1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "gc/shared/barrierSet.hpp"
26 #include "gc/shared/c2/barrierSetC2.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "oops/compressedOops.hpp"
30 #include "opto/ad.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/callnode.hpp"
33 #include "opto/idealGraphPrinter.hpp"
34 #include "opto/matcher.hpp"
35 #include "opto/memnode.hpp"
36 #include "opto/movenode.hpp"
37 #include "opto/opcodes.hpp"
38 #include "opto/regmask.hpp"
39 #include "opto/rootnode.hpp"
40 #include "opto/runtime.hpp"
41 #include "opto/type.hpp"
42 #include "opto/vectornode.hpp"
43 #include "runtime/os.inline.hpp"
44 #include "runtime/sharedRuntime.hpp"
45 #include "utilities/align.hpp"
46
47 OptoReg::Name OptoReg::c_frame_pointer;
48
49 const RegMask *Matcher::idealreg2regmask[_last_machine_leaf];
50 RegMask Matcher::mreg2regmask[_last_Mach_Reg];
51 RegMask Matcher::caller_save_regmask;
52 RegMask Matcher::caller_save_regmask_exclude_soe;
53 RegMask Matcher::mh_caller_save_regmask;
54 RegMask Matcher::mh_caller_save_regmask_exclude_soe;
55 RegMask Matcher::STACK_ONLY_mask;
56 RegMask Matcher::c_frame_ptr_mask;
57 const uint Matcher::_begin_rematerialize = _BEGIN_REMATERIALIZE;
58 const uint Matcher::_end_rematerialize = _END_REMATERIALIZE;
59
60 //---------------------------Matcher-------------------------------------------
61 Matcher::Matcher()
62 : PhaseTransform( Phase::Ins_Select ),
63 _states_arena(Chunk::medium_size, mtCompiler, Arena::Tag::tag_states),
64 _new_nodes(C->comp_arena()),
65 _visited(&_states_arena),
66 _shared(&_states_arena),
67 _dontcare(&_states_arena),
68 _reduceOp(reduceOp), _leftOp(leftOp), _rightOp(rightOp),
69 _swallowed(swallowed),
70 _begin_inst_chain_rule(_BEGIN_INST_CHAIN_RULE),
71 _end_inst_chain_rule(_END_INST_CHAIN_RULE),
72 _must_clone(must_clone),
73 _shared_nodes(C->comp_arena()),
74 #ifndef PRODUCT
75 _old2new_map(C->comp_arena()),
76 _new2old_map(C->comp_arena()),
77 _reused(C->comp_arena()),
78 #endif // !PRODUCT
79 _allocation_started(false),
80 _ruleName(ruleName),
81 _register_save_policy(register_save_policy),
82 _c_reg_save_policy(c_reg_save_policy),
83 _register_save_type(register_save_type) {
84 C->set_matcher(this);
85
86 idealreg2spillmask [Op_RegI] = nullptr;
87 idealreg2spillmask [Op_RegN] = nullptr;
88 idealreg2spillmask [Op_RegL] = nullptr;
89 idealreg2spillmask [Op_RegF] = nullptr;
90 idealreg2spillmask [Op_RegD] = nullptr;
91 idealreg2spillmask [Op_RegP] = nullptr;
92 idealreg2spillmask [Op_VecA] = nullptr;
93 idealreg2spillmask [Op_VecS] = nullptr;
94 idealreg2spillmask [Op_VecD] = nullptr;
95 idealreg2spillmask [Op_VecX] = nullptr;
96 idealreg2spillmask [Op_VecY] = nullptr;
97 idealreg2spillmask [Op_VecZ] = nullptr;
98 idealreg2spillmask [Op_RegFlags] = nullptr;
99 idealreg2spillmask [Op_RegVectMask] = nullptr;
100
101 idealreg2debugmask [Op_RegI] = nullptr;
102 idealreg2debugmask [Op_RegN] = nullptr;
103 idealreg2debugmask [Op_RegL] = nullptr;
104 idealreg2debugmask [Op_RegF] = nullptr;
105 idealreg2debugmask [Op_RegD] = nullptr;
106 idealreg2debugmask [Op_RegP] = nullptr;
107 idealreg2debugmask [Op_VecA] = nullptr;
108 idealreg2debugmask [Op_VecS] = nullptr;
109 idealreg2debugmask [Op_VecD] = nullptr;
110 idealreg2debugmask [Op_VecX] = nullptr;
111 idealreg2debugmask [Op_VecY] = nullptr;
112 idealreg2debugmask [Op_VecZ] = nullptr;
113 idealreg2debugmask [Op_RegFlags] = nullptr;
114 idealreg2debugmask [Op_RegVectMask] = nullptr;
115
116 idealreg2mhdebugmask[Op_RegI] = nullptr;
117 idealreg2mhdebugmask[Op_RegN] = nullptr;
118 idealreg2mhdebugmask[Op_RegL] = nullptr;
119 idealreg2mhdebugmask[Op_RegF] = nullptr;
120 idealreg2mhdebugmask[Op_RegD] = nullptr;
121 idealreg2mhdebugmask[Op_RegP] = nullptr;
122 idealreg2mhdebugmask[Op_VecA] = nullptr;
123 idealreg2mhdebugmask[Op_VecS] = nullptr;
124 idealreg2mhdebugmask[Op_VecD] = nullptr;
125 idealreg2mhdebugmask[Op_VecX] = nullptr;
126 idealreg2mhdebugmask[Op_VecY] = nullptr;
127 idealreg2mhdebugmask[Op_VecZ] = nullptr;
128 idealreg2mhdebugmask[Op_RegFlags] = nullptr;
129 idealreg2mhdebugmask[Op_RegVectMask] = nullptr;
130
131 debug_only(_mem_node = nullptr;) // Ideal memory node consumed by mach node
132 }
133
134 //------------------------------warp_incoming_stk_arg------------------------
135 // This warps a VMReg into an OptoReg::Name
136 OptoReg::Name Matcher::warp_incoming_stk_arg( VMReg reg ) {
137 OptoReg::Name warped;
138 if( reg->is_stack() ) { // Stack slot argument?
139 warped = OptoReg::add(_old_SP, reg->reg2stack() );
140 warped = OptoReg::add(warped, C->out_preserve_stack_slots());
141 if( warped >= _in_arg_limit )
142 _in_arg_limit = OptoReg::add(warped, 1); // Bump max stack slot seen
143 if (!RegMask::can_represent_arg(warped)) {
144 // the compiler cannot represent this method's calling sequence
145 // Bailout. We do not have space to represent all arguments.
146 C->record_method_not_compilable("unsupported incoming calling sequence");
147 return OptoReg::Bad;
148 }
149 return warped;
150 }
151 return OptoReg::as_OptoReg(reg);
152 }
153
154 //---------------------------compute_old_SP------------------------------------
155 OptoReg::Name Compile::compute_old_SP() {
156 int fixed = fixed_slots();
157 int preserve = in_preserve_stack_slots();
158 return OptoReg::stack2reg(align_up(fixed + preserve, (int)Matcher::stack_alignment_in_slots()));
159 }
160
161
162
163 #ifdef ASSERT
164 void Matcher::verify_new_nodes_only(Node* xroot) {
165 // Make sure that the new graph only references new nodes
166 ResourceMark rm;
167 Unique_Node_List worklist;
168 VectorSet visited;
169 worklist.push(xroot);
170 while (worklist.size() > 0) {
171 Node* n = worklist.pop();
172 if (visited.test_set(n->_idx)) {
173 continue;
174 }
175 assert(C->node_arena()->contains(n), "dead node");
176 for (uint j = 0; j < n->req(); j++) {
177 Node* in = n->in(j);
178 if (in != nullptr) {
179 worklist.push(in);
180 }
181 }
182 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
183 worklist.push(n->fast_out(j));
184 }
185 }
186 }
187 #endif
188
189 // Array of RegMask, one per returned values (inline type instances can
190 // be returned as multiple return values, one per field)
191 RegMask* Matcher::return_values_mask(const TypeFunc* tf) {
192 const TypeTuple* range = tf->range_cc();
193 uint cnt = range->cnt() - TypeFunc::Parms;
194 if (cnt == 0) {
195 return nullptr;
196 }
197 RegMask* mask = NEW_RESOURCE_ARRAY(RegMask, cnt);
198 BasicType* sig_bt = NEW_RESOURCE_ARRAY(BasicType, cnt);
199 VMRegPair* vm_parm_regs = NEW_RESOURCE_ARRAY(VMRegPair, cnt);
200 for (uint i = 0; i < cnt; i++) {
201 sig_bt[i] = range->field_at(i+TypeFunc::Parms)->basic_type();
202 }
203
204 int regs = SharedRuntime::java_return_convention(sig_bt, vm_parm_regs, cnt);
205 if (regs <= 0) {
206 // We ran out of registers to store the IsInit information for a nullable inline type return.
207 // Since it is only set in the 'call_epilog', we can simply put it on the stack.
208 assert(tf->returns_inline_type_as_fields(), "should have been tested during graph construction");
209 // TODO 8284443 Can we teach the register allocator to reserve a stack slot instead?
210 // mask[--cnt] = STACK_ONLY_mask does not work (test with -XX:+StressGCM)
211 int slot = C->fixed_slots() - 2;
212 if (C->needs_stack_repair()) {
213 slot -= 2; // Account for stack increment value
214 }
215 mask[--cnt].Clear();
216 mask[cnt].Insert(OptoReg::stack2reg(slot));
217 }
218 for (uint i = 0; i < cnt; i++) {
219 mask[i].Clear();
220
221 OptoReg::Name reg1 = OptoReg::as_OptoReg(vm_parm_regs[i].first());
222 if (OptoReg::is_valid(reg1)) {
223 mask[i].Insert(reg1);
224 }
225 OptoReg::Name reg2 = OptoReg::as_OptoReg(vm_parm_regs[i].second());
226 if (OptoReg::is_valid(reg2)) {
227 mask[i].Insert(reg2);
228 }
229 }
230
231 return mask;
232 }
233
234 //---------------------------match---------------------------------------------
235 void Matcher::match( ) {
236 if( MaxLabelRootDepth < 100 ) { // Too small?
237 assert(false, "invalid MaxLabelRootDepth, increase it to 100 minimum");
238 MaxLabelRootDepth = 100;
239 }
240 // One-time initialization of some register masks.
241 init_spill_mask( C->root()->in(1) );
242 if (C->failing()) {
243 return;
244 }
245 _return_addr_mask = return_addr();
246 #ifdef _LP64
247 // Pointers take 2 slots in 64-bit land
248 _return_addr_mask.Insert(OptoReg::add(return_addr(),1));
249 #endif
250
251 // Map Java-signature return types into return register-value
252 // machine registers.
253 _return_values_mask = return_values_mask(C->tf());
254
255 // ---------------
256 // Frame Layout
257
258 // Need the method signature to determine the incoming argument types,
259 // because the types determine which registers the incoming arguments are
260 // in, and this affects the matched code.
261 const TypeTuple *domain = C->tf()->domain_cc();
262 uint argcnt = domain->cnt() - TypeFunc::Parms;
263 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt );
264 VMRegPair *vm_parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt );
265 _parm_regs = NEW_RESOURCE_ARRAY( OptoRegPair, argcnt );
266 _calling_convention_mask = NEW_RESOURCE_ARRAY( RegMask, argcnt );
267 uint i;
268 for( i = 0; i<argcnt; i++ ) {
269 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type();
270 }
271
272 // Pass array of ideal registers and length to USER code (from the AD file)
273 // that will convert this to an array of register numbers.
274 const StartNode *start = C->start();
275 start->calling_convention( sig_bt, vm_parm_regs, argcnt );
276 #ifdef ASSERT
277 // Sanity check users' calling convention. Real handy while trying to
278 // get the initial port correct.
279 { for (uint i = 0; i<argcnt; i++) {
280 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) {
281 assert(domain->field_at(i+TypeFunc::Parms)==Type::HALF, "only allowed on halve" );
282 _parm_regs[i].set_bad();
283 continue;
284 }
285 VMReg parm_reg = vm_parm_regs[i].first();
286 assert(parm_reg->is_valid(), "invalid arg?");
287 if (parm_reg->is_reg()) {
288 OptoReg::Name opto_parm_reg = OptoReg::as_OptoReg(parm_reg);
289 assert(can_be_java_arg(opto_parm_reg) ||
290 C->stub_function() == CAST_FROM_FN_PTR(address, OptoRuntime::rethrow_C) ||
291 opto_parm_reg == inline_cache_reg(),
292 "parameters in register must be preserved by runtime stubs");
293 }
294 for (uint j = 0; j < i; j++) {
295 assert(parm_reg != vm_parm_regs[j].first(),
296 "calling conv. must produce distinct regs");
297 }
298 }
299 }
300 #endif
301
302 // Do some initial frame layout.
303
304 // Compute the old incoming SP (may be called FP) as
305 // OptoReg::stack0() + locks + in_preserve_stack_slots + pad2.
306 _old_SP = C->compute_old_SP();
307 assert( is_even(_old_SP), "must be even" );
308
309 // Compute highest incoming stack argument as
310 // _old_SP + out_preserve_stack_slots + incoming argument size.
311 _in_arg_limit = OptoReg::add(_old_SP, C->out_preserve_stack_slots());
312 assert( is_even(_in_arg_limit), "out_preserve must be even" );
313 for( i = 0; i < argcnt; i++ ) {
314 // Permit args to have no register
315 _calling_convention_mask[i].Clear();
316 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) {
317 _parm_regs[i].set_bad();
318 continue;
319 }
320 // calling_convention returns stack arguments as a count of
321 // slots beyond OptoReg::stack0()/VMRegImpl::stack0. We need to convert this to
322 // the allocators point of view, taking into account all the
323 // preserve area, locks & pad2.
324
325 OptoReg::Name reg1 = warp_incoming_stk_arg(vm_parm_regs[i].first());
326 if (C->failing()) {
327 return;
328 }
329 if( OptoReg::is_valid(reg1))
330 _calling_convention_mask[i].Insert(reg1);
331
332 OptoReg::Name reg2 = warp_incoming_stk_arg(vm_parm_regs[i].second());
333 if (C->failing()) {
334 return;
335 }
336 if( OptoReg::is_valid(reg2))
337 _calling_convention_mask[i].Insert(reg2);
338
339 // Saved biased stack-slot register number
340 _parm_regs[i].set_pair(reg2, reg1);
341 }
342
343 // Finally, make sure the incoming arguments take up an even number of
344 // words, in case the arguments or locals need to contain doubleword stack
345 // slots. The rest of the system assumes that stack slot pairs (in
346 // particular, in the spill area) which look aligned will in fact be
347 // aligned relative to the stack pointer in the target machine. Double
348 // stack slots will always be allocated aligned.
349 _new_SP = OptoReg::Name(align_up(_in_arg_limit, (int)RegMask::SlotsPerLong));
350
351 // Compute highest outgoing stack argument as
352 // _new_SP + out_preserve_stack_slots + max(outgoing argument size).
353 _out_arg_limit = OptoReg::add(_new_SP, C->out_preserve_stack_slots());
354 assert( is_even(_out_arg_limit), "out_preserve must be even" );
355
356 if (!RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1))) {
357 // the compiler cannot represent this method's calling sequence
358 // Bailout. We do not have space to represent all arguments.
359 C->record_method_not_compilable("must be able to represent all call arguments in reg mask");
360 }
361
362 if (C->failing()) return; // bailed out on incoming arg failure
363
364 // ---------------
365 // Collect roots of matcher trees. Every node for which
366 // _shared[_idx] is cleared is guaranteed to not be shared, and thus
367 // can be a valid interior of some tree.
368 find_shared( C->root() );
369 find_shared( C->top() );
370
371 C->print_method(PHASE_BEFORE_MATCHING, 1);
372
373 // Create new ideal node ConP #null even if it does exist in old space
374 // to avoid false sharing if the corresponding mach node is not used.
375 // The corresponding mach node is only used in rare cases for derived
376 // pointers.
377 Node* new_ideal_null = ConNode::make(TypePtr::NULL_PTR);
378
379 // Swap out to old-space; emptying new-space
380 Arena* old = C->swap_old_and_new();
381
382 // Save debug and profile information for nodes in old space:
383 _old_node_note_array = C->node_note_array();
384 if (_old_node_note_array != nullptr) {
385 C->set_node_note_array(new(C->comp_arena()) GrowableArray<Node_Notes*>
386 (C->comp_arena(), _old_node_note_array->length(),
387 0, nullptr));
388 }
389
390 // Pre-size the new_node table to avoid the need for range checks.
391 grow_new_node_array(C->unique());
392
393 // Reset node counter so MachNodes start with _idx at 0
394 int live_nodes = C->live_nodes();
395 C->set_unique(0);
396 C->reset_dead_node_list();
397
398 // Recursively match trees from old space into new space.
399 // Correct leaves of new-space Nodes; they point to old-space.
400 _visited.clear();
401 Node* const n = xform(C->top(), live_nodes);
402 if (C->failing()) return;
403 C->set_cached_top_node(n);
404 if (!C->failing()) {
405 Node* xroot = xform( C->root(), 1 );
406 if (C->failing()) return;
407 if (xroot == nullptr) {
408 Matcher::soft_match_failure(); // recursive matching process failed
409 assert(false, "instruction match failed");
410 C->record_method_not_compilable("instruction match failed");
411 } else {
412 // During matching shared constants were attached to C->root()
413 // because xroot wasn't available yet, so transfer the uses to
414 // the xroot.
415 for( DUIterator_Fast jmax, j = C->root()->fast_outs(jmax); j < jmax; j++ ) {
416 Node* n = C->root()->fast_out(j);
417 if (C->node_arena()->contains(n)) {
418 assert(n->in(0) == C->root(), "should be control user");
419 n->set_req(0, xroot);
420 --j;
421 --jmax;
422 }
423 }
424
425 // Generate new mach node for ConP #null
426 assert(new_ideal_null != nullptr, "sanity");
427 _mach_null = match_tree(new_ideal_null);
428 // Don't set control, it will confuse GCM since there are no uses.
