Set the node's latency
2264 _current_latency[n->_idx] = _bundle_cycle_number;
2266 // Now merge the functional unit information
2267 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2268 _bundle_use.add_usage(node_usage);
2270 // Increment the number of instructions in this bundle
2271 _bundle_instr_count += instruction_count;
2273 // Remember this node for later
2274 if (n->is_Mach())
2275 _next_node = n;
2276 }
2278 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2279 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2280 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2281 // into the block. All other scheduled nodes get put in the schedule here.
2282 int op = n->Opcode();
2283 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2284 (op != Op_Node && // Not an unused antidepedence node and
2285 // not an unallocated boxlock
2286 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2288 // Push any trailing projections
2289 if( bb->_nodes[bb->_nodes.size()-1] != n ) {
2290 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2291 Node *foi = n->fast_out(i);
2292 if( foi->is_Proj() )
2293 _scheduled.push(foi);
2294 }
2295 }
2297 // Put the instruction in the schedule list
2298 _scheduled.push(n);
2299 }
2301 #ifndef PRODUCT
2302 if (_cfg->C->trace_opto_output())
2303 dump_available();
2304 #endif
2306 // Walk all the definitions, decrementing use counts, and
2307 // if a definition has a 0 use count, place it in the available list.
2308 DecrementUseCounts(n,bb);
2309 }
2311 // This method sets the use count within a basic block. We will ignore all
2312 // uses outside the current basic block. As we are doing a backwards walk,
2313 // any node we reach that has a use count of 0 may be scheduled. This also
2314 // avoids the problem of cyclic references from phi nodes, as long as phi
2315 // nodes are at the front of the basic block. This method also initializes
2316 // the available list to the set of instructions that have no uses within this
2317 // basic block.
2318 void Scheduling::ComputeUseCount(const Block *bb) {
2319 #ifndef PRODUCT
2320 if (_cfg->C->trace_opto_output())
2321 tty->print("# -> ComputeUseCount\n");
2322 #endif
2324 // Clear the list of available and scheduled instructions, just in case
2325 _available.clear();
2326 _scheduled.clear();
2328 // No delay slot specified
2329 _unconditional_delay_slot = NULL;
2331 #ifdef ASSERT
2332 for( uint i=0; i < bb->_nodes.size(); i++ )
2333 assert( _uses[bb->_nodes[i]->_idx] == 0, "_use array not clean" );
2334 #endif
2336 // Force the _uses count to never go to zero for unscheduable pieces
2337 // of the block
2338 for( uint k = 0; k < _bb_start; k++ )
2339 _uses[bb->_nodes[k]->_idx] = 1;
2340 for( uint l = _bb_end; l < bb->_nodes.size(); l++ )
2341 _uses[bb->_nodes[l]->_idx] = 1;
2343 // Iterate backwards over the instructions in the block. Don't count the
2344 // branch projections at end or the block header instructions.
2345 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2346 Node *n = bb->_nodes[j];
2347 if( n->is_Proj() ) continue; // Projections handled another way
2349 // Account for all uses
2350 for ( uint k = 0; k < n->len(); k++ ) {
2351 Node *inp = n->in(k);
2352 if (!inp) continue;
2353 assert(inp != n, "no cycles allowed" );
2354 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2355 if (inp->is_Proj()) { // Skip through Proj's
2356 inp = inp->in(0);
2357 }
2358 ++_uses[inp->_idx]; // Count 1 block-local use
2359 }
2360 }
2362 // If this instruction has a 0 use count, then it is available
2363 if (!_uses[n->_idx]) {
2364 _current_latency[n->_idx] = _bundle_cycle_number;
2365 AddNodeToAvailableList(n);
2366 }
2368 #ifndef PRODUCT
2369 if (_cfg->C->trace_opto_output()) {
2370 tty->print("# uses: %3d: ", _uses[n->_idx]);
2371 n->dump();
2372 }
2373 #endif
2374 }
2376 #ifndef PRODUCT
2377 if (_cfg->C->trace_opto_output())
2378 tty->print("# <- ComputeUseCount\n");
2379 #endif
2380 }
2382 // This routine performs scheduling on each basic block in reverse order,
2383 // using instruction latencies and taking into account function unit
2384 // availability.
