1 /*
2 * Copyright (c) 1997, 2022, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/shared/barrierSet.hpp"
27 #include "gc/shared/c2/barrierSetC2.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "memory/resourceArea.hpp"
30 #include "oops/compressedOops.hpp"
31 #include "opto/ad.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/callnode.hpp"
34 #include "opto/idealGraphPrinter.hpp"
35 #include "opto/matcher.hpp"
36 #include "opto/memnode.hpp"
37 #include "opto/movenode.hpp"
38 #include "opto/opcodes.hpp"
39 #include "opto/regmask.hpp"
40 #include "opto/rootnode.hpp"
41 #include "opto/runtime.hpp"
42 #include "opto/type.hpp"
43 #include "opto/vectornode.hpp"
44 #include "runtime/os.inline.hpp"
45 #include "runtime/sharedRuntime.hpp"
46 #include "utilities/align.hpp"
47
48 OptoReg::Name OptoReg::c_frame_pointer;
49
50 const RegMask *Matcher::idealreg2regmask[_last_machine_leaf];
51 RegMask Matcher::mreg2regmask[_last_Mach_Reg];
52 RegMask Matcher::caller_save_regmask;
53 RegMask Matcher::caller_save_regmask_exclude_soe;
54 RegMask Matcher::mh_caller_save_regmask;
55 RegMask Matcher::mh_caller_save_regmask_exclude_soe;
56 RegMask Matcher::STACK_ONLY_mask;
57 RegMask Matcher::c_frame_ptr_mask;
58 const uint Matcher::_begin_rematerialize = _BEGIN_REMATERIALIZE;
59 const uint Matcher::_end_rematerialize = _END_REMATERIALIZE;
60
61 //---------------------------Matcher-------------------------------------------
62 Matcher::Matcher()
63 : PhaseTransform( Phase::Ins_Select ),
64 _states_arena(Chunk::medium_size, mtCompiler),
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] = NULL;
87 idealreg2spillmask [Op_RegN] = NULL;
88 idealreg2spillmask [Op_RegL] = NULL;
89 idealreg2spillmask [Op_RegF] = NULL;
90 idealreg2spillmask [Op_RegD] = NULL;
91 idealreg2spillmask [Op_RegP] = NULL;
92 idealreg2spillmask [Op_VecA] = NULL;
93 idealreg2spillmask [Op_VecS] = NULL;
94 idealreg2spillmask [Op_VecD] = NULL;
95 idealreg2spillmask [Op_VecX] = NULL;
96 idealreg2spillmask [Op_VecY] = NULL;
97 idealreg2spillmask [Op_VecZ] = NULL;
98 idealreg2spillmask [Op_RegFlags] = NULL;
99 idealreg2spillmask [Op_RegVectMask] = NULL;
100
101 idealreg2debugmask [Op_RegI] = NULL;
102 idealreg2debugmask [Op_RegN] = NULL;
103 idealreg2debugmask [Op_RegL] = NULL;
104 idealreg2debugmask [Op_RegF] = NULL;
105 idealreg2debugmask [Op_RegD] = NULL;
106 idealreg2debugmask [Op_RegP] = NULL;
107 idealreg2debugmask [Op_VecA] = NULL;
108 idealreg2debugmask [Op_VecS] = NULL;
109 idealreg2debugmask [Op_VecD] = NULL;
110 idealreg2debugmask [Op_VecX] = NULL;
111 idealreg2debugmask [Op_VecY] = NULL;
112 idealreg2debugmask [Op_VecZ] = NULL;
113 idealreg2debugmask [Op_RegFlags] = NULL;
114 idealreg2debugmask [Op_RegVectMask] = NULL;
115
116 idealreg2mhdebugmask[Op_RegI] = NULL;
117 idealreg2mhdebugmask[Op_RegN] = NULL;
118 idealreg2mhdebugmask[Op_RegL] = NULL;
119 idealreg2mhdebugmask[Op_RegF] = NULL;
120 idealreg2mhdebugmask[Op_RegD] = NULL;
121 idealreg2mhdebugmask[Op_RegP] = NULL;
122 idealreg2mhdebugmask[Op_VecA] = NULL;
123 idealreg2mhdebugmask[Op_VecS] = NULL;
124 idealreg2mhdebugmask[Op_VecD] = NULL;
125 idealreg2mhdebugmask[Op_VecX] = NULL;
126 idealreg2mhdebugmask[Op_VecY] = NULL;
127 idealreg2mhdebugmask[Op_VecZ] = NULL;
128 idealreg2mhdebugmask[Op_RegFlags] = NULL;
129 idealreg2mhdebugmask[Op_RegVectMask] = NULL;
130
131 debug_only(_mem_node = NULL;) // 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 C->record_method_not_compilable("unsupported incoming calling sequence");