429 // The control will be set when this node is used first time
430 // in find_base_for_derived().
431 assert(_mach_null != nullptr || C->failure_is_artificial(), ""); // bailouts are handled below.
432
433 C->set_root(xroot->is_Root() ? xroot->as_Root() : nullptr);
434
435 #ifdef ASSERT
436 verify_new_nodes_only(xroot);
437 #endif
438 }
439 }
440 if (C->top() == nullptr || C->root() == nullptr) {
441 // New graph lost. This is due to a compilation failure we encountered earlier.
442 stringStream ss;
443 if (C->failure_reason() != nullptr) {
444 ss.print("graph lost: %s", C->failure_reason());
445 } else {
446 assert(C->failure_reason() != nullptr, "graph lost: reason unknown");
447 ss.print("graph lost: reason unknown");
448 }
449 C->record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
450 }
451 if (C->failing()) {
452 // delete old;
453 old->destruct_contents();
454 return;
455 }
456 assert( C->top(), "" );
457 assert( C->root(), "" );
458 validate_null_checks();
459
460 // Now smoke old-space
461 NOT_DEBUG( old->destruct_contents() );
462
463 // ------------------------
464 // Set up save-on-entry registers.
465 Fixup_Save_On_Entry( );
466
467 { // Cleanup mach IR after selection phase is over.
468 Compile::TracePhase tp(_t_postselect_cleanup);
469 do_postselect_cleanup();
470 if (C->failing()) return;
471 assert(verify_after_postselect_cleanup(), "");
472 }
473 }
474
475 //------------------------------Fixup_Save_On_Entry----------------------------
476 // The stated purpose of this routine is to take care of save-on-entry
477 // registers. However, the overall goal of the Match phase is to convert into
478 // machine-specific instructions which have RegMasks to guide allocation.
479 // So what this procedure really does is put a valid RegMask on each input
480 // to the machine-specific variations of all Return, TailCall and Halt
481 // instructions. It also adds edgs to define the save-on-entry values (and of
482 // course gives them a mask).
483
484 static RegMask *init_input_masks( uint size, RegMask &ret_adr, RegMask &fp ) {
485 RegMask *rms = NEW_RESOURCE_ARRAY( RegMask, size );
486 // Do all the pre-defined register masks
487 rms[TypeFunc::Control ] = RegMask::Empty;
488 rms[TypeFunc::I_O ] = RegMask::Empty;
489 rms[TypeFunc::Memory ] = RegMask::Empty;
490 rms[TypeFunc::ReturnAdr] = ret_adr;
491 rms[TypeFunc::FramePtr ] = fp;
492 return rms;
493 }
494
495 int Matcher::scalable_predicate_reg_slots() {
496 assert(Matcher::has_predicated_vectors() && Matcher::supports_scalable_vector(),
497 "scalable predicate vector should be supported");
498 int vector_reg_bit_size = Matcher::scalable_vector_reg_size(T_BYTE) << LogBitsPerByte;
499 // We assume each predicate register is one-eighth of the size of
500 // scalable vector register, one mask bit per vector byte.
501 int predicate_reg_bit_size = vector_reg_bit_size >> 3;
502 // Compute number of slots which is required when scalable predicate
503 // register is spilled. E.g. if scalable vector register is 640 bits,
504 // predicate register is 80 bits, which is 2.5 * slots.
505 // We will round up the slot number to power of 2, which is required
506 // by find_first_set().
507 int slots = predicate_reg_bit_size & (BitsPerInt - 1)
508 ? (predicate_reg_bit_size >> LogBitsPerInt) + 1
509 : predicate_reg_bit_size >> LogBitsPerInt;
510 return round_up_power_of_2(slots);
511 }
512
513 #define NOF_STACK_MASKS (3*13)
514
515 // Create the initial stack mask used by values spilling to the stack.
516 // Disallow any debug info in outgoing argument areas by setting the
517 // initial mask accordingly.
518 void Matcher::init_first_stack_mask() {
519
520 // Allocate storage for spill masks as masks for the appropriate load type.
521 RegMask *rms = (RegMask*)C->comp_arena()->AmallocWords(sizeof(RegMask) * NOF_STACK_MASKS);
522
523 // Initialize empty placeholder masks into the newly allocated arena
524 for (int i = 0; i < NOF_STACK_MASKS; i++) {
525 new (rms + i) RegMask();
526 }
527
528 idealreg2spillmask [Op_RegN] = &rms[0];
529 idealreg2spillmask [Op_RegI] = &rms[1];
530 idealreg2spillmask [Op_RegL] = &rms[2];
531 idealreg2spillmask [Op_RegF] = &rms[3];
532 idealreg2spillmask [Op_RegD] = &rms[4];
533 idealreg2spillmask [Op_RegP] = &rms[5];
534
535 idealreg2debugmask [Op_RegN] = &rms[6];
536 idealreg2debugmask [Op_RegI] = &rms[7];
537 idealreg2debugmask [Op_RegL] = &rms[8];
538 idealreg2debugmask [Op_RegF] = &rms[9];
539 idealreg2debugmask [Op_RegD] = &rms[10];
540 idealreg2debugmask [Op_RegP] = &rms[11];
541
542 idealreg2mhdebugmask[Op_RegN] = &rms[12];
543 idealreg2mhdebugmask[Op_RegI] = &rms[13];
544 idealreg2mhdebugmask[Op_RegL] = &rms[14];
545 idealreg2mhdebugmask[Op_RegF] = &rms[15];
546 idealreg2mhdebugmask[Op_RegD] = &rms[16];
547 idealreg2mhdebugmask[Op_RegP] = &rms[17];
548
549 idealreg2spillmask [Op_VecA] = &rms[18];
550 idealreg2spillmask [Op_VecS] = &rms[19];
551 idealreg2spillmask [Op_VecD] = &rms[20];
552 idealreg2spillmask [Op_VecX] = &rms[21];
553 idealreg2spillmask [Op_VecY] = &rms[22];
554 idealreg2spillmask [Op_VecZ] = &rms[23];
555
556 idealreg2debugmask [Op_VecA] = &rms[24];
557 idealreg2debugmask [Op_VecS] = &rms[25];
558 idealreg2debugmask [Op_VecD] = &rms[26];
559 idealreg2debugmask [Op_VecX] = &rms[27];
560 idealreg2debugmask [Op_VecY] = &rms[28];
561 idealreg2debugmask [Op_VecZ] = &rms[29];
562
563 idealreg2mhdebugmask[Op_VecA] = &rms[30];
564 idealreg2mhdebugmask[Op_VecS] = &rms[31];
565 idealreg2mhdebugmask[Op_VecD] = &rms[32];
566 idealreg2mhdebugmask[Op_VecX] = &rms[33];
567 idealreg2mhdebugmask[Op_VecY] = &rms[34];
568 idealreg2mhdebugmask[Op_VecZ] = &rms[35];
569
570 idealreg2spillmask [Op_RegVectMask] = &rms[36];
571 idealreg2debugmask [Op_RegVectMask] = &rms[37];
572 idealreg2mhdebugmask[Op_RegVectMask] = &rms[38];
573
574 OptoReg::Name i;
575
576 // At first, start with the empty mask
577 C->FIRST_STACK_mask().Clear();
578
579 // Add in the incoming argument area
580 OptoReg::Name init_in = OptoReg::add(_old_SP, C->out_preserve_stack_slots());
581 for (i = init_in; i < _in_arg_limit; i = OptoReg::add(i,1)) {
582 C->FIRST_STACK_mask().Insert(i);
583 }
584
585 // Add in all bits past the outgoing argument area
586 guarantee(RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1)),
587 "must be able to represent all call arguments in reg mask");
588 OptoReg::Name init = _out_arg_limit;
589 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) {
590 C->FIRST_STACK_mask().Insert(i);
591 }
592 // Finally, set the "infinite stack" bit.
593 C->FIRST_STACK_mask().set_AllStack();
594
595 // Make spill masks. Registers for their class, plus FIRST_STACK_mask.
596 RegMask aligned_stack_mask = C->FIRST_STACK_mask();
597 // Keep spill masks aligned.
598 aligned_stack_mask.clear_to_pairs();
599 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack");
600 RegMask scalable_stack_mask = aligned_stack_mask;
601
602 *idealreg2spillmask[Op_RegP] = *idealreg2regmask[Op_RegP];
603 #ifdef _LP64
604 *idealreg2spillmask[Op_RegN] = *idealreg2regmask[Op_RegN];
605 idealreg2spillmask[Op_RegN]->OR(C->FIRST_STACK_mask());
606 idealreg2spillmask[Op_RegP]->OR(aligned_stack_mask);
607 #else
608 idealreg2spillmask[Op_RegP]->OR(C->FIRST_STACK_mask());
609 #endif
610 *idealreg2spillmask[Op_RegI] = *idealreg2regmask[Op_RegI];
611 idealreg2spillmask[Op_RegI]->OR(C->FIRST_STACK_mask());
612 *idealreg2spillmask[Op_RegL] = *idealreg2regmask[Op_RegL];
613 idealreg2spillmask[Op_RegL]->OR(aligned_stack_mask);
614 *idealreg2spillmask[Op_RegF] = *idealreg2regmask[Op_RegF];
615 idealreg2spillmask[Op_RegF]->OR(C->FIRST_STACK_mask());
616 *idealreg2spillmask[Op_RegD] = *idealreg2regmask[Op_RegD];
617 idealreg2spillmask[Op_RegD]->OR(aligned_stack_mask);
618
619 if (Matcher::has_predicated_vectors()) {
620 *idealreg2spillmask[Op_RegVectMask] = *idealreg2regmask[Op_RegVectMask];
621 idealreg2spillmask[Op_RegVectMask]->OR(aligned_stack_mask);
622 } else {
623 *idealreg2spillmask[Op_RegVectMask] = RegMask::Empty;
624 }
625
626 if (Matcher::vector_size_supported(T_BYTE,4)) {
627 *idealreg2spillmask[Op_VecS] = *idealreg2regmask[Op_VecS];
628 idealreg2spillmask[Op_VecS]->OR(C->FIRST_STACK_mask());
629 } else {
630 *idealreg2spillmask[Op_VecS] = RegMask::Empty;
631 }
632
633 if (Matcher::vector_size_supported(T_FLOAT,2)) {
634 // For VecD we need dual alignment and 8 bytes (2 slots) for spills.
635 // RA guarantees such alignment since it is needed for Double and Long values.
636 *idealreg2spillmask[Op_VecD] = *idealreg2regmask[Op_VecD];
637 idealreg2spillmask[Op_VecD]->OR(aligned_stack_mask);
638 } else {
639 *idealreg2spillmask[Op_VecD] = RegMask::Empty;
640 }
641
642 if (Matcher::vector_size_supported(T_FLOAT,4)) {
643 // For VecX we need quadro alignment and 16 bytes (4 slots) for spills.
644 //
645 // RA can use input arguments stack slots for spills but until RA
646 // we don't know frame size and offset of input arg stack slots.
647 //
648 // Exclude last input arg stack slots to avoid spilling vectors there
649 // otherwise vector spills could stomp over stack slots in caller frame.
650 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1);
651 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecX); k++) {
652 aligned_stack_mask.Remove(in);
653 in = OptoReg::add(in, -1);
654 }
655 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecX);
656 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack");
657 *idealreg2spillmask[Op_VecX] = *idealreg2regmask[Op_VecX];
658 idealreg2spillmask[Op_VecX]->OR(aligned_stack_mask);
659 } else {
660 *idealreg2spillmask[Op_VecX] = RegMask::Empty;
661 }
662
663 if (Matcher::vector_size_supported(T_FLOAT,8)) {
664 // For VecY we need octo alignment and 32 bytes (8 slots) for spills.
665 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1);
666 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecY); k++) {
667 aligned_stack_mask.Remove(in);
668 in = OptoReg::add(in, -1);
669 }
670 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecY);
671 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack");
672 *idealreg2spillmask[Op_VecY] = *idealreg2regmask[Op_VecY];
673 idealreg2spillmask[Op_VecY]->OR(aligned_stack_mask);
674 } else {
675 *idealreg2spillmask[Op_VecY] = RegMask::Empty;
676 }
677
678 if (Matcher::vector_size_supported(T_FLOAT,16)) {
679 // For VecZ we need enough alignment and 64 bytes (16 slots) for spills.
680 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1);
681 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecZ); k++) {
682 aligned_stack_mask.Remove(in);
683 in = OptoReg::add(in, -1);
684 }
685 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecZ);
686 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack");
687 *idealreg2spillmask[Op_VecZ] = *idealreg2regmask[Op_VecZ];
688 idealreg2spillmask[Op_VecZ]->OR(aligned_stack_mask);
689 } else {
690 *idealreg2spillmask[Op_VecZ] = RegMask::Empty;
691 }
692
693 if (Matcher::supports_scalable_vector()) {
694 int k = 1;
695 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1);
696 if (Matcher::has_predicated_vectors()) {
697 // Exclude last input arg stack slots to avoid spilling vector register there,
698 // otherwise RegVectMask spills could stomp over stack slots in caller frame.
699 for (; (in >= init_in) && (k < scalable_predicate_reg_slots()); k++) {
700 scalable_stack_mask.Remove(in);
701 in = OptoReg::add(in, -1);
702 }
703
704 // For RegVectMask
705 scalable_stack_mask.clear_to_sets(scalable_predicate_reg_slots());
706 assert(scalable_stack_mask.is_AllStack(), "should be infinite stack");
707 *idealreg2spillmask[Op_RegVectMask] = *idealreg2regmask[Op_RegVectMask];
708 idealreg2spillmask[Op_RegVectMask]->OR(scalable_stack_mask);
709 }
710
711 // Exclude last input arg stack slots to avoid spilling vector register there,
712 // otherwise vector spills could stomp over stack slots in caller frame.
713 for (; (in >= init_in) && (k < scalable_vector_reg_size(T_FLOAT)); k++) {
714 scalable_stack_mask.Remove(in);
715 in = OptoReg::add(in, -1);
716 }
717
718 // For VecA
719 scalable_stack_mask.clear_to_sets(RegMask::SlotsPerVecA);
720 assert(scalable_stack_mask.is_AllStack(), "should be infinite stack");
721 *idealreg2spillmask[Op_VecA] = *idealreg2regmask[Op_VecA];
722 idealreg2spillmask[Op_VecA]->OR(scalable_stack_mask);
723 } else {
724 *idealreg2spillmask[Op_VecA] = RegMask::Empty;
725 }
726
727 if (UseFPUForSpilling) {
728 // This mask logic assumes that the spill operations are
729 // symmetric and that the registers involved are the same size.
730 // On sparc for instance we may have to use 64 bit moves will
731 // kill 2 registers when used with F0-F31.
732 idealreg2spillmask[Op_RegI]->OR(*idealreg2regmask[Op_RegF]);
733 idealreg2spillmask[Op_RegF]->OR(*idealreg2regmask[Op_RegI]);
734 #ifdef _LP64
735 idealreg2spillmask[Op_RegN]->OR(*idealreg2regmask[Op_RegF]);
736 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]);
737 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]);
738 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegD]);
739 #else
740 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegF]);
741 #ifdef ARM
742 // ARM has support for moving 64bit values between a pair of
743 // integer registers and a double register
744 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]);
745 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]);
746 #endif
747 #endif
748 }
749
750 // Make up debug masks. Any spill slot plus callee-save (SOE) registers.
751 // Caller-save (SOC, AS) registers are assumed to be trashable by the various
752 // inline-cache fixup routines.