2385 void Scheduling::DoScheduling() {
2386 #ifndef PRODUCT
2387 if (_cfg->C->trace_opto_output())
2388 tty->print("# -> DoScheduling\n");
2389 #endif
2391 Block *succ_bb = NULL;
2392 Block *bb;
2394 // Walk over all the basic blocks in reverse order
2395 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2396 bb = _cfg->get_block(i);
2398 #ifndef PRODUCT
2399 if (_cfg->C->trace_opto_output()) {
2400 tty->print("# Schedule BB#%03d (initial)\n", i);
2401 for (uint j = 0; j < bb->_nodes.size(); j++) {
2402 bb->_nodes[j]->dump();
2403 }
2404 }
2405 #endif
2407 // On the head node, skip processing
2408 if (bb == _cfg->get_root_block()) {
2409 continue;
2410 }
2412 // Skip empty, connector blocks
2413 if (bb->is_connector())
2414 continue;
2416 // If the following block is not the sole successor of
2417 // this one, then reset the pipeline information
2418 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2419 #ifndef PRODUCT
2420 if (_cfg->C->trace_opto_output()) {
2421 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2422 _next_node->_idx, _bundle_instr_count);
2423 }
2424 #endif
2425 step_and_clear();
2426 }
2428 // Leave untouched the starting instruction, any Phis, a CreateEx node
2429 // or Top. bb->_nodes[_bb_start] is the first schedulable instruction.
2430 _bb_end = bb->_nodes.size()-1;
2431 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2432 Node *n = bb->_nodes[_bb_start];
2433 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2434 // Also, MachIdealNodes do not get scheduled
2435 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2436 MachNode *mach = n->as_Mach();
2437 int iop = mach->ideal_Opcode();
2438 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2439 if( iop == Op_Con ) continue; // Do not schedule Top
2440 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2441 mach->pipeline() == MachNode::pipeline_class() &&
2442 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc
2443 continue;
2444 break; // Funny loop structure to be sure...
2445 }
2446 // Compute last "interesting" instruction in block - last instruction we
2447 // might schedule. _bb_end points just after last schedulable inst. We
2448 // normally schedule conditional branches (despite them being forced last
2449 // in the block), because they have delay slots we can fill. Calls all
2450 // have their delay slots filled in the template expansions, so we don't
2451 // bother scheduling them.
2452 Node *last = bb->_nodes[_bb_end];
2453 // Ignore trailing NOPs.
2454 while (_bb_end > 0 && last->is_Mach() &&
2455 last->as_Mach()->ideal_Opcode() == Op_Con) {
2456 last = bb->_nodes[--_bb_end];
2457 }
2458 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2459 if( last->is_Catch() ||
2460 // Exclude unreachable path case when Halt node is in a separate block.
2461 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2462 // There must be a prior call. Skip it.
2463 while( !bb->_nodes[--_bb_end]->is_MachCall() ) {
2464 assert( bb->_nodes[_bb_end]->is_MachProj(), "skipping projections after expected call" );
2465 }
2466 } else if( last->is_MachNullCheck() ) {
2467 // Backup so the last null-checked memory instruction is
2468 // outside the schedulable range. Skip over the nullcheck,
2469 // projection, and the memory nodes.
2470 Node *mem = last->in(1);
2471 do {
2472 _bb_end--;
2473 } while (mem != bb->_nodes[_bb_end]);
2474 } else {
2475 // Set _bb_end to point after last schedulable inst.
2476 _bb_end++;
2477 }
2479 assert( _bb_start <= _bb_end, "inverted block ends" );
2481 // Compute the register antidependencies for the basic block
2482 ComputeRegisterAntidependencies(bb);
2483 if (_cfg->C->failing()) return; // too many D-U pinch points
2485 // Compute intra-bb latencies for the nodes
2486 ComputeLocalLatenciesForward(bb);
2488 // Compute the usage within the block, and set the list of all nodes
2489 // in the block that have no uses within the block.