146 return OptoReg::Bad;
147 }
148 return warped;
149 }
150 return OptoReg::as_OptoReg(reg);
151 }
152
153 //---------------------------compute_old_SP------------------------------------
154 OptoReg::Name Compile::compute_old_SP() {
155 int fixed = fixed_slots();
156 int preserve = in_preserve_stack_slots();
157 return OptoReg::stack2reg(align_up(fixed + preserve, (int)Matcher::stack_alignment_in_slots()));
158 }
159
160
161
162 #ifdef ASSERT
163 void Matcher::verify_new_nodes_only(Node* xroot) {
164 // Make sure that the new graph only references new nodes
165 ResourceMark rm;
166 Unique_Node_List worklist;
167 VectorSet visited;
168 worklist.push(xroot);
169 while (worklist.size() > 0) {
170 Node* n = worklist.pop();
171 visited.set(n->_idx);
172 assert(C->node_arena()->contains(n), "dead node");
173 for (uint j = 0; j < n->req(); j++) {
174 Node* in = n->in(j);
175 if (in != NULL) {
176 assert(C->node_arena()->contains(in), "dead node");
177 if (!visited.test(in->_idx)) {
178 worklist.push(in);
179 }
180 }
181 }
182 }
183 }
184 #endif
185
186 // Array of RegMask, one per returned values (inline type instances can
187 // be returned as multiple return values, one per field)
188 RegMask* Matcher::return_values_mask(const TypeFunc* tf) {
189 const TypeTuple* range = tf->range_cc();
190 uint cnt = range->cnt() - TypeFunc::Parms;
191 if (cnt == 0) {
192 return NULL;
193 }
194 RegMask* mask = NEW_RESOURCE_ARRAY(RegMask, cnt);
195 BasicType* sig_bt = NEW_RESOURCE_ARRAY(BasicType, cnt);
196 VMRegPair* vm_parm_regs = NEW_RESOURCE_ARRAY(VMRegPair, cnt);
197 for (uint i = 0; i < cnt; i++) {
198 sig_bt[i] = range->field_at(i+TypeFunc::Parms)->basic_type();
199 }
200
201 int regs = SharedRuntime::java_return_convention(sig_bt, vm_parm_regs, cnt);
202 if (regs <= 0) {
203 // We ran out of registers to store the IsInit information for a nullable inline type return.
204 // Since it is only set in the 'call_epilog', we can simply put it on the stack.
205 assert(tf->returns_inline_type_as_fields(), "should have been tested during graph construction");
206 // TODO 8284443 Can we teach the register allocator to reserve a stack slot instead?
207 // mask[--cnt] = STACK_ONLY_mask does not work (test with -XX:+StressGCM)
208 int slot = C->fixed_slots() - 2;
209 if (C->needs_stack_repair()) {
210 slot -= 2; // Account for stack increment value
211 }
212 mask[--cnt].Clear();
213 mask[cnt].Insert(OptoReg::stack2reg(slot));
214 }
215 for (uint i = 0; i < cnt; i++) {
216 mask[i].Clear();
217
218 OptoReg::Name reg1 = OptoReg::as_OptoReg(vm_parm_regs[i].first());
219 if (OptoReg::is_valid(reg1)) {
220 mask[i].Insert(reg1);
221 }
222 OptoReg::Name reg2 = OptoReg::as_OptoReg(vm_parm_regs[i].second());
223 if (OptoReg::is_valid(reg2)) {
224 mask[i].Insert(reg2);
225 }
226 }
227
228 return mask;
229 }
230
231 //---------------------------match---------------------------------------------
232 void Matcher::match( ) {
233 if( MaxLabelRootDepth < 100 ) { // Too small?
234 assert(false, "invalid MaxLabelRootDepth, increase it to 100 minimum");
235 MaxLabelRootDepth = 100;
236 }
237 // One-time initialization of some register masks.
238 init_spill_mask( C->root()->in(1) );
239 _return_addr_mask = return_addr();
240 #ifdef _LP64
241 // Pointers take 2 slots in 64-bit land
242 _return_addr_mask.Insert(OptoReg::add(return_addr(),1));
243 #endif
244
245 // Map Java-signature return types into return register-value
246 // machine registers.