753 *idealreg2debugmask [Op_RegN] = *idealreg2spillmask[Op_RegN];
754 *idealreg2debugmask [Op_RegI] = *idealreg2spillmask[Op_RegI];
755 *idealreg2debugmask [Op_RegL] = *idealreg2spillmask[Op_RegL];
756 *idealreg2debugmask [Op_RegF] = *idealreg2spillmask[Op_RegF];
757 *idealreg2debugmask [Op_RegD] = *idealreg2spillmask[Op_RegD];
758 *idealreg2debugmask [Op_RegP] = *idealreg2spillmask[Op_RegP];
759 *idealreg2debugmask [Op_RegVectMask] = *idealreg2spillmask[Op_RegVectMask];
760
761 *idealreg2debugmask [Op_VecA] = *idealreg2spillmask[Op_VecA];
762 *idealreg2debugmask [Op_VecS] = *idealreg2spillmask[Op_VecS];
763 *idealreg2debugmask [Op_VecD] = *idealreg2spillmask[Op_VecD];
764 *idealreg2debugmask [Op_VecX] = *idealreg2spillmask[Op_VecX];
765 *idealreg2debugmask [Op_VecY] = *idealreg2spillmask[Op_VecY];
766 *idealreg2debugmask [Op_VecZ] = *idealreg2spillmask[Op_VecZ];
767
768 *idealreg2mhdebugmask[Op_RegN] = *idealreg2spillmask[Op_RegN];
769 *idealreg2mhdebugmask[Op_RegI] = *idealreg2spillmask[Op_RegI];
770 *idealreg2mhdebugmask[Op_RegL] = *idealreg2spillmask[Op_RegL];
771 *idealreg2mhdebugmask[Op_RegF] = *idealreg2spillmask[Op_RegF];
772 *idealreg2mhdebugmask[Op_RegD] = *idealreg2spillmask[Op_RegD];
773 *idealreg2mhdebugmask[Op_RegP] = *idealreg2spillmask[Op_RegP];
774 *idealreg2mhdebugmask[Op_RegVectMask] = *idealreg2spillmask[Op_RegVectMask];
775
776 *idealreg2mhdebugmask[Op_VecA] = *idealreg2spillmask[Op_VecA];
777 *idealreg2mhdebugmask[Op_VecS] = *idealreg2spillmask[Op_VecS];
778 *idealreg2mhdebugmask[Op_VecD] = *idealreg2spillmask[Op_VecD];
779 *idealreg2mhdebugmask[Op_VecX] = *idealreg2spillmask[Op_VecX];
780 *idealreg2mhdebugmask[Op_VecY] = *idealreg2spillmask[Op_VecY];
781 *idealreg2mhdebugmask[Op_VecZ] = *idealreg2spillmask[Op_VecZ];
782
783 // Prevent stub compilations from attempting to reference
784 // callee-saved (SOE) registers from debug info
785 bool exclude_soe = !Compile::current()->is_method_compilation();
786 RegMask* caller_save_mask = exclude_soe ? &caller_save_regmask_exclude_soe : &caller_save_regmask;
787 RegMask* mh_caller_save_mask = exclude_soe ? &mh_caller_save_regmask_exclude_soe : &mh_caller_save_regmask;
788
789 idealreg2debugmask[Op_RegN]->SUBTRACT(*caller_save_mask);
790 idealreg2debugmask[Op_RegI]->SUBTRACT(*caller_save_mask);
791 idealreg2debugmask[Op_RegL]->SUBTRACT(*caller_save_mask);
792 idealreg2debugmask[Op_RegF]->SUBTRACT(*caller_save_mask);
793 idealreg2debugmask[Op_RegD]->SUBTRACT(*caller_save_mask);
794 idealreg2debugmask[Op_RegP]->SUBTRACT(*caller_save_mask);
795 idealreg2debugmask[Op_RegVectMask]->SUBTRACT(*caller_save_mask);
796
797 idealreg2debugmask[Op_VecA]->SUBTRACT(*caller_save_mask);
798 idealreg2debugmask[Op_VecS]->SUBTRACT(*caller_save_mask);
799 idealreg2debugmask[Op_VecD]->SUBTRACT(*caller_save_mask);
800 idealreg2debugmask[Op_VecX]->SUBTRACT(*caller_save_mask);
801 idealreg2debugmask[Op_VecY]->SUBTRACT(*caller_save_mask);
802 idealreg2debugmask[Op_VecZ]->SUBTRACT(*caller_save_mask);
803
804 idealreg2mhdebugmask[Op_RegN]->SUBTRACT(*mh_caller_save_mask);
805 idealreg2mhdebugmask[Op_RegI]->SUBTRACT(*mh_caller_save_mask);
806 idealreg2mhdebugmask[Op_RegL]->SUBTRACT(*mh_caller_save_mask);
807 idealreg2mhdebugmask[Op_RegF]->SUBTRACT(*mh_caller_save_mask);
808 idealreg2mhdebugmask[Op_RegD]->SUBTRACT(*mh_caller_save_mask);
809 idealreg2mhdebugmask[Op_RegP]->SUBTRACT(*mh_caller_save_mask);
810 idealreg2mhdebugmask[Op_RegVectMask]->SUBTRACT(*mh_caller_save_mask);
811
812 idealreg2mhdebugmask[Op_VecA]->SUBTRACT(*mh_caller_save_mask);
813 idealreg2mhdebugmask[Op_VecS]->SUBTRACT(*mh_caller_save_mask);
814 idealreg2mhdebugmask[Op_VecD]->SUBTRACT(*mh_caller_save_mask);
815 idealreg2mhdebugmask[Op_VecX]->SUBTRACT(*mh_caller_save_mask);
816 idealreg2mhdebugmask[Op_VecY]->SUBTRACT(*mh_caller_save_mask);
817 idealreg2mhdebugmask[Op_VecZ]->SUBTRACT(*mh_caller_save_mask);
818 }
819
820 //---------------------------is_save_on_entry----------------------------------
821 bool Matcher::is_save_on_entry(int reg) {
822 return
823 _register_save_policy[reg] == 'E' ||
824 _register_save_policy[reg] == 'A'; // Save-on-entry register?
825 }
826
827 //---------------------------Fixup_Save_On_Entry-------------------------------
828 void Matcher::Fixup_Save_On_Entry( ) {
829 init_first_stack_mask();
830
831 Node *root = C->root(); // Short name for root
832 // Count number of save-on-entry registers.
833 uint soe_cnt = number_of_saved_registers();
834 uint i;
835
836 // Find the procedure Start Node
837 StartNode *start = C->start();
838 assert( start, "Expect a start node" );
839
840 // Input RegMask array shared by all Returns.
841 // The type for doubles and longs has a count of 2, but
842 // there is only 1 returned value
843 uint ret_edge_cnt = C->tf()->range_cc()->cnt();
844 RegMask *ret_rms = init_input_masks( ret_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
845 for (i = TypeFunc::Parms; i < ret_edge_cnt; i++) {
846 ret_rms[i] = _return_values_mask[i-TypeFunc::Parms];
847 }
848
849 // Input RegMask array shared by all ForwardExceptions
850 uint forw_exc_edge_cnt = TypeFunc::Parms;
851 RegMask* forw_exc_rms = init_input_masks( forw_exc_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
852
853 // Input RegMask array shared by all Rethrows.
854 uint reth_edge_cnt = TypeFunc::Parms+1;
855 RegMask *reth_rms = init_input_masks( reth_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
856 // Rethrow takes exception oop only, but in the argument 0 slot.
857 OptoReg::Name reg = find_receiver();
858 if (reg >= 0) {
859 reth_rms[TypeFunc::Parms] = mreg2regmask[reg];
860 #ifdef _LP64
861 // Need two slots for ptrs in 64-bit land
862 reth_rms[TypeFunc::Parms].Insert(OptoReg::add(OptoReg::Name(reg), 1));
863 #endif
864 }
865
866 // Input RegMask array shared by all TailCalls
867 uint tail_call_edge_cnt = TypeFunc::Parms+2;
868 RegMask *tail_call_rms = init_input_masks( tail_call_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
869
870 // Input RegMask array shared by all TailJumps
871 uint tail_jump_edge_cnt = TypeFunc::Parms+2;
872 RegMask *tail_jump_rms = init_input_masks( tail_jump_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
873
874 // TailCalls have 2 returned values (target & moop), whose masks come
875 // from the usual MachNode/MachOper mechanism. Find a sample
876 // TailCall to extract these masks and put the correct masks into
877 // the tail_call_rms array.
878 for( i=1; i < root->req(); i++ ) {
879 MachReturnNode *m = root->in(i)->as_MachReturn();
880 if( m->ideal_Opcode() == Op_TailCall ) {
881 tail_call_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0);
882 tail_call_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1);
883 break;
884 }
885 }
886
887 // TailJumps have 2 returned values (target & ex_oop), whose masks come
888 // from the usual MachNode/MachOper mechanism. Find a sample
889 // TailJump to extract these masks and put the correct masks into
890 // the tail_jump_rms array.
891 for( i=1; i < root->req(); i++ ) {
892 MachReturnNode *m = root->in(i)->as_MachReturn();
893 if( m->ideal_Opcode() == Op_TailJump ) {
894 tail_jump_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0);
895 tail_jump_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1);
896 break;
897 }
898 }
899
900 // Input RegMask array shared by all Halts
901 uint halt_edge_cnt = TypeFunc::Parms;
902 RegMask *halt_rms = init_input_masks( halt_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
903
904 // Capture the return input masks into each exit flavor
905 for( i=1; i < root->req(); i++ ) {
906 MachReturnNode *exit = root->in(i)->as_MachReturn();
907 switch( exit->ideal_Opcode() ) {
908 case Op_Return : exit->_in_rms = ret_rms; break;
909 case Op_Rethrow : exit->_in_rms = reth_rms; break;
910 case Op_TailCall : exit->_in_rms = tail_call_rms; break;
911 case Op_TailJump : exit->_in_rms = tail_jump_rms; break;
912 case Op_ForwardException: exit->_in_rms = forw_exc_rms; break;
913 case Op_Halt : exit->_in_rms = halt_rms; break;
914 default : ShouldNotReachHere();
915 }
916 }
917
918 // Next unused projection number from Start.
919 int proj_cnt = C->tf()->domain_cc()->cnt();
920
921 // Do all the save-on-entry registers. Make projections from Start for
922 // them, and give them a use at the exit points. To the allocator, they
923 // look like incoming register arguments.
924 for( i = 0; i < _last_Mach_Reg; i++ ) {
925 if( is_save_on_entry(i) ) {
926
927 // Add the save-on-entry to the mask array
928 ret_rms [ ret_edge_cnt] = mreg2regmask[i];
929 reth_rms [ reth_edge_cnt] = mreg2regmask[i];
930 tail_call_rms[tail_call_edge_cnt] = mreg2regmask[i];
931 tail_jump_rms[tail_jump_edge_cnt] = mreg2regmask[i];
932 forw_exc_rms [ forw_exc_edge_cnt] = mreg2regmask[i];
933 // Halts need the SOE registers, but only in the stack as debug info.
934 // A just-prior uncommon-trap or deoptimization will use the SOE regs.
935 halt_rms [ halt_edge_cnt] = *idealreg2spillmask[_register_save_type[i]];
936
937 Node *mproj;
938
939 // Is this a RegF low half of a RegD? Double up 2 adjacent RegF's
940 // into a single RegD.
941 if( (i&1) == 0 &&
942 _register_save_type[i ] == Op_RegF &&
943 _register_save_type[i+1] == Op_RegF &&
944 is_save_on_entry(i+1) ) {
945 // Add other bit for double
946 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1));
947 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1));
948 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1));
949 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1));
950 forw_exc_rms [ forw_exc_edge_cnt].Insert(OptoReg::Name(i+1));
951 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1));
952 mproj = new MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegD );
953 proj_cnt += 2; // Skip 2 for doubles
954 }
955 else if( (i&1) == 1 && // Else check for high half of double
956 _register_save_type[i-1] == Op_RegF &&
957 _register_save_type[i ] == Op_RegF &&
958 is_save_on_entry(i-1) ) {
959 ret_rms [ ret_edge_cnt] = RegMask::Empty;
960 reth_rms [ reth_edge_cnt] = RegMask::Empty;
961 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty;
962 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty;
963 forw_exc_rms [ forw_exc_edge_cnt] = RegMask::Empty;
964 halt_rms [ halt_edge_cnt] = RegMask::Empty;
965 mproj = C->top();
966 }
967 // Is this a RegI low half of a RegL? Double up 2 adjacent RegI's
968 // into a single RegL.
969 else if( (i&1) == 0 &&
970 _register_save_type[i ] == Op_RegI &&
971 _register_save_type[i+1] == Op_RegI &&
972 is_save_on_entry(i+1) ) {
973 // Add other bit for long
974 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1));
975 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1));
976 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1));
977 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1));
978 forw_exc_rms [ forw_exc_edge_cnt].Insert(OptoReg::Name(i+1));
979 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1));
980 mproj = new MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegL );
981 proj_cnt += 2; // Skip 2 for longs
982 }
983 else if( (i&1) == 1 && // Else check for high half of long
984 _register_save_type[i-1] == Op_RegI &&
985 _register_save_type[i ] == Op_RegI &&
986 is_save_on_entry(i-1) ) {
987 ret_rms [ ret_edge_cnt] = RegMask::Empty;
988 reth_rms [ reth_edge_cnt] = RegMask::Empty;
989 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty;
990 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty;
991 forw_exc_rms [ forw_exc_edge_cnt] = RegMask::Empty;
992 halt_rms [ halt_edge_cnt] = RegMask::Empty;
993 mproj = C->top();
994 } else {
995 // Make a projection for it off the Start
996 mproj = new MachProjNode( start, proj_cnt++, ret_rms[ret_edge_cnt], _register_save_type[i] );
997 }
998
999 ret_edge_cnt ++;
1000 reth_edge_cnt ++;
1001 tail_call_edge_cnt ++;
1002 tail_jump_edge_cnt ++;
1003 forw_exc_edge_cnt++;
1004 halt_edge_cnt ++;
1005
1006 // Add a use of the SOE register to all exit paths
1007 for (uint j=1; j < root->req(); j++) {
1008 root->in(j)->add_req(mproj);
1009 }
1010 } // End of if a save-on-entry register
1011 } // End of for all machine registers
1012 }
1013
1014 //------------------------------init_spill_mask--------------------------------
1015 void Matcher::init_spill_mask( Node *ret ) {
1016 if( idealreg2regmask[Op_RegI] ) return; // One time only init
1017
1018 OptoReg::c_frame_pointer = c_frame_pointer();
1019 c_frame_ptr_mask = c_frame_pointer();
1020 #ifdef _LP64
1021 // pointers are twice as big
1022 c_frame_ptr_mask.Insert(OptoReg::add(c_frame_pointer(),1));
1023 #endif
1024
1025 // Start at OptoReg::stack0()
1026 STACK_ONLY_mask.Clear();
1027 OptoReg::Name init = OptoReg::stack2reg(0);
1028 // STACK_ONLY_mask is all stack bits
1029 OptoReg::Name i;
1030 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1))
1031 STACK_ONLY_mask.Insert(i);
1032 // Also set the "infinite stack" bit.
1033 STACK_ONLY_mask.set_AllStack();
1034
1035 for (i = OptoReg::Name(0); i < OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i, 1)) {
1036 // Copy the register names over into the shared world.
1037 // SharedInfo::regName[i] = regName[i];
1038 // Handy RegMasks per machine register
1039 mreg2regmask[i].Insert(i);
1040
1041 // Set up regmasks used to exclude save-on-call (and always-save) registers from debug masks.
1042 if (_register_save_policy[i] == 'C' ||
1043 _register_save_policy[i] == 'A') {
1044 caller_save_regmask.Insert(i);
1045 mh_caller_save_regmask.Insert(i);
1046 }
1047 // Exclude save-on-entry registers from debug masks for stub compilations.
1048 if (_register_save_policy[i] == 'C' ||
1049 _register_save_policy[i] == 'A' ||
1050 _register_save_policy[i] == 'E') {
1051 caller_save_regmask_exclude_soe.Insert(i);
1052 mh_caller_save_regmask_exclude_soe.Insert(i);
1053 }
1054 }
1055
1056 // Also exclude the register we use to save the SP for MethodHandle
1057 // invokes to from the corresponding MH debug masks
1058 const RegMask sp_save_mask = method_handle_invoke_SP_save_mask();
1059 mh_caller_save_regmask.OR(sp_save_mask);
1060 mh_caller_save_regmask_exclude_soe.OR(sp_save_mask);
1061
1062 // Grab the Frame Pointer
1063 Node *fp = ret->in(TypeFunc::FramePtr);
1064 // Share frame pointer while making spill ops
1065 set_shared(fp);
1066
1067 // Get the ADLC notion of the right regmask, for each basic type.
1068 #ifdef _LP64
1069 idealreg2regmask[Op_RegN] = regmask_for_ideal_register(Op_RegN, ret);
1070 #endif
1071 idealreg2regmask[Op_RegI] = regmask_for_ideal_register(Op_RegI, ret);
1072 idealreg2regmask[Op_RegP] = regmask_for_ideal_register(Op_RegP, ret);
1073 idealreg2regmask[Op_RegF] = regmask_for_ideal_register(Op_RegF, ret);
1074 idealreg2regmask[Op_RegD] = regmask_for_ideal_register(Op_RegD, ret);
1075 idealreg2regmask[Op_RegL] = regmask_for_ideal_register(Op_RegL, ret);
1076 idealreg2regmask[Op_VecA] = regmask_for_ideal_register(Op_VecA, ret);
1077 idealreg2regmask[Op_VecS] = regmask_for_ideal_register(Op_VecS, ret);
1078 idealreg2regmask[Op_VecD] = regmask_for_ideal_register(Op_VecD, ret);
1079 idealreg2regmask[Op_VecX] = regmask_for_ideal_register(Op_VecX, ret);
1080 idealreg2regmask[Op_VecY] = regmask_for_ideal_register(Op_VecY, ret);
1081 idealreg2regmask[Op_VecZ] = regmask_for_ideal_register(Op_VecZ, ret);
1082 idealreg2regmask[Op_RegVectMask] = regmask_for_ideal_register(Op_RegVectMask, ret);
1083 }
1084
1085 #ifdef ASSERT
1086 static void match_alias_type(Compile* C, Node* n, Node* m) {
1087 if (!VerifyAliases) return; // do not go looking for trouble by default
1088 const TypePtr* nat = n->adr_type();
1089 const TypePtr* mat = m->adr_type();
1090 int nidx = C->get_alias_index(nat);
1091 int midx = C->get_alias_index(mat);
1092 // Detune the assert for cases like (AndI 0xFF (LoadB p)).
1093 if (nidx == Compile::AliasIdxTop && midx >= Compile::AliasIdxRaw) {
1094 for (uint i = 1; i < n->req(); i++) {
1095 Node* n1 = n->in(i);
1096 const TypePtr* n1at = n1->adr_type();
1097 if (n1at != nullptr) {
1098 nat = n1at;
1099 nidx = C->get_alias_index(n1at);
1100 }
1101 }
1102 }
1103 // %%% Kludgery. Instead, fix ideal adr_type methods for all these cases:
1104 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxRaw) {
1105 switch (n->Opcode()) {
1106 case Op_PrefetchAllocation:
1107 nidx = Compile::AliasIdxRaw;
1108 nat = TypeRawPtr::BOTTOM;
1109 break;
1110 }
1111 }
1112 if (nidx == Compile::AliasIdxRaw && midx == Compile::AliasIdxTop) {
1113 switch (n->Opcode()) {
1114 case Op_ClearArray:
1115 midx = Compile::AliasIdxRaw;
1116 mat = TypeRawPtr::BOTTOM;
1117 break;
1118 }
1119 }
1120 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxBot) {
1121 switch (n->Opcode()) {
1122 case Op_Return:
1123 case Op_Rethrow:
1124 case Op_Halt:
1125 case Op_TailCall:
1126 case Op_TailJump:
1127 case Op_ForwardException:
1128 nidx = Compile::AliasIdxBot;
1129 nat = TypePtr::BOTTOM;
1130 break;
1131 }
1132 }
1133 if (nidx == Compile::AliasIdxBot && midx == Compile::AliasIdxTop) {
1134 switch (n->Opcode()) {
1135 case Op_StrComp:
1136 case Op_StrEquals:
1137 case Op_StrIndexOf:
1138 case Op_StrIndexOfChar:
1139 case Op_AryEq:
1140 case Op_VectorizedHashCode:
1141 case Op_CountPositives:
1142 case Op_MemBarVolatile:
1143 case Op_MemBarCPUOrder: // %%% these ideals should have narrower adr_type?