2490 ComputeUseCount(bb);
2492 // Schedule the remaining instructions in the block
2493 while ( _available.size() > 0 ) {
2494 Node *n = ChooseNodeToBundle();
2495 guarantee(n != NULL, "no nodes available");
2496 AddNodeToBundle(n,bb);
2497 }
2499 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2500 #ifdef ASSERT
2501 for( uint l = _bb_start; l < _bb_end; l++ ) {
2502 Node *n = bb->_nodes[l];
2503 uint m;
2504 for( m = 0; m < _bb_end-_bb_start; m++ )
2505 if( _scheduled[m] == n )
2506 break;
2507 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2508 }
2509 #endif
2511 // Now copy the instructions (in reverse order) back to the block
2512 for ( uint k = _bb_start; k < _bb_end; k++ )
2513 bb->_nodes.map(k, _scheduled[_bb_end-k-1]);
2515 #ifndef PRODUCT
2516 if (_cfg->C->trace_opto_output()) {
2517 tty->print("# Schedule BB#%03d (final)\n", i);
2518 uint current = 0;
2519 for (uint j = 0; j < bb->_nodes.size(); j++) {
2520 Node *n = bb->_nodes[j];
2521 if( valid_bundle_info(n) ) {
2522 Bundle *bundle = node_bundling(n);
2523 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2524 tty->print("*** Bundle: ");
2525 bundle->dump();
2526 }
2527 n->dump();
2528 }
2529 }
2530 }
2531 #endif
2532 #ifdef ASSERT
2533 verify_good_schedule(bb,"after block local scheduling");
2534 #endif
2535 }
2537 #ifndef PRODUCT
2538 if (_cfg->C->trace_opto_output())
2539 tty->print("# <- DoScheduling\n");
2540 #endif
2542 // Record final node-bundling array location
2543 _regalloc->C->set_node_bundling_base(_node_bundling_base);
2545 } // end DoScheduling
2547 // Verify that no live-range used in the block is killed in the block by a
2548 // wrong DEF. This doesn't verify live-ranges that span blocks.
2550 // Check for edge existence. Used to avoid adding redundant precedence edges.
2551 static bool edge_from_to( Node *from, Node *to ) {
2552 for( uint i=0; i<from->len(); i++ )
2553 if( from->in(i) == to )
2554 return true;
2555 return false;
2556 }
2558 #ifdef ASSERT
2559 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2560 // Check for bad kills
2561 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2562 Node *prior_use = _reg_node[def];
2563 if( prior_use && !edge_from_to(prior_use,n) ) {
2564 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2565 n->dump();
2566 tty->print_cr("...");
2567 prior_use->dump();
2568 assert(edge_from_to(prior_use,n),msg);
2569 }
2570 _reg_node.map(def,NULL); // Kill live USEs
2571 }
2572 }
2574 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2576 // Zap to something reasonable for the verify code
2577 _reg_node.clear();
2579 // Walk over the block backwards. Check to make sure each DEF doesn't
2580 // kill a live value (other than the one it's supposed to). Add each
2581 // USE to the live set.
2582 for( uint i = b->_nodes.size()-1; i >= _bb_start; i-- ) {
2583 Node *n = b->_nodes[i];
2584 int n_op = n->Opcode();
2585 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2586 // Fat-proj kills a slew of registers
2587 RegMask rm = n->out_RegMask();// Make local copy
2588 while( rm.is_NotEmpty() ) {
2589 OptoReg::Name kill = rm.find_first_elem();
2590 rm.Remove(kill);
2591 verify_do_def( n, kill, msg );
2592 }
2593 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2594 // Get DEF'd registers the normal way
2595 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2596 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2597 }
2599 // Now make all USEs live
2600 for( uint i=1; i<n->req(); i++ ) {
2601 Node *def = n->in(i);
2602 assert(def != 0, "input edge required");
2603 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2604 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2605 if( OptoReg::is_valid(reg_lo) ) {
2606 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg);
2607 _reg_node.map(reg_lo,n);
2608 }
2609 if( OptoReg::is_valid(reg_hi) ) {
2610 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg);
2611 _reg_node.map(reg_hi,n);
2612 }
2613 }
2615 }
2617 // Zap to something reasonable for the Antidependence code
2618 _reg_node.clear();
2619 }
2620 #endif
2622 // Conditionally add precedence edges. Avoid putting edges on Projs.
2623 static void add_prec_edge_from_to( Node *from, Node *to ) {
2624 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2625 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2626 from = from->in(0);
2627 }
2628 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2629 !edge_from_to( from, to ) ) // Avoid duplicate edge
2630 from->add_prec(to);
2631 }
2633 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2634 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2635 return;
2637 Node *pinch = _reg_node[def_reg]; // Get pinch point
2638 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2639 is_def ) { // Check for a true def (not a kill)
2640 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2641 return;
2642 }
2644 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2645 debug_only( def = (Node*)0xdeadbeef; )
2647 // After some number of kills there _may_ be a later def
2648 Node *later_def = NULL;
2650 // Finding a kill requires a real pinch-point.
2651 // Check for not already having a pinch-point.
2652 // Pinch points are Op_Node's.
2653 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2654 later_def = pinch; // Must be def/kill as optimistic pinch-point
2655 if ( _pinch_free_list.size() > 0) {
2656 pinch = _pinch_free_list.pop();
2657 } else {
2658 pinch = new (_cfg->C) Node(1); // Pinch point to-be
2659 }
2660 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2661 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2662 return;
2663 }
2664 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
2665 _reg_node.map(def_reg,pinch); // Record pinch-point
2666 //_regalloc->set_bad(pinch->_idx); // Already initialized this way.
2667 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2668 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call
2669 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2670 later_def = NULL; // and no later def
2671 }
2672 pinch->set_req(0,later_def); // Hook later def so we can find it
2673 } else { // Else have valid pinch point
2674 if( pinch->in(0) ) // If there is a later-def
2675 later_def = pinch->in(0); // Get it
2676 }
2678 // Add output-dependence edge from later def to kill
2679 if( later_def ) // If there is some original def
2680 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2682 // See if current kill is also a use, and so is forced to be the pinch-point.
2683 if( pinch->Opcode() == Op_Node ) {
2684 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2685 for( uint i=1; i<uses->req(); i++ ) {
2686 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2687 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2688 // Yes, found a use/kill pinch-point
2689 pinch->set_req(0,NULL); //
2690 pinch->replace_by(kill); // Move anti-dep edges up
2691 pinch = kill;
2692 _reg_node.map(def_reg,pinch);
2693 return;
2694 }
2695 }
2696 }
2698 // Add edge from kill to pinch-point
2699 add_prec_edge_from_to(kill,pinch);
2700 }
2702 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2703 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2704 return;
2705 Node *pinch = _reg_node[use_reg]; // Get pinch point
2706 // Check for no later def_reg/kill in block
2707 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b &&
2708 // Use has to be block-local as well
2709 _cfg->get_block_for_node(use) == b) {
2710 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2711 pinch->req() == 1 ) { // pinch not yet in block?
2712 pinch->del_req(0); // yank pointer to later-def, also set flag
2713 // Insert the pinch-point in the block just after the last use
2714 b->_nodes.insert(b->find_node(use)+1,pinch);
2715 _bb_end++; // Increase size scheduled region in block
2716 }
2718 add_prec_edge_from_to(pinch,use);
2719 }
2720 }
2722 // We insert antidependences between the reads and following write of
2723 // allocated registers to prevent illegal code motion. Hopefully, the
2724 // number of added references should be fairly small, especially as we
2725 // are only adding references within the current basic block.
2726 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2728 #ifdef ASSERT
2729 verify_good_schedule(b,"before block local scheduling");
2730 #endif
2732 // A valid schedule, for each register independently, is an endless cycle
2733 // of: a def, then some uses (connected to the def by true dependencies),
2734 // then some kills (defs with no uses), finally the cycle repeats with a new
2735 // def. The uses are allowed to float relative to each other, as are the
2736 // kills. No use is allowed to slide past a kill (or def). This requires
2737 // antidependencies between all uses of a single def and all kills that
2738 // follow, up to the next def. More edges are redundant, because later defs
2739 // & kills are already serialized with true or antidependencies. To keep
2740 // the edge count down, we add a 'pinch point' node if there's more than
2741 // one use or more than one kill/def.