1144 case Op_StrInflatedCopy:
1145 case Op_StrCompressedCopy:
1146 case Op_OnSpinWait:
1147 case Op_EncodeISOArray:
1148 nidx = Compile::AliasIdxTop;
1149 nat = nullptr;
1150 break;
1151 }
1152 }
1153 if (nidx != midx) {
1154 if (PrintOpto || (PrintMiscellaneous && (WizardMode || Verbose))) {
1155 tty->print_cr("==== Matcher alias shift %d => %d", nidx, midx);
1156 n->dump();
1157 m->dump();
1158 }
1159 assert(C->subsume_loads() && C->must_alias(nat, midx),
1160 "must not lose alias info when matching");
1161 }
1162 }
1163 #endif
1164
1165 //------------------------------xform------------------------------------------
1166 // Given a Node in old-space, Match him (Label/Reduce) to produce a machine
1167 // Node in new-space. Given a new-space Node, recursively walk his children.
1168 Node *Matcher::transform( Node *n ) { ShouldNotCallThis(); return n; }
1169 Node *Matcher::xform( Node *n, int max_stack ) {
1170 // Use one stack to keep both: child's node/state and parent's node/index
1171 MStack mstack(max_stack * 2 * 2); // usually: C->live_nodes() * 2 * 2
1172 mstack.push(n, Visit, nullptr, -1); // set null as parent to indicate root
1173 while (mstack.is_nonempty()) {
1174 C->check_node_count(NodeLimitFudgeFactor, "too many nodes matching instructions");
1175 if (C->failing()) return nullptr;
1176 n = mstack.node(); // Leave node on stack
1177 Node_State nstate = mstack.state();
1178 if (nstate == Visit) {
1179 mstack.set_state(Post_Visit);
1180 Node *oldn = n;
1181 // Old-space or new-space check
1182 if (!C->node_arena()->contains(n)) {
1183 // Old space!
1184 Node* m;
1185 if (has_new_node(n)) { // Not yet Label/Reduced
1186 m = new_node(n);
1187 } else {
1188 if (!is_dontcare(n)) { // Matcher can match this guy
1189 // Calls match special. They match alone with no children.
1190 // Their children, the incoming arguments, match normally.
1191 m = n->is_SafePoint() ? match_sfpt(n->as_SafePoint()):match_tree(n);
1192 if (C->failing()) return nullptr;
1193 if (m == nullptr) { Matcher::soft_match_failure(); return nullptr; }
1194 if (n->is_MemBar()) {
1195 m->as_MachMemBar()->set_adr_type(n->adr_type());
1196 }
1197 } else { // Nothing the matcher cares about
1198 if (n->is_Proj() && n->in(0) != nullptr && n->in(0)->is_Multi()) { // Projections?
1199 // Convert to machine-dependent projection
1200 RegMask* mask = nullptr;
1201 if (n->in(0)->is_Call() && n->in(0)->as_Call()->tf()->returns_inline_type_as_fields()) {
1202 mask = return_values_mask(n->in(0)->as_Call()->tf());
1203 }
1204 m = n->in(0)->as_Multi()->match(n->as_Proj(), this, mask);
1205 NOT_PRODUCT(record_new2old(m, n);)
1206 if (m->in(0) != nullptr) // m might be top
1207 collect_null_checks(m, n);
1208 } else { // Else just a regular 'ol guy
1209 m = n->clone(); // So just clone into new-space
1210 NOT_PRODUCT(record_new2old(m, n);)
1211 // Def-Use edges will be added incrementally as Uses
1212 // of this node are matched.
1213 assert(m->outcnt() == 0, "no Uses of this clone yet");
1214 }
1215 }
1216
1217 set_new_node(n, m); // Map old to new
1218 if (_old_node_note_array != nullptr) {
1219 Node_Notes* nn = C->locate_node_notes(_old_node_note_array,
1220 n->_idx);
1221 C->set_node_notes_at(m->_idx, nn);
1222 }
1223 debug_only(match_alias_type(C, n, m));
1224 }
1225 n = m; // n is now a new-space node
1226 mstack.set_node(n);
1227 }
1228
1229 // New space!
1230 if (_visited.test_set(n->_idx)) continue; // while(mstack.is_nonempty())
1231
1232 int i;
1233 // Put precedence edges on stack first (match them last).
1234 for (i = oldn->req(); (uint)i < oldn->len(); i++) {
1235 Node *m = oldn->in(i);
1236 if (m == nullptr) break;
1237 // set -1 to call add_prec() instead of set_req() during Step1
1238 mstack.push(m, Visit, n, -1);
1239 }
1240
1241 // Handle precedence edges for interior nodes
1242 for (i = n->len()-1; (uint)i >= n->req(); i--) {
1243 Node *m = n->in(i);
1244 if (m == nullptr || C->node_arena()->contains(m)) continue;
1245 n->rm_prec(i);
1246 // set -1 to call add_prec() instead of set_req() during Step1
1247 mstack.push(m, Visit, n, -1);
1248 }
1249
1250 // For constant debug info, I'd rather have unmatched constants.
1251 int cnt = n->req();
1252 JVMState* jvms = n->jvms();
1253 int debug_cnt = jvms ? jvms->debug_start() : cnt;
1254
1255 // Now do only debug info. Clone constants rather than matching.
1256 // Constants are represented directly in the debug info without
1257 // the need for executable machine instructions.
1258 // Monitor boxes are also represented directly.
1259 for (i = cnt - 1; i >= debug_cnt; --i) { // For all debug inputs do
1260 Node *m = n->in(i); // Get input
1261 int op = m->Opcode();
1262 assert((op == Op_BoxLock) == jvms->is_monitor_use(i), "boxes only at monitor sites");
1263 if( op == Op_ConI || op == Op_ConP || op == Op_ConN || op == Op_ConNKlass ||
1264 op == Op_ConF || op == Op_ConD || op == Op_ConL
1265 // || op == Op_BoxLock // %%%% enable this and remove (+++) in chaitin.cpp
1266 ) {
1267 m = m->clone();
1268 NOT_PRODUCT(record_new2old(m, n));
1269 mstack.push(m, Post_Visit, n, i); // Don't need to visit
1270 mstack.push(m->in(0), Visit, m, 0);
1271 } else {
1272 mstack.push(m, Visit, n, i);
1273 }
1274 }
1275
1276 // And now walk his children, and convert his inputs to new-space.
1277 for( ; i >= 0; --i ) { // For all normal inputs do
1278 Node *m = n->in(i); // Get input
1279 if(m != nullptr)
1280 mstack.push(m, Visit, n, i);
1281 }
1282
1283 }
1284 else if (nstate == Post_Visit) {
1285 // Set xformed input
1286 Node *p = mstack.parent();
1287 if (p != nullptr) { // root doesn't have parent
1288 int i = (int)mstack.index();
1289 if (i >= 0)
1290 p->set_req(i, n); // required input
1291 else if (i == -1)
1292 p->add_prec(n); // precedence input
1293 else
1294 ShouldNotReachHere();
1295 }
1296 mstack.pop(); // remove processed node from stack
1297 }
1298 else {
1299 ShouldNotReachHere();
1300 }
1301 } // while (mstack.is_nonempty())
1302 return n; // Return new-space Node
1303 }
1304
1305 //------------------------------warp_outgoing_stk_arg------------------------
1306 OptoReg::Name Matcher::warp_outgoing_stk_arg( VMReg reg, OptoReg::Name begin_out_arg_area, OptoReg::Name &out_arg_limit_per_call ) {
1307 // Convert outgoing argument location to a pre-biased stack offset
1308 if (reg->is_stack()) {
1309 OptoReg::Name warped = reg->reg2stack();
1310 // Adjust the stack slot offset to be the register number used
1311 // by the allocator.
1312 warped = OptoReg::add(begin_out_arg_area, warped);
1313 // Keep track of the largest numbered stack slot used for an arg.
1314 // Largest used slot per call-site indicates the amount of stack
1315 // that is killed by the call.
1316 if( warped >= out_arg_limit_per_call )
1317 out_arg_limit_per_call = OptoReg::add(warped,1);
1318 if (!RegMask::can_represent_arg(warped)) {
1319 // Bailout. For example not enough space on stack for all arguments. Happens for methods with too many arguments.
1320 C->record_method_not_compilable("unsupported calling sequence");
1321 return OptoReg::Bad;
1322 }
1323 return warped;
1324 }
1325 return OptoReg::as_OptoReg(reg);
1326 }
1327
1328
1329 //------------------------------match_sfpt-------------------------------------
1330 // Helper function to match call instructions. Calls match special.
1331 // They match alone with no children. Their children, the incoming
1332 // arguments, match normally.
1333 MachNode *Matcher::match_sfpt( SafePointNode *sfpt ) {
1334 MachSafePointNode *msfpt = nullptr;
1335 MachCallNode *mcall = nullptr;
1336 uint cnt;
1337 // Split out case for SafePoint vs Call
1338 CallNode *call;
1339 const TypeTuple *domain;
1340 ciMethod* method = nullptr;
1341 bool is_method_handle_invoke = false; // for special kill effects
1342 if( sfpt->is_Call() ) {
1343 call = sfpt->as_Call();
1344 domain = call->tf()->domain_cc();
1345 cnt = domain->cnt();
1346
1347 // Match just the call, nothing else
1348 MachNode *m = match_tree(call);
1349 if (C->failing()) return nullptr;
1350 if( m == nullptr ) { Matcher::soft_match_failure(); return nullptr; }
1351
1352 // Copy data from the Ideal SafePoint to the machine version
1353 mcall = m->as_MachCall();
1354
1355 mcall->set_tf( call->tf());
1356 mcall->set_entry_point( call->entry_point());
1357 mcall->set_cnt( call->cnt());
1358 mcall->set_guaranteed_safepoint(call->guaranteed_safepoint());
1359
1360 if( mcall->is_MachCallJava() ) {
1361 MachCallJavaNode *mcall_java = mcall->as_MachCallJava();
1362 const CallJavaNode *call_java = call->as_CallJava();
1363 assert(call_java->validate_symbolic_info(), "inconsistent info");
1364 method = call_java->method();
1365 mcall_java->_method = method;
1366 mcall_java->_optimized_virtual = call_java->is_optimized_virtual();
1367 is_method_handle_invoke = call_java->is_method_handle_invoke();
1368 mcall_java->_method_handle_invoke = is_method_handle_invoke;
1369 mcall_java->_override_symbolic_info = call_java->override_symbolic_info();
1370 mcall_java->_arg_escape = call_java->arg_escape();
1371 if (is_method_handle_invoke) {
1372 C->set_has_method_handle_invokes(true);
1373 }
1374 if( mcall_java->is_MachCallStaticJava() )
1375 mcall_java->as_MachCallStaticJava()->_name =
1376 call_java->as_CallStaticJava()->_name;
1377 if( mcall_java->is_MachCallDynamicJava() )
1378 mcall_java->as_MachCallDynamicJava()->_vtable_index =
1379 call_java->as_CallDynamicJava()->_vtable_index;
1380 }
1381 else if( mcall->is_MachCallRuntime() ) {
1382 MachCallRuntimeNode* mach_call_rt = mcall->as_MachCallRuntime();
1383 mach_call_rt->_name = call->as_CallRuntime()->_name;
1384 mach_call_rt->_leaf_no_fp = call->is_CallLeafNoFP();
1385 }
1386 msfpt = mcall;
1387 }
1388 // This is a non-call safepoint
1389 else {
1390 call = nullptr;
1391 domain = nullptr;
1392 MachNode *mn = match_tree(sfpt);
1393 if (C->failing()) return nullptr;
1394 msfpt = mn->as_MachSafePoint();
1395 cnt = TypeFunc::Parms;
1396 }
1397 msfpt->_has_ea_local_in_scope = sfpt->has_ea_local_in_scope();
1398
1399 // Advertise the correct memory effects (for anti-dependence computation).
1400 msfpt->set_adr_type(sfpt->adr_type());
1401
1402 // Allocate a private array of RegMasks. These RegMasks are not shared.
1403 msfpt->_in_rms = NEW_RESOURCE_ARRAY( RegMask, cnt );
1404 // Empty them all.
1405 for (uint i = 0; i < cnt; i++) ::new (&(msfpt->_in_rms[i])) RegMask();
1406
1407 // Do all the pre-defined non-Empty register masks
1408 msfpt->_in_rms[TypeFunc::ReturnAdr] = _return_addr_mask;
1409 msfpt->_in_rms[TypeFunc::FramePtr ] = c_frame_ptr_mask;
1410
1411 // Place first outgoing argument can possibly be put.
1412 OptoReg::Name begin_out_arg_area = OptoReg::add(_new_SP, C->out_preserve_stack_slots());
1413 assert( is_even(begin_out_arg_area), "" );
1414 // Compute max outgoing register number per call site.
1415 OptoReg::Name out_arg_limit_per_call = begin_out_arg_area;
1416 // Calls to C may hammer extra stack slots above and beyond any arguments.
1417 // These are usually backing store for register arguments for varargs.
1418 if( call != nullptr && call->is_CallRuntime() )
1419 out_arg_limit_per_call = OptoReg::add(out_arg_limit_per_call,C->varargs_C_out_slots_killed());
1420
1421
1422 // Do the normal argument list (parameters) register masks
1423 // Null entry point is a special cast where the target of the call
1424 // is in a register.
1425 int adj = (call != nullptr && call->entry_point() == nullptr) ? 1 : 0;
1426 int argcnt = cnt - TypeFunc::Parms - adj;
1427 if( argcnt > 0 ) { // Skip it all if we have no args
1428 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt );
1429 VMRegPair *parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt );
1430 int i;
1431 for( i = 0; i < argcnt; i++ ) {
1432 sig_bt[i] = domain->field_at(i+TypeFunc::Parms+adj)->basic_type();
1433 }
1434 // V-call to pick proper calling convention
1435 call->calling_convention( sig_bt, parm_regs, argcnt );
1436
1437 #ifdef ASSERT
1438 // Sanity check users' calling convention. Really handy during
1439 // the initial porting effort. Fairly expensive otherwise.
1440 { for (int i = 0; i<argcnt; i++) {
1441 if( !parm_regs[i].first()->is_valid() &&
1442 !parm_regs[i].second()->is_valid() ) continue;
1443 VMReg reg1 = parm_regs[i].first();
1444 VMReg reg2 = parm_regs[i].second();
1445 for (int j = 0; j < i; j++) {
1446 if( !parm_regs[j].first()->is_valid() &&
1447 !parm_regs[j].second()->is_valid() ) continue;
1448 VMReg reg3 = parm_regs[j].first();
1449 VMReg reg4 = parm_regs[j].second();
1450 if( !reg1->is_valid() ) {
1451 assert( !reg2->is_valid(), "valid halvsies" );
1452 } else if( !reg3->is_valid() ) {
1453 assert( !reg4->is_valid(), "valid halvsies" );
1454 } else {
1455 assert( reg1 != reg2, "calling conv. must produce distinct regs");
1456 assert( reg1 != reg3, "calling conv. must produce distinct regs");
1457 assert( reg1 != reg4, "calling conv. must produce distinct regs");
1458 assert( reg2 != reg3, "calling conv. must produce distinct regs");
1459 assert( reg2 != reg4 || !reg2->is_valid(), "calling conv. must produce distinct regs");
1460 assert( reg3 != reg4, "calling conv. must produce distinct regs");
1461 }
1462 }
1463 }
1464 }
1465 #endif
1466
1467 // Visit each argument. Compute its outgoing register mask.
1468 // Return results now can have 2 bits returned.
1469 // Compute max over all outgoing arguments both per call-site
1470 // and over the entire method.
1471 for( i = 0; i < argcnt; i++ ) {
1472 // Address of incoming argument mask to fill in
1473 RegMask *rm = &mcall->_in_rms[i+TypeFunc::Parms+adj];
1474 VMReg first = parm_regs[i].first();
1475 VMReg second = parm_regs[i].second();
1476 if(!first->is_valid() &&
1477 !second->is_valid()) {
1478 continue; // Avoid Halves
1479 }
1480 // Handle case where arguments are in vector registers.
1481 if(call->in(TypeFunc::Parms + i)->bottom_type()->isa_vect()) {
1482 OptoReg::Name reg_fst = OptoReg::as_OptoReg(first);
1483 OptoReg::Name reg_snd = OptoReg::as_OptoReg(second);
1484 assert (reg_fst <= reg_snd, "fst=%d snd=%d", reg_fst, reg_snd);
1485 for (OptoReg::Name r = reg_fst; r <= reg_snd; r++) {
1486 rm->Insert(r);
1487 }
1488 }
1489 // Grab first register, adjust stack slots and insert in mask.
1490 OptoReg::Name reg1 = warp_outgoing_stk_arg(first, begin_out_arg_area, out_arg_limit_per_call );
1491 if (C->failing()) {
1492 return nullptr;
1493 }
1494 if (OptoReg::is_valid(reg1)) {
1495 rm->Insert( reg1 );
1496 }
1497 // Grab second register (if any), adjust stack slots and insert in mask.
1498 OptoReg::Name reg2 = warp_outgoing_stk_arg(second, begin_out_arg_area, out_arg_limit_per_call );
1499 if (C->failing()) {
1500 return nullptr;
1501 }
1502 if (OptoReg::is_valid(reg2)) {
1503 rm->Insert( reg2 );
1504 }
1505 } // End of for all arguments
1506 }
1507
1508 // Compute the max stack slot killed by any call. These will not be
1509 // available for debug info, and will be used to adjust FIRST_STACK_mask
1510 // after all call sites have been visited.
1511 if( _out_arg_limit < out_arg_limit_per_call)
1512 _out_arg_limit = out_arg_limit_per_call;
1513
1514 if (mcall) {
1515 // Kill the outgoing argument area, including any non-argument holes and
1516 // any legacy C-killed slots. Use Fat-Projections to do the killing.