2743 // We add dependencies in one bottom-up pass.
2745 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2747 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2748 // register. If not, we record the DEF/KILL in _reg_node, the
2749 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
2750 // "pinch point", a new Node that's in the graph but not in the block.
2751 // We put edges from the prior and current DEF/KILLs to the pinch point.
2752 // We put the pinch point in _reg_node. If there's already a pinch point
2753 // we merely add an edge from the current DEF/KILL to the pinch point.
2755 // After doing the DEF/KILLs, we handle USEs. For each used register, we
2756 // put an edge from the pinch point to the USE.
2758 // To be expedient, the _reg_node array is pre-allocated for the whole
2759 // compilation. _reg_node is lazily initialized; it either contains a NULL,
2760 // or a valid def/kill/pinch-point, or a leftover node from some prior
2761 // block. Leftover node from some prior block is treated like a NULL (no
2762 // prior def, so no anti-dependence needed). Valid def is distinguished by
2763 // it being in the current block.
2764 bool fat_proj_seen = false;
2765 uint last_safept = _bb_end-1;
2766 Node* end_node = (_bb_end-1 >= _bb_start) ? b->_nodes[last_safept] : NULL;
2767 Node* last_safept_node = end_node;
2768 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2769 Node *n = b->_nodes[i];
2770 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
2771 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
2772 // Fat-proj kills a slew of registers
2773 // This can add edges to 'n' and obscure whether or not it was a def,
2774 // hence the is_def flag.
2775 fat_proj_seen = true;
2776 RegMask rm = n->out_RegMask();// Make local copy
2777 while( rm.is_NotEmpty() ) {
2778 OptoReg::Name kill = rm.find_first_elem();
2779 rm.Remove(kill);
2780 anti_do_def( b, n, kill, is_def );
2781 }
2782 } else {
2783 // Get DEF'd registers the normal way
2784 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2785 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2786 }
2788 // Kill projections on a branch should appear to occur on the
2789 // branch, not afterwards, so grab the masks from the projections
2790 // and process them.
2791 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) {
2792 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2793 Node* use = n->fast_out(i);
2794 if (use->is_Proj()) {
2795 RegMask rm = use->out_RegMask();// Make local copy
2796 while( rm.is_NotEmpty() ) {
2797 OptoReg::Name kill = rm.find_first_elem();
2798 rm.Remove(kill);
2799 anti_do_def( b, n, kill, false );
2800 }
2801 }
2802 }
2803 }
2805 // Check each register used by this instruction for a following DEF/KILL
2806 // that must occur afterward and requires an anti-dependence edge.
2807 for( uint j=0; j<n->req(); j++ ) {
2808 Node *def = n->in(j);
2809 if( def ) {
2810 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
2811 anti_do_use( b, n, _regalloc->get_reg_first(def) );
2812 anti_do_use( b, n, _regalloc->get_reg_second(def) );
2813 }
2814 }
2815 // Do not allow defs of new derived values to float above GC
2816 // points unless the base is definitely available at the GC point.
2818 Node *m = b->_nodes[i];
2820 // Add precedence edge from following safepoint to use of derived pointer
2821 if( last_safept_node != end_node &&
2822 m != last_safept_node) {
2823 for (uint k = 1; k < m->req(); k++) {
2824 const Type *t = m->in(k)->bottom_type();
2825 if( t->isa_oop_ptr() &&
2826 t->is_ptr()->offset() != 0 ) {
2827 last_safept_node->add_prec( m );
2828 break;
2829 }
2830 }
2831 }
2833 if( n->jvms() ) { // Precedence edge from derived to safept
2834 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2835 if( b->_nodes[last_safept] != last_safept_node ) {
2836 last_safept = b->find_node(last_safept_node);
2837 }
2838 for( uint j=last_safept; j > i; j-- ) {
2839 Node *mach = b->_nodes[j];
2840 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2841 mach->add_prec( n );
2842 }
2843 last_safept = i;
2844 last_safept_node = m;
2845 }
2846 }
2848 if (fat_proj_seen) {
2849 // Garbage collect pinch nodes that were not consumed.
2850 // They are usually created by a fat kill MachProj for a call.