1517 // Since the max-per-method covers the max-per-call-site and debug info
1518 // is excluded on the max-per-method basis, debug info cannot land in
1519 // this killed area.
1520 uint r_cnt = mcall->tf()->range_sig()->cnt();
1521 MachProjNode *proj = new MachProjNode( mcall, r_cnt+10000, RegMask::Empty, MachProjNode::fat_proj );
1522 if (!RegMask::can_represent_arg(OptoReg::Name(out_arg_limit_per_call-1))) {
1523 // Bailout. We do not have space to represent all arguments.
1524 C->record_method_not_compilable("unsupported outgoing calling sequence");
1525 } else {
1526 for (int i = begin_out_arg_area; i < out_arg_limit_per_call; i++)
1527 proj->_rout.Insert(OptoReg::Name(i));
1528 }
1529 if (proj->_rout.is_NotEmpty()) {
1530 push_projection(proj);
1531 }
1532 }
1533 // Transfer the safepoint information from the call to the mcall
1534 // Move the JVMState list
1535 msfpt->set_jvms(sfpt->jvms());
1536 for (JVMState* jvms = msfpt->jvms(); jvms; jvms = jvms->caller()) {
1537 jvms->set_map(sfpt);
1538 }
1539
1540 // Debug inputs begin just after the last incoming parameter
1541 assert((mcall == nullptr) || (mcall->jvms() == nullptr) ||
1542 (mcall->jvms()->debug_start() + mcall->_jvmadj == mcall->tf()->domain_cc()->cnt()), "");
1543
1544 // Add additional edges.
1545 if (msfpt->mach_constant_base_node_input() != (uint)-1 && !msfpt->is_MachCallLeaf()) {
1546 // For these calls we can not add MachConstantBase in expand(), as the
1547 // ins are not complete then.
1548 msfpt->ins_req(msfpt->mach_constant_base_node_input(), C->mach_constant_base_node());
1549 if (msfpt->jvms() &&
1550 msfpt->mach_constant_base_node_input() <= msfpt->jvms()->debug_start() + msfpt->_jvmadj) {
1551 // We added an edge before jvms, so we must adapt the position of the ins.
1552 msfpt->jvms()->adapt_position(+1);
1553 }
1554 }
1555
1556 // Registers killed by the call are set in the local scheduling pass
1557 // of Global Code Motion.
1558 return msfpt;
1559 }
1560
1561 //---------------------------match_tree----------------------------------------
1562 // Match a Ideal Node DAG - turn it into a tree; Label & Reduce. Used as part
1563 // of the whole-sale conversion from Ideal to Mach Nodes. Also used for
1564 // making GotoNodes while building the CFG and in init_spill_mask() to identify
1565 // a Load's result RegMask for memoization in idealreg2regmask[]
1566 MachNode *Matcher::match_tree( const Node *n ) {
1567 assert( n->Opcode() != Op_Phi, "cannot match" );
1568 assert( !n->is_block_start(), "cannot match" );
1569 // Set the mark for all locally allocated State objects.
1570 // When this call returns, the _states_arena arena will be reset
1571 // freeing all State objects.
1572 ResourceMark rm( &_states_arena );
1573
1574 LabelRootDepth = 0;
1575
1576 // StoreNodes require their Memory input to match any LoadNodes
1577 Node *mem = n->is_Store() ? n->in(MemNode::Memory) : (Node*)1 ;
1578 #ifdef ASSERT
1579 Node* save_mem_node = _mem_node;
1580 _mem_node = n->is_Store() ? (Node*)n : nullptr;
1581 #endif
1582 // State object for root node of match tree
1583 // Allocate it on _states_arena - stack allocation can cause stack overflow.
1584 State *s = new (&_states_arena) State;
1585 s->_kids[0] = nullptr;
1586 s->_kids[1] = nullptr;
1587 s->_leaf = (Node*)n;
1588 // Label the input tree, allocating labels from top-level arena
1589 Node* root_mem = mem;
1590 Label_Root(n, s, n->in(0), root_mem);
1591 if (C->failing()) return nullptr;
1592
1593 // The minimum cost match for the whole tree is found at the root State
1594 uint mincost = max_juint;
1595 uint cost = max_juint;
1596 uint i;
1597 for (i = 0; i < NUM_OPERANDS; i++) {
1598 if (s->valid(i) && // valid entry and
1599 s->cost(i) < cost && // low cost and
1600 s->rule(i) >= NUM_OPERANDS) {// not an operand
1601 mincost = i;
1602 cost = s->cost(i);
1603 }
1604 }
1605 if (mincost == max_juint) {
1606 #ifndef PRODUCT
1607 tty->print("No matching rule for:");
1608 s->dump();
1609 #endif
1610 Matcher::soft_match_failure();
1611 return nullptr;
1612 }
1613 // Reduce input tree based upon the state labels to machine Nodes
1614 MachNode *m = ReduceInst(s, s->rule(mincost), mem);
1615 // New-to-old mapping is done in ReduceInst, to cover complex instructions.
1616 NOT_PRODUCT(_old2new_map.map(n->_idx, m);)
1617
1618 // Add any Matcher-ignored edges
1619 uint cnt = n->req();
1620 uint start = 1;
1621 if( mem != (Node*)1 ) start = MemNode::Memory+1;
1622 if( n->is_AddP() ) {
1623 assert( mem == (Node*)1, "" );
1624 start = AddPNode::Base+1;
1625 }
1626 for( i = start; i < cnt; i++ ) {
1627 if( !n->match_edge(i) ) {
1628 if( i < m->req() )
1629 m->ins_req( i, n->in(i) );
1630 else
1631 m->add_req( n->in(i) );
1632 }
1633 }
1634
1635 debug_only( _mem_node = save_mem_node; )
1636 return m;
1637 }
1638
1639
1640 //------------------------------match_into_reg---------------------------------
1641 // Choose to either match this Node in a register or part of the current
1642 // match tree. Return true for requiring a register and false for matching
1643 // as part of the current match tree.
1644 static bool match_into_reg( const Node *n, Node *m, Node *control, int i, bool shared ) {
1645
1646 const Type *t = m->bottom_type();
1647
1648 if (t->singleton()) {
1649 // Never force constants into registers. Allow them to match as
1650 // constants or registers. Copies of the same value will share
1651 // the same register. See find_shared_node.
1652 return false;
1653 } else { // Not a constant
1654 if (!shared && Matcher::is_encode_and_store_pattern(n, m)) {
1655 // Make it possible to match "encode and store" patterns with non-shared
1656 // encode operations that are pinned to a control node (e.g. by CastPP
1657 // node removal in final graph reshaping). The matched instruction cannot
1658 // float above the encode's control node because it is pinned to the
1659 // store's control node.
1660 return false;
1661 }
1662 // Stop recursion if they have different Controls.
1663 Node* m_control = m->in(0);
1664 // Control of load's memory can post-dominates load's control.
1665 // So use it since load can't float above its memory.
1666 Node* mem_control = (m->is_Load()) ? m->in(MemNode::Memory)->in(0) : nullptr;
1667 if (control && m_control && control != m_control && control != mem_control) {
1668
1669 // Actually, we can live with the most conservative control we
1670 // find, if it post-dominates the others. This allows us to
1671 // pick up load/op/store trees where the load can float a little
1672 // above the store.
1673 Node *x = control;
1674 const uint max_scan = 6; // Arbitrary scan cutoff
1675 uint j;
1676 for (j=0; j<max_scan; j++) {
1677 if (x->is_Region()) // Bail out at merge points
1678 return true;
1679 x = x->in(0);
1680 if (x == m_control) // Does 'control' post-dominate
1681 break; // m->in(0)? If so, we can use it
1682 if (x == mem_control) // Does 'control' post-dominate
1683 break; // mem_control? If so, we can use it
1684 }
1685 if (j == max_scan) // No post-domination before scan end?
1686 return true; // Then break the match tree up
1687 }
1688 if ((m->is_DecodeN() && Matcher::narrow_oop_use_complex_address()) ||
1689 (m->is_DecodeNKlass() && Matcher::narrow_klass_use_complex_address())) {
1690 // These are commonly used in address expressions and can
1691 // efficiently fold into them on X64 in some cases.
1692 return false;
1693 }
1694 }
1695
1696 // Not forceable cloning. If shared, put it into a register.
1697 return shared;
1698 }
1699
1700
1701 //------------------------------Instruction Selection--------------------------
1702 // Label method walks a "tree" of nodes, using the ADLC generated DFA to match
1703 // ideal nodes to machine instructions. Trees are delimited by shared Nodes,
1704 // things the Matcher does not match (e.g., Memory), and things with different
1705 // Controls (hence forced into different blocks). We pass in the Control
1706 // selected for this entire State tree.
1707
1708 // The Matcher works on Trees, but an Intel add-to-memory requires a DAG: the
1709 // Store and the Load must have identical Memories (as well as identical
1710 // pointers). Since the Matcher does not have anything for Memory (and
1711 // does not handle DAGs), I have to match the Memory input myself. If the
1712 // Tree root is a Store or if there are multiple Loads in the tree, I require
1713 // all Loads to have the identical memory.
1714 Node* Matcher::Label_Root(const Node* n, State* svec, Node* control, Node*& mem) {
1715 // Since Label_Root is a recursive function, its possible that we might run
1716 // out of stack space. See bugs 6272980 & 6227033 for more info.
1717 LabelRootDepth++;
1718 if (LabelRootDepth > MaxLabelRootDepth) {
1719 // Bailout. Can for example be hit with a deep chain of operations.
1720 C->record_method_not_compilable("Out of stack space, increase MaxLabelRootDepth");
1721 return nullptr;
1722 }
1723 uint care = 0; // Edges matcher cares about
1724 uint cnt = n->req();
1725 uint i = 0;
1726
1727 // Examine children for memory state
1728 // Can only subsume a child into your match-tree if that child's memory state
1729 // is not modified along the path to another input.
1730 // It is unsafe even if the other inputs are separate roots.
1731 Node *input_mem = nullptr;
1732 for( i = 1; i < cnt; i++ ) {
1733 if( !n->match_edge(i) ) continue;
1734 Node *m = n->in(i); // Get ith input
1735 assert( m, "expect non-null children" );
1736 if( m->is_Load() ) {
1737 if( input_mem == nullptr ) {
1738 input_mem = m->in(MemNode::Memory);
1739 if (mem == (Node*)1) {
1740 // Save this memory to bail out if there's another memory access
1741 // to a different memory location in the same tree.
1742 mem = input_mem;
1743 }
1744 } else if( input_mem != m->in(MemNode::Memory) ) {
1745 input_mem = NodeSentinel;
1746 }
1747 }
1748 }
1749
1750 for( i = 1; i < cnt; i++ ){// For my children
1751 if( !n->match_edge(i) ) continue;
1752 Node *m = n->in(i); // Get ith input
1753 // Allocate states out of a private arena
1754 State *s = new (&_states_arena) State;
1755 svec->_kids[care++] = s;
1756 assert( care <= 2, "binary only for now" );
1757
1758 // Recursively label the State tree.
1759 s->_kids[0] = nullptr;
1760 s->_kids[1] = nullptr;
1761 s->_leaf = m;
1762
1763 // Check for leaves of the State Tree; things that cannot be a part of
1764 // the current tree. If it finds any, that value is matched as a
1765 // register operand. If not, then the normal matching is used.
1766 if( match_into_reg(n, m, control, i, is_shared(m)) ||
1767 // Stop recursion if this is a LoadNode and there is another memory access
1768 // to a different memory location in the same tree (for example, a StoreNode
1769 // at the root of this tree or another LoadNode in one of the children).
1770 ((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) ||
1771 // Can NOT include the match of a subtree when its memory state
1772 // is used by any of the other subtrees
1773 (input_mem == NodeSentinel) ) {
1774 // Print when we exclude matching due to different memory states at input-loads
1775 if (PrintOpto && (Verbose && WizardMode) && (input_mem == NodeSentinel)
1776 && !((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem)) {
1777 tty->print_cr("invalid input_mem");
1778 }
1779 // Switch to a register-only opcode; this value must be in a register
1780 // and cannot be subsumed as part of a larger instruction.
1781 s->DFA( m->ideal_reg(), m );
1782
1783 } else {
1784 // If match tree has no control and we do, adopt it for entire tree
1785 if( control == nullptr && m->in(0) != nullptr && m->req() > 1 )
1786 control = m->in(0); // Pick up control
1787 // Else match as a normal part of the match tree.
1788 control = Label_Root(m, s, control, mem);
1789 if (C->failing()) return nullptr;
1790 }
1791 }
1792
1793 // Call DFA to match this node, and return
1794 svec->DFA( n->Opcode(), n );
1795
1796 uint x;
1797 for( x = 0; x < _LAST_MACH_OPER; x++ )
1798 if( svec->valid(x) )
1799 break;
1800
1801 if (x >= _LAST_MACH_OPER) {
1802 #ifdef ASSERT
1803 n->dump();
1804 svec->dump();
1805 #endif
1806 assert( false, "bad AD file" );
1807 C->record_failure("bad AD file");
1808 }
1809 return control;
1810 }
1811
1812
1813 // Con nodes reduced using the same rule can share their MachNode
1814 // which reduces the number of copies of a constant in the final
1815 // program. The register allocator is free to split uses later to
1816 // split live ranges.
1817 MachNode* Matcher::find_shared_node(Node* leaf, uint rule) {
1818 if (!leaf->is_Con() && !leaf->is_DecodeNarrowPtr()) return nullptr;
1819
1820 // See if this Con has already been reduced using this rule.
1821 if (_shared_nodes.max() <= leaf->_idx) return nullptr;
1822 MachNode* last = (MachNode*)_shared_nodes.at(leaf->_idx);
1823 if (last != nullptr && rule == last->rule()) {
1824 // Don't expect control change for DecodeN
1825 if (leaf->is_DecodeNarrowPtr())
1826 return last;
1827 // Get the new space root.
1828 Node* xroot = new_node(C->root());
1829 if (xroot == nullptr) {
1830 // This shouldn't happen give the order of matching.
1831 return nullptr;
1832 }
1833
1834 // Shared constants need to have their control be root so they
1835 // can be scheduled properly.
1836 Node* control = last->in(0);
1837 if (control != xroot) {
1838 if (control == nullptr || control == C->root()) {
1839 last->set_req(0, xroot);
1840 } else {
1841 assert(false, "unexpected control");
1842 return nullptr;
1843 }
1844 }
1845 return last;
1846 }
1847 return nullptr;
1848 }
1849
1850
1851 //------------------------------ReduceInst-------------------------------------
1852 // Reduce a State tree (with given Control) into a tree of MachNodes.
1853 // This routine (and it's cohort ReduceOper) convert Ideal Nodes into
1854 // complicated machine Nodes. Each MachNode covers some tree of Ideal Nodes.
1855 // Each MachNode has a number of complicated MachOper operands; each
1856 // MachOper also covers a further tree of Ideal Nodes.
1857
1858 // The root of the Ideal match tree is always an instruction, so we enter
1859 // the recursion here. After building the MachNode, we need to recurse
1860 // the tree checking for these cases:
1861 // (1) Child is an instruction -
1862 // Build the instruction (recursively), add it as an edge.
1863 // Build a simple operand (register) to hold the result of the instruction.
1864 // (2) Child is an interior part of an instruction -
1865 // Skip over it (do nothing)
1866 // (3) Child is the start of a operand -
1867 // Build the operand, place it inside the instruction
1868 // Call ReduceOper.
1869 MachNode *Matcher::ReduceInst( State *s, int rule, Node *&mem ) {
1870 assert( rule >= NUM_OPERANDS, "called with operand rule" );
1871
1872 MachNode* shared_node = find_shared_node(s->_leaf, rule);
1873 if (shared_node != nullptr) {
1874 return shared_node;
1875 }
1876
1877 // Build the object to represent this state & prepare for recursive calls
1878 MachNode *mach = s->MachNodeGenerator(rule);
1879 guarantee(mach != nullptr, "Missing MachNode");
1880 mach->_opnds[0] = s->MachOperGenerator(_reduceOp[rule]);
1881 assert( mach->_opnds[0] != nullptr, "Missing result operand" );
1882 Node *leaf = s->_leaf;
1883 NOT_PRODUCT(record_new2old(mach, leaf);)
1884 // Check for instruction or instruction chain rule
1885 if( rule >= _END_INST_CHAIN_RULE || rule < _BEGIN_INST_CHAIN_RULE ) {
1886 assert(C->node_arena()->contains(s->_leaf) || !has_new_node(s->_leaf),
1887 "duplicating node that's already been matched");
1888 // Instruction
1889 mach->add_req( leaf->in(0) ); // Set initial control
1890 // Reduce interior of complex instruction
1891 ReduceInst_Interior( s, rule, mem, mach, 1 );
1892 } else {
1893 // Instruction chain rules are data-dependent on their inputs
1894 mach->add_req(nullptr); // Set initial control to none
1895 ReduceInst_Chain_Rule( s, rule, mem, mach );
1896 }
1897
1898 // If a Memory was used, insert a Memory edge
1899 if( mem != (Node*)1 ) {
1900 mach->ins_req(MemNode::Memory,mem);
1901 #ifdef ASSERT
1902 // Verify adr type after matching memory operation
1903 const MachOper* oper = mach->memory_operand();
1904 if (oper != nullptr && oper != (MachOper*)-1) {
1905 // It has a unique memory operand. Find corresponding ideal mem node.
1906 Node* m = nullptr;
1907 if (leaf->is_Mem()) {
1908 m = leaf;
1909 } else {
1910 m = _mem_node;
1911 assert(m != nullptr && m->is_Mem(), "expecting memory node");
1912 }
1913 const Type* mach_at = mach->adr_type();
1914 // DecodeN node consumed by an address may have different type
1915 // than its input. Don't compare types for such case.