2851 garbage_collect_pinch_nodes();
2852 }
2853 }
2855 // Garbage collect pinch nodes for reuse by other blocks.
2856 //
2857 // The block scheduler's insertion of anti-dependence
2858 // edges creates many pinch nodes when the block contains
2859 // 2 or more Calls. A pinch node is used to prevent a
2860 // combinatorial explosion of edges. If a set of kills for a
2861 // register is anti-dependent on a set of uses (or defs), rather
2862 // than adding an edge in the graph between each pair of kill
2863 // and use (or def), a pinch is inserted between them:
2864 //
2865 // use1 use2 use3
2866 // \ | /
2867 // \ | /
2868 // pinch
2869 // / | \
2870 // / | \
2871 // kill1 kill2 kill3
2872 //
2873 // One pinch node is created per register killed when
2874 // the second call is encountered during a backwards pass
2875 // over the block. Most of these pinch nodes are never
2876 // wired into the graph because the register is never
2877 // used or def'ed in the block.
2878 //
2879 void Scheduling::garbage_collect_pinch_nodes() {
2880 #ifndef PRODUCT
2881 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
2882 #endif
2883 int trace_cnt = 0;
2884 for (uint k = 0; k < _reg_node.Size(); k++) {
2885 Node* pinch = _reg_node[k];
2886 if ((pinch != NULL) && pinch->Opcode() == Op_Node &&
2887 // no predecence input edges
2888 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
2889 cleanup_pinch(pinch);
2890 _pinch_free_list.push(pinch);
2891 _reg_node.map(k, NULL);
2892 #ifndef PRODUCT
2893 if (_cfg->C->trace_opto_output()) {
2894 trace_cnt++;
2895 if (trace_cnt > 40) {
2896 tty->print("\n");
2897 trace_cnt = 0;
2898 }
2899 tty->print(" %d", pinch->_idx);
2900 }
2901 #endif
2902 }
2903 }
2904 #ifndef PRODUCT
2905 if (_cfg->C->trace_opto_output()) tty->print("\n");
2906 #endif
2907 }
2909 // Clean up a pinch node for reuse.
2910 void Scheduling::cleanup_pinch( Node *pinch ) {
2911 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
2913 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
2914 Node* use = pinch->last_out(i);
2915 uint uses_found = 0;
2916 for (uint j = use->req(); j < use->len(); j++) {
2917 if (use->in(j) == pinch) {
2918 use->rm_prec(j);
2919 uses_found++;
2920 }
2921 }
2922 assert(uses_found > 0, "must be a precedence edge");
2923 i -= uses_found; // we deleted 1 or more copies of this edge
2924 }
2925 // May have a later_def entry
2926 pinch->set_req(0, NULL);
2927 }
2929 #ifndef PRODUCT
2931 void Scheduling::dump_available() const {
2932 tty->print("#Availist ");
2933 for (uint i = 0; i < _available.size(); i++)
2934 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
2935 tty->cr();
2936 }
2938 // Print Scheduling Statistics
2939 void Scheduling::print_statistics() {
2940 // Print the size added by nops for bundling
2941 tty->print("Nops added %d bytes to total of %d bytes",
2942 _total_nop_size, _total_method_size);
2943 if (_total_method_size > 0)
2944 tty->print(", for %.2f%%",
2945 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
2946 tty->print("\n");
2948 // Print the number of branch shadows filled
2949 if (Pipeline::_branch_has_delay_slot) {
2950 tty->print("Of %d branches, %d had unconditional delay slots filled",
2951 _total_branches, _total_unconditional_delays);
2952 if (_total_branches > 0)
2953 tty->print(", for %.2f%%",
2954 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
2955 tty->print("\n");
2956 }
2958 uint total_instructions = 0, total_bundles = 0;
2960 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
2961 uint bundle_count = _total_instructions_per_bundle[i];
2962 total_instructions += bundle_count * i;
2963 total_bundles += bundle_count;
2964 }
2966 if (total_bundles > 0)
2967 tty->print("Average ILP (excluding nops) is %.2f\n",
2968 ((double)total_instructions) / ((double)total_bundles));
2969 }
2970 #endif