1916 if (m->adr_type() != mach_at &&
1917 (m->in(MemNode::Address)->is_DecodeNarrowPtr() ||
1918 (m->in(MemNode::Address)->is_AddP() &&
1919 m->in(MemNode::Address)->in(AddPNode::Address)->is_DecodeNarrowPtr()) ||
1920 (m->in(MemNode::Address)->is_AddP() &&
1921 m->in(MemNode::Address)->in(AddPNode::Address)->is_AddP() &&
1922 m->in(MemNode::Address)->in(AddPNode::Address)->in(AddPNode::Address)->is_DecodeNarrowPtr()))) {
1923 mach_at = m->adr_type();
1924 }
1925 if (m->adr_type() != mach_at) {
1926 m->dump();
1927 tty->print_cr("mach:");
1928 mach->dump(1);
1929 }
1930 assert(m->adr_type() == mach_at, "matcher should not change adr type");
1931 }
1932 #endif
1933 }
1934
1935 // If the _leaf is an AddP, insert the base edge
1936 if (leaf->is_AddP()) {
1937 mach->ins_req(AddPNode::Base,leaf->in(AddPNode::Base));
1938 }
1939
1940 uint number_of_projections_prior = number_of_projections();
1941
1942 // Perform any 1-to-many expansions required
1943 MachNode *ex = mach->Expand(s, _projection_list, mem);
1944 if (ex != mach) {
1945 assert(ex->ideal_reg() == mach->ideal_reg(), "ideal types should match");
1946 if( ex->in(1)->is_Con() )
1947 ex->in(1)->set_req(0, C->root());
1948 // Remove old node from the graph
1949 for( uint i=0; i<mach->req(); i++ ) {
1950 mach->set_req(i,nullptr);
1951 }
1952 NOT_PRODUCT(record_new2old(ex, s->_leaf);)
1953 }
1954
1955 // PhaseChaitin::fixup_spills will sometimes generate spill code
1956 // via the matcher. By the time, nodes have been wired into the CFG,
1957 // and any further nodes generated by expand rules will be left hanging
1958 // in space, and will not get emitted as output code. Catch this.
1959 // Also, catch any new register allocation constraints ("projections")
1960 // generated belatedly during spill code generation.
1961 if (_allocation_started) {
1962 guarantee(ex == mach, "no expand rules during spill generation");
1963 guarantee(number_of_projections_prior == number_of_projections(), "no allocation during spill generation");
1964 }
1965
1966 if (leaf->is_Con() || leaf->is_DecodeNarrowPtr()) {
1967 // Record the con for sharing
1968 _shared_nodes.map(leaf->_idx, ex);
1969 }
1970
1971 // Have mach nodes inherit GC barrier data
1972 mach->set_barrier_data(MemNode::barrier_data(leaf));
1973
1974 return ex;
1975 }
1976
1977 void Matcher::handle_precedence_edges(Node* n, MachNode *mach) {
1978 for (uint i = n->req(); i < n->len(); i++) {
1979 if (n->in(i) != nullptr) {
1980 mach->add_prec(n->in(i));
1981 }
1982 }
1983 }
1984
1985 void Matcher::ReduceInst_Chain_Rule(State* s, int rule, Node* &mem, MachNode* mach) {
1986 // 'op' is what I am expecting to receive
1987 int op = _leftOp[rule];
1988 // Operand type to catch childs result
1989 // This is what my child will give me.
1990 unsigned int opnd_class_instance = s->rule(op);
1991 // Choose between operand class or not.
1992 // This is what I will receive.
1993 int catch_op = (FIRST_OPERAND_CLASS <= op && op < NUM_OPERANDS) ? opnd_class_instance : op;
1994 // New rule for child. Chase operand classes to get the actual rule.
1995 unsigned int newrule = s->rule(catch_op);
1996
1997 if (newrule < NUM_OPERANDS) {
1998 // Chain from operand or operand class, may be output of shared node
1999 assert(opnd_class_instance < NUM_OPERANDS, "Bad AD file: Instruction chain rule must chain from operand");
2000 // Insert operand into array of operands for this instruction
2001 mach->_opnds[1] = s->MachOperGenerator(opnd_class_instance);
2002
2003 ReduceOper(s, newrule, mem, mach);
2004 } else {
2005 // Chain from the result of an instruction
2006 assert(newrule >= _LAST_MACH_OPER, "Do NOT chain from internal operand");
2007 mach->_opnds[1] = s->MachOperGenerator(_reduceOp[catch_op]);
2008 Node *mem1 = (Node*)1;
2009 debug_only(Node *save_mem_node = _mem_node;)
2010 mach->add_req( ReduceInst(s, newrule, mem1) );
2011 debug_only(_mem_node = save_mem_node;)
2012 }
2013 return;
2014 }
2015
2016
2017 uint Matcher::ReduceInst_Interior( State *s, int rule, Node *&mem, MachNode *mach, uint num_opnds ) {
2018 handle_precedence_edges(s->_leaf, mach);
2019
2020 if( s->_leaf->is_Load() ) {
2021 Node *mem2 = s->_leaf->in(MemNode::Memory);
2022 assert( mem == (Node*)1 || mem == mem2, "multiple Memories being matched at once?" );
2023 debug_only( if( mem == (Node*)1 ) _mem_node = s->_leaf;)
2024 mem = mem2;
2025 }
2026 if( s->_leaf->in(0) != nullptr && s->_leaf->req() > 1) {
2027 if( mach->in(0) == nullptr )
2028 mach->set_req(0, s->_leaf->in(0));
2029 }
2030
2031 // Now recursively walk the state tree & add operand list.
2032 for( uint i=0; i<2; i++ ) { // binary tree
2033 State *newstate = s->_kids[i];
2034 if( newstate == nullptr ) break; // Might only have 1 child
2035 // 'op' is what I am expecting to receive
2036 int op;
2037 if( i == 0 ) {
2038 op = _leftOp[rule];
2039 } else {
2040 op = _rightOp[rule];
2041 }
2042 // Operand type to catch childs result
2043 // This is what my child will give me.
2044 int opnd_class_instance = newstate->rule(op);
2045 // Choose between operand class or not.
2046 // This is what I will receive.
2047 int catch_op = (op >= FIRST_OPERAND_CLASS && op < NUM_OPERANDS) ? opnd_class_instance : op;
2048 // New rule for child. Chase operand classes to get the actual rule.
2049 int newrule = newstate->rule(catch_op);
2050
2051 if (newrule < NUM_OPERANDS) { // Operand/operandClass or internalOp/instruction?
2052 // Operand/operandClass
2053 // Insert operand into array of operands for this instruction
2054 mach->_opnds[num_opnds++] = newstate->MachOperGenerator(opnd_class_instance);
2055 ReduceOper(newstate, newrule, mem, mach);
2056
2057 } else { // Child is internal operand or new instruction
2058 if (newrule < _LAST_MACH_OPER) { // internal operand or instruction?
2059 // internal operand --> call ReduceInst_Interior
2060 // Interior of complex instruction. Do nothing but recurse.
2061 num_opnds = ReduceInst_Interior(newstate, newrule, mem, mach, num_opnds);
2062 } else {
2063 // instruction --> call build operand( ) to catch result
2064 // --> ReduceInst( newrule )
2065 mach->_opnds[num_opnds++] = s->MachOperGenerator(_reduceOp[catch_op]);
2066 Node *mem1 = (Node*)1;
2067 debug_only(Node *save_mem_node = _mem_node;)
2068 mach->add_req( ReduceInst( newstate, newrule, mem1 ) );
2069 debug_only(_mem_node = save_mem_node;)
2070 }
2071 }
2072 assert( mach->_opnds[num_opnds-1], "" );
2073 }
2074 return num_opnds;
2075 }
2076
2077 // This routine walks the interior of possible complex operands.
2078 // At each point we check our children in the match tree:
2079 // (1) No children -
2080 // We are a leaf; add _leaf field as an input to the MachNode
2081 // (2) Child is an internal operand -
2082 // Skip over it ( do nothing )
2083 // (3) Child is an instruction -
2084 // Call ReduceInst recursively and
2085 // and instruction as an input to the MachNode
2086 void Matcher::ReduceOper( State *s, int rule, Node *&mem, MachNode *mach ) {
2087 assert( rule < _LAST_MACH_OPER, "called with operand rule" );
2088 State *kid = s->_kids[0];
2089 assert( kid == nullptr || s->_leaf->in(0) == nullptr, "internal operands have no control" );
2090
2091 // Leaf? And not subsumed?
2092 if( kid == nullptr && !_swallowed[rule] ) {
2093 mach->add_req( s->_leaf ); // Add leaf pointer
2094 return; // Bail out
2095 }
2096
2097 if( s->_leaf->is_Load() ) {
2098 assert( mem == (Node*)1, "multiple Memories being matched at once?" );
2099 mem = s->_leaf->in(MemNode::Memory);
2100 debug_only(_mem_node = s->_leaf;)
2101 }
2102
2103 handle_precedence_edges(s->_leaf, mach);
2104
2105 if( s->_leaf->in(0) && s->_leaf->req() > 1) {
2106 if( !mach->in(0) )
2107 mach->set_req(0,s->_leaf->in(0));
2108 else {
2109 assert( s->_leaf->in(0) == mach->in(0), "same instruction, differing controls?" );
2110 }
2111 }
2112
2113 for (uint i = 0; kid != nullptr && i < 2; kid = s->_kids[1], i++) { // binary tree
2114 int newrule;
2115 if( i == 0) {
2116 newrule = kid->rule(_leftOp[rule]);
2117 } else {
2118 newrule = kid->rule(_rightOp[rule]);
2119 }
2120
2121 if (newrule < _LAST_MACH_OPER) { // Operand or instruction?
2122 // Internal operand; recurse but do nothing else
2123 ReduceOper(kid, newrule, mem, mach);
2124
2125 } else { // Child is a new instruction
2126 // Reduce the instruction, and add a direct pointer from this
2127 // machine instruction to the newly reduced one.
2128 Node *mem1 = (Node*)1;
2129 debug_only(Node *save_mem_node = _mem_node;)
2130 mach->add_req( ReduceInst( kid, newrule, mem1 ) );
2131 debug_only(_mem_node = save_mem_node;)
2132 }
2133 }
2134 }
2135
2136
2137 // -------------------------------------------------------------------------
2138 // Java-Java calling convention
2139 // (what you use when Java calls Java)
2140
2141 //------------------------------find_receiver----------------------------------
2142 // For a given signature, return the OptoReg for parameter 0.
2143 OptoReg::Name Matcher::find_receiver() {
2144 VMRegPair regs;
2145 BasicType sig_bt = T_OBJECT;
2146 SharedRuntime::java_calling_convention(&sig_bt, ®s, 1);
2147 // Return argument 0 register. In the LP64 build pointers
2148 // take 2 registers, but the VM wants only the 'main' name.
2149 return OptoReg::as_OptoReg(regs.first());
2150 }
2151
2152 bool Matcher::is_vshift_con_pattern(Node* n, Node* m) {
2153 if (n != nullptr && m != nullptr) {
2154 return VectorNode::is_vector_shift(n) &&
2155 VectorNode::is_vector_shift_count(m) && m->in(1)->is_Con();
2156 }
2157 return false;
2158 }
2159
2160 bool Matcher::clone_node(Node* n, Node* m, Matcher::MStack& mstack) {
2161 // Must clone all producers of flags, or we will not match correctly.
2162 // Suppose a compare setting int-flags is shared (e.g., a switch-tree)
2163 // then it will match into an ideal Op_RegFlags. Alas, the fp-flags
2164 // are also there, so we may match a float-branch to int-flags and
2165 // expect the allocator to haul the flags from the int-side to the
2166 // fp-side. No can do.
2167 if (_must_clone[m->Opcode()]) {
2168 mstack.push(m, Visit);
2169 return true;
2170 }
2171 return pd_clone_node(n, m, mstack);
2172 }
2173
2174 bool Matcher::clone_base_plus_offset_address(AddPNode* m, Matcher::MStack& mstack, VectorSet& address_visited) {
2175 Node *off = m->in(AddPNode::Offset);
2176 if (off->is_Con()) {
2177 address_visited.test_set(m->_idx); // Flag as address_visited
2178 mstack.push(m->in(AddPNode::Address), Pre_Visit);
2179 // Clone X+offset as it also folds into most addressing expressions
2180 mstack.push(off, Visit);
2181 mstack.push(m->in(AddPNode::Base), Pre_Visit);
2182 return true;
2183 }
2184 return false;
2185 }
2186
2187 // A method-klass-holder may be passed in the inline_cache_reg
2188 // and then expanded into the inline_cache_reg and a method_ptr register
2189 // defined in ad_<arch>.cpp
2190
2191 //------------------------------find_shared------------------------------------
2192 // Set bits if Node is shared or otherwise a root
2193 void Matcher::find_shared(Node* n) {
2194 // Allocate stack of size C->live_nodes() * 2 to avoid frequent realloc
2195 MStack mstack(C->live_nodes() * 2);
2196 // Mark nodes as address_visited if they are inputs to an address expression
2197 VectorSet address_visited;
2198 mstack.push(n, Visit); // Don't need to pre-visit root node
2199 while (mstack.is_nonempty()) {
2200 n = mstack.node(); // Leave node on stack
2201 Node_State nstate = mstack.state();
2202 uint nop = n->Opcode();
2203 if (nstate == Pre_Visit) {
2204 if (address_visited.test(n->_idx)) { // Visited in address already?
2205 // Flag as visited and shared now.
2206 set_visited(n);
2207 }
2208 if (is_visited(n)) { // Visited already?
2209 // Node is shared and has no reason to clone. Flag it as shared.
2210 // This causes it to match into a register for the sharing.
2211 set_shared(n); // Flag as shared and
2212 if (n->is_DecodeNarrowPtr()) {
2213 // Oop field/array element loads must be shared but since
2214 // they are shared through a DecodeN they may appear to have
2215 // a single use so force sharing here.
2216 set_shared(n->in(1));
2217 }
2218 mstack.pop(); // remove node from stack
2219 continue;
2220 }
2221 nstate = Visit; // Not already visited; so visit now
2222 }
2223 if (nstate == Visit) {
2224 mstack.set_state(Post_Visit);
2225 set_visited(n); // Flag as visited now
2226 bool mem_op = false;
2227 int mem_addr_idx = MemNode::Address;
2228 if (find_shared_visit(mstack, n, nop, mem_op, mem_addr_idx)) {
2229 continue;
2230 }
2231 for (int i = n->len() - 1; i >= 0; --i) { // For my children
2232 Node* m = n->in(i); // Get ith input
2233 if (m == nullptr) {
2234 continue; // Ignore nulls
2235 }
2236 if (clone_node(n, m, mstack)) {
2237 continue;
2238 }
2239
2240 // Clone addressing expressions as they are "free" in memory access instructions
2241 if (mem_op && i == mem_addr_idx && m->is_AddP() &&
2242 // When there are other uses besides address expressions
2243 // put it on stack and mark as shared.
2244 !is_visited(m)) {
2245 // Some inputs for address expression are not put on stack
2246 // to avoid marking them as shared and forcing them into register
2247 // if they are used only in address expressions.
2248 // But they should be marked as shared if there are other uses
2249 // besides address expressions.
2250
2251 if (pd_clone_address_expressions(m->as_AddP(), mstack, address_visited)) {
2252 continue;
2253 }
2254 } // if( mem_op &&
2255 mstack.push(m, Pre_Visit);
2256 } // for(int i = ...)
2257 }
2258 else if (nstate == Alt_Post_Visit) {
2259 mstack.pop(); // Remove node from stack
2260 // We cannot remove the Cmp input from the Bool here, as the Bool may be
2261 // shared and all users of the Bool need to move the Cmp in parallel.
2262 // This leaves both the Bool and the If pointing at the Cmp. To
2263 // prevent the Matcher from trying to Match the Cmp along both paths
2264 // BoolNode::match_edge always returns a zero.
2265
2266 // We reorder the Op_If in a pre-order manner, so we can visit without
2267 // accidentally sharing the Cmp (the Bool and the If make 2 users).
2268 n->add_req( n->in(1)->in(1) ); // Add the Cmp next to the Bool
2269 }
2270 else if (nstate == Post_Visit) {
2271 mstack.pop(); // Remove node from stack
2272
2273 // Now hack a few special opcodes
2274 uint opcode = n->Opcode();
2275 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->matcher_find_shared_post_visit(this, n, opcode);
2276 if (!gc_handled) {
2277 find_shared_post_visit(n, opcode);
2278 }
2279 }
2280 else {
2281 ShouldNotReachHere();
2282 }
2283 } // end of while (mstack.is_nonempty())
2284 }
2285
2286 bool Matcher::find_shared_visit(MStack& mstack, Node* n, uint opcode, bool& mem_op, int& mem_addr_idx) {
2287 switch(opcode) { // Handle some opcodes special
2288 case Op_Phi: // Treat Phis as shared roots
2289 case Op_Parm:
2290 case Op_Proj: // All handled specially during matching
2291 case Op_SafePointScalarObject:
2292 set_shared(n);
2293 set_dontcare(n);
2294 break;
2295 case Op_If:
2296 case Op_CountedLoopEnd:
2297 mstack.set_state(Alt_Post_Visit); // Alternative way
2298 // Convert (If (Bool (CmpX A B))) into (If (Bool) (CmpX A B)). Helps
2299 // with matching cmp/branch in 1 instruction. The Matcher needs the
2300 // Bool and CmpX side-by-side, because it can only get at constants
2301 // that are at the leaves of Match trees, and the Bool's condition acts
2302 // as a constant here.
2303 mstack.push(n->in(1), Visit); // Clone the Bool
2304 mstack.push(n->in(0), Pre_Visit); // Visit control input
2305 return true; // while (mstack.is_nonempty())
2306 case Op_ConvI2D: // These forms efficiently match with a prior
2307 case Op_ConvI2F: // Load but not a following Store
2308 if( n->in(1)->is_Load() && // Prior load
2309 n->outcnt() == 1 && // Not already shared
2310 n->unique_out()->is_Store() ) // Following store
2311 set_shared(n); // Force it to be a root
2312 break;
2313 case Op_ReverseBytesI:
2314 case Op_ReverseBytesL:
2315 if( n->in(1)->is_Load() && // Prior load
2316 n->outcnt() == 1 ) // Not already shared
2317 set_shared(n); // Force it to be a root
2318 break;
2319 case Op_BoxLock: // Can't match until we get stack-regs in ADLC
2320 case Op_IfFalse:
2321 case Op_IfTrue:
2322 case Op_MachProj:
2323 case Op_MergeMem:
2324 case Op_Catch:
2325 case Op_CatchProj:
2326 case Op_CProj:
2327 case Op_JumpProj:
2328 case Op_JProj:
2329 case Op_NeverBranch:
2330 set_dontcare(n);
2331 break;
2332 case Op_Jump:
2333 mstack.push(n->in(1), Pre_Visit); // Switch Value (could be shared)
2334 mstack.push(n->in(0), Pre_Visit); // Visit Control input
2335 return true; // while (mstack.is_nonempty())
2336 case Op_StrComp:
2337 case Op_StrEquals:
2338 case Op_StrIndexOf:
2339 case Op_StrIndexOfChar:
2340 case Op_AryEq:
2341 case Op_VectorizedHashCode:
2342 case Op_CountPositives:
2343 case Op_StrInflatedCopy:
2344 case Op_StrCompressedCopy:
2345 case Op_EncodeISOArray:
2346 case Op_FmaD:
2347 case Op_FmaF:
2348 case Op_FmaHF:
2349 case Op_FmaVD:
2350 case Op_FmaVF:
2351 case Op_MacroLogicV:
2352 case Op_VectorCmpMasked:
2353 case Op_CompressV:
2354 case Op_CompressM:
2355 case Op_ExpandV:
2356 case Op_VectorLoadMask:
2357 set_shared(n); // Force result into register (it will be anyways)
2358 break;
2359 case Op_ConP: { // Convert pointers above the centerline to NUL
2360 TypeNode *tn = n->as_Type(); // Constants derive from type nodes
2361 const TypePtr* tp = tn->type()->is_ptr();
2362 if (tp->_ptr == TypePtr::AnyNull) {
2363 tn->set_type(TypePtr::NULL_PTR);
2364 }
2365 break;
2366 }
2367 case Op_ConN: { // Convert narrow pointers above the centerline to NUL
2368 TypeNode *tn = n->as_Type(); // Constants derive from type nodes
2369 const TypePtr* tp = tn->type()->make_ptr();
2370 if (tp && tp->_ptr == TypePtr::AnyNull) {
2371 tn->set_type(TypeNarrowOop::NULL_PTR);
2372 }
2373 break;
2374 }
2375 case Op_Binary: // These are introduced in the Post_Visit state.
2376 ShouldNotReachHere();
2377 break;
2378 case Op_ClearArray:
2379 case Op_SafePoint:
2380 mem_op = true;
2381 break;
2382 default:
2383 if( n->is_Store() ) {
2384 // Do match stores, despite no ideal reg
2385 mem_op = true;
2386 break;
2387 }
2388 if( n->is_Mem() ) { // Loads and LoadStores
2389 mem_op = true;
2390 // Loads must be root of match tree due to prior load conflict
2391 if( C->subsume_loads() == false )
2392 set_shared(n);
2393 }
2394 // Fall into default case
2395 if( !n->ideal_reg() )
2396 set_dontcare(n); // Unmatchable Nodes
2397 } // end_switch
2398 return false;
2399 }
2400
2401 void Matcher::find_shared_post_visit(Node* n, uint opcode) {
2402 if (n->is_predicated_vector()) {
2403 // Restructure into binary trees for Matching.
2404 if (n->req() == 4) {
2405 n->set_req(1, new BinaryNode(n->in(1), n->in(2)));
2406 n->set_req(2, n->in(3));
2407 n->del_req(3);
2408 } else if (n->req() == 5) {
2409 n->set_req(1, new BinaryNode(n->in(1), n->in(2)));
2410 n->set_req(2, new BinaryNode(n->in(3), n->in(4)));
2411 n->del_req(4);
2412 n->del_req(3);
2413 } else if (n->req() == 6) {
2414 Node* b3 = new BinaryNode(n->in(4), n->in(5));
2415 Node* b2 = new BinaryNode(n->in(3), b3);
2416 Node* b1 = new BinaryNode(n->in(2), b2);
2417 n->set_req(2, b1);
2418 n->del_req(5);
2419 n->del_req(4);
2420 n->del_req(3);
2421 }
2422 return;
2423 }
2424
2425 switch(opcode) { // Handle some opcodes special
2426 case Op_CompareAndExchangeB:
2427 case Op_CompareAndExchangeS:
2428 case Op_CompareAndExchangeI:
2429 case Op_CompareAndExchangeL:
2430 case Op_CompareAndExchangeP:
2431 case Op_CompareAndExchangeN:
2432 case Op_WeakCompareAndSwapB:
2433 case Op_WeakCompareAndSwapS:
2434 case Op_WeakCompareAndSwapI:
2435 case Op_WeakCompareAndSwapL:
2436 case Op_WeakCompareAndSwapP:
2437 case Op_WeakCompareAndSwapN:
2438 case Op_CompareAndSwapB:
2439 case Op_CompareAndSwapS:
2440 case Op_CompareAndSwapI:
2441 case Op_CompareAndSwapL:
2442 case Op_CompareAndSwapP:
2443 case Op_CompareAndSwapN: { // Convert trinary to binary-tree
2444 Node* newval = n->in(MemNode::ValueIn);
2445 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
2446 Node* pair = new BinaryNode(oldval, newval);
2447 n->set_req(MemNode::ValueIn, pair);
2448 n->del_req(LoadStoreConditionalNode::ExpectedIn);
2449 break;
2450 }
2451 case Op_CMoveD: // Convert trinary to binary-tree
2452 case Op_CMoveF:
2453 case Op_CMoveI:
2454 case Op_CMoveL:
2455 case Op_CMoveN:
2456 case Op_CMoveP: {
2457 // Restructure into a binary tree for Matching. It's possible that
2458 // we could move this code up next to the graph reshaping for IfNodes
2459 // or vice-versa, but I do not want to debug this for Ladybird.
2460 // 10/2/2000 CNC.
2461 Node* pair1 = new BinaryNode(n->in(1), n->in(1)->in(1));
2462 n->set_req(1, pair1);
2463 Node* pair2 = new BinaryNode(n->in(2), n->in(3));
2464 n->set_req(2, pair2);
2465 n->del_req(3);
2466 break;
2467 }
2468 case Op_MacroLogicV: {
2469 Node* pair1 = new BinaryNode(n->in(1), n->in(2));
2470 Node* pair2 = new BinaryNode(n->in(3), n->in(4));
2471 n->set_req(1, pair1);
2472 n->set_req(2, pair2);
2473 n->del_req(4);
2474 n->del_req(3);
2475 break;
2476 }
2477 case Op_StoreVectorMasked: {
2478 Node* pair = new BinaryNode(n->in(3), n->in(4));
2479 n->set_req(3, pair);
2480 n->del_req(4);
2481 break;
2482 }
2483 case Op_SelectFromTwoVector:
2484 case Op_LoopLimit: {
2485 Node* pair1 = new BinaryNode(n->in(1), n->in(2));
2486 n->set_req(1, pair1);
2487 n->set_req(2, n->in(3));
2488 n->del_req(3);
2489 break;
2490 }
2491 case Op_StrEquals:
2492 case Op_StrIndexOfChar: {
2493 Node* pair1 = new BinaryNode(n->in(2), n->in(3));
2494 n->set_req(2, pair1);
2495 n->set_req(3, n->in(4));
2496 n->del_req(4);
2497 break;
2498 }
2499 case Op_StrComp:
2500 case Op_StrIndexOf:
2501 case Op_VectorizedHashCode: {
2502 Node* pair1 = new BinaryNode(n->in(2), n->in(3));
2503 n->set_req(2, pair1);
2504 Node* pair2 = new BinaryNode(n->in(4),n->in(5));
2505 n->set_req(3, pair2);
2506 n->del_req(5);
2507 n->del_req(4);
2508 break;
2509 }
2510 case Op_EncodeISOArray:
2511 case Op_StrCompressedCopy:
2512 case Op_StrInflatedCopy: {
2513 // Restructure into a binary tree for Matching.
2514 Node* pair = new BinaryNode(n->in(3), n->in(4));
2515 n->set_req(3, pair);
2516 n->del_req(4);
2517 break;
2518 }
2519 case Op_FmaD:
2520 case Op_FmaF:
2521 case Op_FmaHF:
2522 case Op_FmaVD:
2523 case Op_FmaVF: {
2524 // Restructure into a binary tree for Matching.
2525 Node* pair = new BinaryNode(n->in(1), n->in(2));
2526 n->set_req(2, pair);
2527 n->set_req(1, n->in(3));
2528 n->del_req(3);
2529 break;
2530 }
2531 case Op_MulAddS2I: {
2532 Node* pair1 = new BinaryNode(n->in(1), n->in(2));
2533 Node* pair2 = new BinaryNode(n->in(3), n->in(4));
2534 n->set_req(1, pair1);
2535 n->set_req(2, pair2);
2536 n->del_req(4);
2537 n->del_req(3);
2538 break;
2539 }
2540 case Op_ClearArray: {
2541 Node* pair = new BinaryNode(n->in(2), n->in(3));
2542 n->set_req(2, pair);
2543 n->set_req(3, n->in(4));
2544 n->del_req(4);
2545 break;
2546 }
2547 case Op_VectorCmpMasked:
2548 case Op_CopySignD:
2549 case Op_SignumVF:
2550 case Op_SignumVD:
2551 case Op_SignumF:
2552 case Op_SignumD: {
2553 Node* pair = new BinaryNode(n->in(2), n->in(3));
2554 n->set_req(2, pair);
2555 n->del_req(3);
2556 break;
2557 }
2558 case Op_VectorBlend:
2559 case Op_VectorInsert: {
2560 Node* pair = new BinaryNode(n->in(1), n->in(2));
2561 n->set_req(1, pair);
2562 n->set_req(2, n->in(3));
2563 n->del_req(3);
2564 break;
2565 }
2566 case Op_LoadVectorGather:
2567 if (is_subword_type(n->bottom_type()->is_vect()->element_basic_type())) {
2568 Node* pair = new BinaryNode(n->in(MemNode::ValueIn), n->in(MemNode::ValueIn+1));
2569 n->set_req(MemNode::ValueIn, pair);
2570 n->del_req(MemNode::ValueIn+1);
2571 }
2572 break;
2573 case Op_LoadVectorGatherMasked:
2574 if (is_subword_type(n->bottom_type()->is_vect()->element_basic_type())) {
2575 Node* pair2 = new BinaryNode(n->in(MemNode::ValueIn + 1), n->in(MemNode::ValueIn + 2));
2576 Node* pair1 = new BinaryNode(n->in(MemNode::ValueIn), pair2);
2577 n->set_req(MemNode::ValueIn, pair1);
2578 n->del_req(MemNode::ValueIn+2);
2579 n->del_req(MemNode::ValueIn+1);
2580 break;
2581 } // fall-through
2582 case Op_StoreVectorScatter: {
2583 Node* pair = new BinaryNode(n->in(MemNode::ValueIn), n->in(MemNode::ValueIn+1));
2584 n->set_req(MemNode::ValueIn, pair);
2585 n->del_req(MemNode::ValueIn+1);
2586 break;
2587 }
2588 case Op_StoreVectorScatterMasked: {
2589 Node* pair = new BinaryNode(n->in(MemNode::ValueIn+1), n->in(MemNode::ValueIn+2));
2590 n->set_req(MemNode::ValueIn+1, pair);
2591 n->del_req(MemNode::ValueIn+2);
2592 pair = new BinaryNode(n->in(MemNode::ValueIn), n->in(MemNode::ValueIn+1));
2593 n->set_req(MemNode::ValueIn, pair);
2594 n->del_req(MemNode::ValueIn+1);
2595 break;
2596 }
2597 case Op_VectorMaskCmp: {
2598 n->set_req(1, new BinaryNode(n->in(1), n->in(2)));
2599 n->set_req(2, n->in(3));
2600 n->del_req(3);
2601 break;
2602 }
2603 case Op_PartialSubtypeCheck: {
2604 if (UseSecondarySupersTable && n->in(2)->is_Con()) {
2605 // PartialSubtypeCheck uses both constant and register operands for superclass input.
2606 n->set_req(2, new BinaryNode(n->in(2), n->in(2)));
2607 break;
2608 }
2609 break;
2610 }
2611 case Op_StoreLSpecial: {
2612 if (n->req() > (MemNode::ValueIn + 1) && n->in(MemNode::ValueIn + 1) != nullptr) {
2613 Node* pair = new BinaryNode(n->in(MemNode::ValueIn), n->in(MemNode::ValueIn + 1));
2614 n->set_req(MemNode::ValueIn, pair);
2615 n->del_req(MemNode::ValueIn + 1);
2616 }
2617 break;
2618 }
2619 default:
2620 break;
2621 }
2622 }
2623
2624 #ifndef PRODUCT
2625 void Matcher::record_new2old(Node* newn, Node* old) {
2626 _new2old_map.map(newn->_idx, old);
2627 if (!_reused.test_set(old->_igv_idx)) {
2628 // Reuse the Ideal-level IGV identifier so that the node can be tracked
2629 // across matching. If there are multiple machine nodes expanded from the
2630 // same Ideal node, only one will reuse its IGV identifier.
2631 newn->_igv_idx = old->_igv_idx;
2632 }
2633 }
2634
2635 // machine-independent root to machine-dependent root
2636 void Matcher::dump_old2new_map() {
2637 _old2new_map.dump();
2638 }
2639 #endif // !PRODUCT
2640
2641 //---------------------------collect_null_checks-------------------------------
2642 // Find null checks in the ideal graph; write a machine-specific node for
2643 // it. Used by later implicit-null-check handling. Actually collects
2644 // either an IfTrue or IfFalse for the common NOT-null path, AND the ideal
2645 // value being tested.
2646 void Matcher::collect_null_checks( Node *proj, Node *orig_proj ) {
2647 Node *iff = proj->in(0);
2648 if( iff->Opcode() == Op_If ) {
2649 // During matching If's have Bool & Cmp side-by-side
2650 BoolNode *b = iff->in(1)->as_Bool();
2651 Node *cmp = iff->in(2);
2652 int opc = cmp->Opcode();
2653 if (opc != Op_CmpP && opc != Op_CmpN) return;
2654
2655 const Type* ct = cmp->in(2)->bottom_type();
2656 if (ct == TypePtr::NULL_PTR ||
2657 (opc == Op_CmpN && ct == TypeNarrowOop::NULL_PTR)) {
2658
2659 bool push_it = false;
2660 if( proj->Opcode() == Op_IfTrue ) {
2661 #ifndef PRODUCT
2662 extern uint all_null_checks_found;
2663 all_null_checks_found++;
2664 #endif
2665 if( b->_test._test == BoolTest::ne ) {
2666 push_it = true;
2667 }
2668 } else {
2669 assert( proj->Opcode() == Op_IfFalse, "" );
2670 if( b->_test._test == BoolTest::eq ) {
2671 push_it = true;
2672 }
2673 }
2674 if( push_it ) {
2675 _null_check_tests.push(proj);
2676 Node* val = cmp->in(1);
2677 #ifdef _LP64
2678 if (val->bottom_type()->isa_narrowoop() &&
2679 !Matcher::narrow_oop_use_complex_address()) {
2680 //
2681 // Look for DecodeN node which should be pinned to orig_proj.
2682 // On platforms (Sparc) which can not handle 2 adds
2683 // in addressing mode we have to keep a DecodeN node and
2684 // use it to do implicit null check in address.
2685 //
2686 // DecodeN node was pinned to non-null path (orig_proj) during
2687 // CastPP transformation in final_graph_reshaping_impl().
2688 //
2689 uint cnt = orig_proj->outcnt();
2690 for (uint i = 0; i < orig_proj->outcnt(); i++) {
2691 Node* d = orig_proj->raw_out(i);
2692 if (d->is_DecodeN() && d->in(1) == val) {
2693 val = d;
2694 val->set_req(0, nullptr); // Unpin now.
2695 // Mark this as special case to distinguish from
2696 // a regular case: CmpP(DecodeN, null).
2697 val = (Node*)(((intptr_t)val) | 1);
2698 break;
2699 }
2700 }
2701 }
2702 #endif
2703 _null_check_tests.push(val);
2704 }
2705 }
2706 }
2707 }
2708
2709 //---------------------------validate_null_checks------------------------------
2710 // Its possible that the value being null checked is not the root of a match
2711 // tree. If so, I cannot use the value in an implicit null check.
2712 void Matcher::validate_null_checks( ) {
2713 uint cnt = _null_check_tests.size();
2714 for( uint i=0; i < cnt; i+=2 ) {
2715 Node *test = _null_check_tests[i];
2716 Node *val = _null_check_tests[i+1];
2717 bool is_decoden = ((intptr_t)val) & 1;
2718 val = (Node*)(((intptr_t)val) & ~1);
2719 if (has_new_node(val)) {
2720 Node* new_val = new_node(val);
2721 if (is_decoden) {
2722 assert(val->is_DecodeNarrowPtr() && val->in(0) == nullptr, "sanity");
2723 // Note: new_val may have a control edge if
2724 // the original ideal node DecodeN was matched before
2725 // it was unpinned in Matcher::collect_null_checks().
2726 // Unpin the mach node and mark it.
2727 new_val->set_req(0, nullptr);
2728 new_val = (Node*)(((intptr_t)new_val) | 1);
2729 }
2730 // Is a match-tree root, so replace with the matched value
2731 _null_check_tests.map(i+1, new_val);
2732 } else {
2733 // Yank from candidate list
2734 _null_check_tests.map(i+1,_null_check_tests[--cnt]);
2735 _null_check_tests.map(i,_null_check_tests[--cnt]);
2736 _null_check_tests.pop();
2737 _null_check_tests.pop();
2738 i-=2;
2739 }
2740 }
2741 }
2742
2743 bool Matcher::gen_narrow_oop_implicit_null_checks() {
2744 // Advice matcher to perform null checks on the narrow oop side.
2745 // Implicit checks are not possible on the uncompressed oop side anyway
2746 // (at least not for read accesses).
2747 // Performs significantly better (especially on Power 6).
2748 if (!os::zero_page_read_protected()) {
2749 return true;
2750 }
2751 return CompressedOops::use_implicit_null_checks() &&
2752 (narrow_oop_use_complex_address() ||
2753 CompressedOops::base() != nullptr);
2754 }
2755
2756 // Compute RegMask for an ideal register.
2757 const RegMask* Matcher::regmask_for_ideal_register(uint ideal_reg, Node* ret) {
2758 assert(!C->failing_internal() || C->failure_is_artificial(), "already failing.");
2759 if (C->failing()) {
2760 return nullptr;
2761 }
2762 const Type* t = Type::mreg2type[ideal_reg];
2763 if (t == nullptr) {
2764 assert(ideal_reg >= Op_VecA && ideal_reg <= Op_VecZ, "not a vector: %d", ideal_reg);
2765 return nullptr; // not supported
2766 }
2767 Node* fp = ret->in(TypeFunc::FramePtr);
2768 Node* mem = ret->in(TypeFunc::Memory);
2769 const TypePtr* atp = TypePtr::BOTTOM;
2770 MemNode::MemOrd mo = MemNode::unordered;
2771
2772 Node* spill;
2773 switch (ideal_reg) {
2774 case Op_RegN: spill = new LoadNNode(nullptr, mem, fp, atp, t->is_narrowoop(), mo); break;
2775 case Op_RegI: spill = new LoadINode(nullptr, mem, fp, atp, t->is_int(), mo); break;
2776 case Op_RegP: spill = new LoadPNode(nullptr, mem, fp, atp, t->is_ptr(), mo); break;
2777 case Op_RegF: spill = new LoadFNode(nullptr, mem, fp, atp, t, mo); break;
2778 case Op_RegD: spill = new LoadDNode(nullptr, mem, fp, atp, t, mo); break;
2779 case Op_RegL: spill = new LoadLNode(nullptr, mem, fp, atp, t->is_long(), mo); break;
2780
2781 case Op_VecA: // fall-through
2782 case Op_VecS: // fall-through
2783 case Op_VecD: // fall-through
2784 case Op_VecX: // fall-through
2785 case Op_VecY: // fall-through
2786 case Op_VecZ: spill = new LoadVectorNode(nullptr, mem, fp, atp, t->is_vect()); break;
2787 case Op_RegVectMask: return Matcher::predicate_reg_mask();
2788
2789 default: ShouldNotReachHere();
2790 }
2791 MachNode* mspill = match_tree(spill);
2792 assert(mspill != nullptr || C->failure_is_artificial(), "matching failed: %d", ideal_reg);
2793 if (C->failing()) {
2794 return nullptr;
2795 }
2796 // Handle generic vector operand case
2797 if (Matcher::supports_generic_vector_operands && t->isa_vect()) {
2798 specialize_mach_node(mspill);
2799 }
2800 return &mspill->out_RegMask();
2801 }
2802
2803 // Process Mach IR right after selection phase is over.
2804 void Matcher::do_postselect_cleanup() {
2805 if (supports_generic_vector_operands) {
2806 specialize_generic_vector_operands();
2807 if (C->failing()) return;
2808 }
2809 }
2810
2811 //----------------------------------------------------------------------
2812 // Generic machine operands elision.
2813 //----------------------------------------------------------------------
2814
2815 // Compute concrete vector operand for a generic TEMP vector mach node based on its user info.
2816 void Matcher::specialize_temp_node(MachTempNode* tmp, MachNode* use, uint idx) {
2817 assert(use->in(idx) == tmp, "not a user");
2818 assert(!Matcher::is_generic_vector(use->_opnds[0]), "use not processed yet");
2819
2820 if ((uint)idx == use->two_adr()) { // DEF_TEMP case
2821 tmp->_opnds[0] = use->_opnds[0]->clone();
2822 } else {
2823 uint ideal_vreg = vector_ideal_reg(C->max_vector_size());
2824 tmp->_opnds[0] = Matcher::pd_specialize_generic_vector_operand(tmp->_opnds[0], ideal_vreg, true /*is_temp*/);
2825 }
2826 }
2827
2828 // Compute concrete vector operand for a generic DEF/USE vector operand (of mach node m at index idx).
2829 MachOper* Matcher::specialize_vector_operand(MachNode* m, uint opnd_idx) {
2830 assert(Matcher::is_generic_vector(m->_opnds[opnd_idx]), "repeated updates");
2831 Node* def = nullptr;
2832 if (opnd_idx == 0) { // DEF
2833 def = m; // use mach node itself to compute vector operand type
2834 } else {
2835 int base_idx = m->operand_index(opnd_idx);
2836 def = m->in(base_idx);
2837 if (def->is_Mach()) {
2838 if (def->is_MachTemp() && Matcher::is_generic_vector(def->as_Mach()->_opnds[0])) {
2839 specialize_temp_node(def->as_MachTemp(), m, base_idx); // MachTemp node use site
2840 } else if (is_reg2reg_move(def->as_Mach())) {
2841 def = def->in(1); // skip over generic reg-to-reg moves
2842 }
2843 }
2844 }
2845 assert(def->bottom_type()->isa_vect(), "not a vector");
2846 uint ideal_vreg = def->bottom_type()->ideal_reg();
2847 return Matcher::pd_specialize_generic_vector_operand(m->_opnds[opnd_idx], ideal_vreg, false /*is_temp*/);
2848 }
2849
2850 void Matcher::specialize_mach_node(MachNode* m) {
2851 assert(!m->is_MachTemp(), "processed along with its user");
2852 // For generic use operands pull specific register class operands from
2853 // its def instruction's output operand (def operand).
2854 for (uint i = 0; i < m->num_opnds(); i++) {
2855 if (Matcher::is_generic_vector(m->_opnds[i])) {
2856 m->_opnds[i] = specialize_vector_operand(m, i);
2857 }
2858 }
2859 }
2860
2861 // Replace generic vector operands with concrete vector operands and eliminate generic reg-to-reg moves from the graph.
2862 void Matcher::specialize_generic_vector_operands() {
2863 assert(supports_generic_vector_operands, "sanity");
2864 ResourceMark rm;
2865
2866 // Replace generic vector operands (vec/legVec) with concrete ones (vec[SDXYZ]/legVec[SDXYZ])
2867 // and remove reg-to-reg vector moves (MoveVec2Leg and MoveLeg2Vec).
2868 Unique_Node_List live_nodes;
2869 C->identify_useful_nodes(live_nodes);
2870
2871 while (live_nodes.size() > 0) {
2872 MachNode* m = live_nodes.pop()->isa_Mach();
2873 if (m != nullptr) {
2874 if (Matcher::is_reg2reg_move(m)) {
2875 // Register allocator properly handles vec <=> leg moves using register masks.
2876 int opnd_idx = m->operand_index(1);
2877 Node* def = m->in(opnd_idx);
2878 m->subsume_by(def, C);
2879 } else if (m->is_MachTemp()) {
2880 // process MachTemp nodes at use site (see Matcher::specialize_vector_operand)
2881 } else {
2882 specialize_mach_node(m);
2883 }
2884 }
2885 }
2886 }
2887
2888 uint Matcher::vector_length(const Node* n) {
2889 const TypeVect* vt = n->bottom_type()->is_vect();
2890 return vt->length();
2891 }
2892
2893 uint Matcher::vector_length(const MachNode* use, const MachOper* opnd) {
2894 int def_idx = use->operand_index(opnd);
2895 Node* def = use->in(def_idx);
2896 return def->bottom_type()->is_vect()->length();
2897 }
2898
2899 uint Matcher::vector_length_in_bytes(const Node* n) {
2900 const TypeVect* vt = n->bottom_type()->is_vect();
2901 return vt->length_in_bytes();
2902 }
2903
2904 uint Matcher::vector_length_in_bytes(const MachNode* use, const MachOper* opnd) {
2905 uint def_idx = use->operand_index(opnd);
2906 Node* def = use->in(def_idx);
2907 return def->bottom_type()->is_vect()->length_in_bytes();
2908 }
2909
2910 BasicType Matcher::vector_element_basic_type(const Node* n) {
2911 const TypeVect* vt = n->bottom_type()->is_vect();
2912 return vt->element_basic_type();
2913 }
2914
2915 BasicType Matcher::vector_element_basic_type(const MachNode* use, const MachOper* opnd) {
2916 int def_idx = use->operand_index(opnd);
2917 Node* def = use->in(def_idx);
2918 return def->bottom_type()->is_vect()->element_basic_type();
2919 }
2920
2921 bool Matcher::is_non_long_integral_vector(const Node* n) {
2922 BasicType bt = vector_element_basic_type(n);
2923 assert(bt != T_CHAR, "char is not allowed in vector");
2924 return is_subword_type(bt) || bt == T_INT;
2925 }
2926
2927 bool Matcher::is_encode_and_store_pattern(const Node* n, const Node* m) {
2928 if (n == nullptr ||
2929 m == nullptr ||
2930 n->Opcode() != Op_StoreN ||
2931 !m->is_EncodeP() ||
2932 n->as_Store()->barrier_data() == 0) {
2933 return false;
2934 }
2935 assert(m == n->in(MemNode::ValueIn), "m should be input to n");
2936 return true;
2937 }
2938
2939 #ifdef ASSERT
2940 bool Matcher::verify_after_postselect_cleanup() {
2941 assert(!C->failing_internal() || C->failure_is_artificial(), "sanity");
2942 if (supports_generic_vector_operands) {
2943 Unique_Node_List useful;
2944 C->identify_useful_nodes(useful);
2945 for (uint i = 0; i < useful.size(); i++) {
2946 MachNode* m = useful.at(i)->isa_Mach();
2947 if (m != nullptr) {
2948 assert(!Matcher::is_reg2reg_move(m), "no MoveVec nodes allowed");
2949 for (uint j = 0; j < m->num_opnds(); j++) {
2950 assert(!Matcher::is_generic_vector(m->_opnds[j]), "no generic vector operands allowed");
2951 }
2952 }
2953 }
2954 }
2955 return true;
2956 }
2957 #endif // ASSERT
2958
2959 // Used by the DFA in dfa_xxx.cpp. Check for a following barrier or
2960 // atomic instruction acting as a store_load barrier without any
2961 // intervening volatile load, and thus we don't need a barrier here.
2962 // We retain the Node to act as a compiler ordering barrier.
2963 bool Matcher::post_store_load_barrier(const Node* vmb) {
2964 Compile* C = Compile::current();
2965 assert(vmb->is_MemBar(), "");
2966 assert(vmb->Opcode() != Op_MemBarAcquire && vmb->Opcode() != Op_LoadFence, "");
2967 const MemBarNode* membar = vmb->as_MemBar();
2968
2969 // Get the Ideal Proj node, ctrl, that can be used to iterate forward
2970 Node* ctrl = nullptr;
2971 for (DUIterator_Fast imax, i = membar->fast_outs(imax); i < imax; i++) {
2972 Node* p = membar->fast_out(i);
2973 assert(p->is_Proj(), "only projections here");
2974 if ((p->as_Proj()->_con == TypeFunc::Control) &&
2975 !C->node_arena()->contains(p)) { // Unmatched old-space only
2976 ctrl = p;
2977 break;
2978 }
2979 }
2980 assert((ctrl != nullptr), "missing control projection");
2981
2982 for (DUIterator_Fast jmax, j = ctrl->fast_outs(jmax); j < jmax; j++) {
2983 Node *x = ctrl->fast_out(j);
2984 int xop = x->Opcode();
2985
2986 // We don't need current barrier if we see another or a lock
2987 // before seeing volatile load.
2988 //
2989 // Op_Fastunlock previously appeared in the Op_* list below.
2990 // With the advent of 1-0 lock operations we're no longer guaranteed
2991 // that a monitor exit operation contains a serializing instruction.
2992
2993 if (xop == Op_MemBarVolatile ||
2994 xop == Op_CompareAndExchangeB ||
2995 xop == Op_CompareAndExchangeS ||
2996 xop == Op_CompareAndExchangeI ||
2997 xop == Op_CompareAndExchangeL ||
2998 xop == Op_CompareAndExchangeP ||
2999 xop == Op_CompareAndExchangeN ||
3000 xop == Op_WeakCompareAndSwapB ||
3001 xop == Op_WeakCompareAndSwapS ||
3002 xop == Op_WeakCompareAndSwapL ||
3003 xop == Op_WeakCompareAndSwapP ||
3004 xop == Op_WeakCompareAndSwapN ||
3005 xop == Op_WeakCompareAndSwapI ||
3006 xop == Op_CompareAndSwapB ||
3007 xop == Op_CompareAndSwapS ||
3008 xop == Op_CompareAndSwapL ||
3009 xop == Op_CompareAndSwapP ||
3010 xop == Op_CompareAndSwapN ||
3011 xop == Op_CompareAndSwapI ||
3012 BarrierSet::barrier_set()->barrier_set_c2()->matcher_is_store_load_barrier(x, xop)) {
3013 return true;
3014 }
3015
3016 // Op_FastLock previously appeared in the Op_* list above.
3017 if (xop == Op_FastLock) {
3018 return true;
3019 }
3020
3021 if (x->is_MemBar()) {
3022 // We must retain this membar if there is an upcoming volatile
3023 // load, which will be followed by acquire membar.
3024 if (xop == Op_MemBarAcquire || xop == Op_LoadFence) {
3025 return false;
3026 } else {
3027 // For other kinds of barriers, check by pretending we
3028 // are them, and seeing if we can be removed.
3029 return post_store_load_barrier(x->as_MemBar());
3030 }
3031 }
3032
3033 // probably not necessary to check for these
3034 if (x->is_Call() || x->is_SafePoint() || x->is_block_proj()) {
3035 return false;
3036 }
3037 }
3038 return false;
3039 }
3040
3041 // Check whether node n is a branch to an uncommon trap that we could
3042 // optimize as test with very high branch costs in case of going to
3043 // the uncommon trap. The code must be able to be recompiled to use
3044 // a cheaper test.
3045 bool Matcher::branches_to_uncommon_trap(const Node *n) {
3046 // Don't do it for natives, adapters, or runtime stubs
3047 Compile *C = Compile::current();
3048 if (!C->is_method_compilation()) return false;
3049
3050 assert(n->is_If(), "You should only call this on if nodes.");
3051 IfNode *ifn = n->as_If();
3052
3053 Node *ifFalse = nullptr;
3054 for (DUIterator_Fast imax, i = ifn->fast_outs(imax); i < imax; i++) {
3055 if (ifn->fast_out(i)->is_IfFalse()) {
3056 ifFalse = ifn->fast_out(i);
3057 break;
3058 }
3059 }
3060 assert(ifFalse, "An If should have an ifFalse. Graph is broken.");
3061
3062 Node *reg = ifFalse;
3063 int cnt = 4; // We must protect against cycles. Limit to 4 iterations.
3064 // Alternatively use visited set? Seems too expensive.
3065 while (reg != nullptr && cnt > 0) {
3066 CallNode *call = nullptr;
3067 RegionNode *nxt_reg = nullptr;
3068 for (DUIterator_Fast imax, i = reg->fast_outs(imax); i < imax; i++) {
3069 Node *o = reg->fast_out(i);
3070 if (o->is_Call()) {
3071 call = o->as_Call();
3072 }
3073 if (o->is_Region()) {
3074 nxt_reg = o->as_Region();
3075 }
3076 }
3077
3078 if (call &&
3079 call->entry_point() == OptoRuntime::uncommon_trap_blob()->entry_point()) {
3080 const Type* trtype = call->in(TypeFunc::Parms)->bottom_type();
3081 if (trtype->isa_int() && trtype->is_int()->is_con()) {
3082 jint tr_con = trtype->is_int()->get_con();
3083 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
3084 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
3085 assert((int)reason < (int)BitsPerInt, "recode bit map");
3086
3087 if (is_set_nth_bit(C->allowed_deopt_reasons(), (int)reason)
3088 && action != Deoptimization::Action_none) {
3089 // This uncommon trap is sure to recompile, eventually.
3090 // When that happens, C->too_many_traps will prevent
3091 // this transformation from happening again.
3092 return true;
3093 }
3094 }
3095 }
3096
3097 reg = nxt_reg;
3098 cnt--;
3099 }
3100
3101 return false;
3102 }
3103
3104 //=============================================================================
3105 //---------------------------State---------------------------------------------
3106 State::State(void) : _rule() {
3107 #ifdef ASSERT
3108 _id = 0;
3109 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe);
3110 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d);
3111 #endif
3112 }
3113
3114 #ifdef ASSERT
3115 State::~State() {
3116 _id = 99;
3117 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe);
3118 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d);
3119 memset(_cost, -3, sizeof(_cost));
3120 memset(_rule, -3, sizeof(_rule));
3121 }
3122 #endif
3123
3124 #ifndef PRODUCT
3125 //---------------------------dump----------------------------------------------
3126 void State::dump() {
3127 tty->print("\n");
3128 dump(0);
3129 }
3130
3131 void State::dump(int depth) {
3132 for (int j = 0; j < depth; j++) {
3133 tty->print(" ");
3134 }
3135 tty->print("--N: ");
3136 _leaf->dump();
3137 uint i;
3138 for (i = 0; i < _LAST_MACH_OPER; i++) {
3139 // Check for valid entry
3140 if (valid(i)) {
3141 for (int j = 0; j < depth; j++) {
3142 tty->print(" ");
3143 }
3144 assert(cost(i) != max_juint, "cost must be a valid value");
3145 assert(rule(i) < _last_Mach_Node, "rule[i] must be valid rule");
3146 tty->print_cr("%s %d %s",
3147 ruleName[i], cost(i), ruleName[rule(i)] );
3148 }
3149 }
3150 tty->cr();
3151
3152 for (i = 0; i < 2; i++) {
3153 if (_kids[i]) {
3154 _kids[i]->dump(depth + 1);
3155 }
3156 }
3157 }
3158 #endif