247 _return_values_mask = return_values_mask(C->tf());
248
249 // ---------------
250 // Frame Layout
251
252 // Need the method signature to determine the incoming argument types,
253 // because the types determine which registers the incoming arguments are
254 // in, and this affects the matched code.
255 const TypeTuple *domain = C->tf()->domain_cc();
256 uint argcnt = domain->cnt() - TypeFunc::Parms;
257 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt );
258 VMRegPair *vm_parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt );
259 _parm_regs = NEW_RESOURCE_ARRAY( OptoRegPair, argcnt );
260 _calling_convention_mask = NEW_RESOURCE_ARRAY( RegMask, argcnt );
261 uint i;
262 for( i = 0; i<argcnt; i++ ) {
263 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type();
264 }
265
266 // Pass array of ideal registers and length to USER code (from the AD file)
267 // that will convert this to an array of register numbers.
268 const StartNode *start = C->start();
269 start->calling_convention( sig_bt, vm_parm_regs, argcnt );
270 #ifdef ASSERT
271 // Sanity check users' calling convention. Real handy while trying to
272 // get the initial port correct.
273 { for (uint i = 0; i<argcnt; i++) {
274 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) {
275 assert(domain->field_at(i+TypeFunc::Parms)==Type::HALF, "only allowed on halve" );
276 _parm_regs[i].set_bad();
277 continue;
278 }
279 VMReg parm_reg = vm_parm_regs[i].first();
280 assert(parm_reg->is_valid(), "invalid arg?");
281 if (parm_reg->is_reg()) {
282 OptoReg::Name opto_parm_reg = OptoReg::as_OptoReg(parm_reg);
283 assert(can_be_java_arg(opto_parm_reg) ||
284 C->stub_function() == CAST_FROM_FN_PTR(address, OptoRuntime::rethrow_C) ||
285 opto_parm_reg == inline_cache_reg(),
286 "parameters in register must be preserved by runtime stubs");
287 }
288 for (uint j = 0; j < i; j++) {
289 assert(parm_reg != vm_parm_regs[j].first(),
290 "calling conv. must produce distinct regs");
291 }
292 }
293 }
294 #endif
295
296 // Do some initial frame layout.
297
298 // Compute the old incoming SP (may be called FP) as
299 // OptoReg::stack0() + locks + in_preserve_stack_slots + pad2.
300 _old_SP = C->compute_old_SP();
301 assert( is_even(_old_SP), "must be even" );
302
303 // Compute highest incoming stack argument as
304 // _old_SP + out_preserve_stack_slots + incoming argument size.
305 _in_arg_limit = OptoReg::add(_old_SP, C->out_preserve_stack_slots());
306 assert( is_even(_in_arg_limit), "out_preserve must be even" );
307 for( i = 0; i < argcnt; i++ ) {
308 // Permit args to have no register
309 _calling_convention_mask[i].Clear();
310 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) {
311 continue;
312 }
313 // calling_convention returns stack arguments as a count of
314 // slots beyond OptoReg::stack0()/VMRegImpl::stack0. We need to convert this to
315 // the allocators point of view, taking into account all the
316 // preserve area, locks & pad2.
317
318 OptoReg::Name reg1 = warp_incoming_stk_arg(vm_parm_regs[i].first());
319 if( OptoReg::is_valid(reg1))
320 _calling_convention_mask[i].Insert(reg1);
321
322 OptoReg::Name reg2 = warp_incoming_stk_arg(vm_parm_regs[i].second());
323 if( OptoReg::is_valid(reg2))
324 _calling_convention_mask[i].Insert(reg2);
325
326 // Saved biased stack-slot register number
327 _parm_regs[i].set_pair(reg2, reg1);
328 }
329
330 // Finally, make sure the incoming arguments take up an even number of
331 // words, in case the arguments or locals need to contain doubleword stack
332 // slots. The rest of the system assumes that stack slot pairs (in
333 // particular, in the spill area) which look aligned will in fact be
334 // aligned relative to the stack pointer in the target machine. Double
335 // stack slots will always be allocated aligned.
336 _new_SP = OptoReg::Name(align_up(_in_arg_limit, (int)RegMask::SlotsPerLong));
337
338 // Compute highest outgoing stack argument as
339 // _new_SP + out_preserve_stack_slots + max(outgoing argument size).
340 _out_arg_limit = OptoReg::add(_new_SP, C->out_preserve_stack_slots());
341 assert( is_even(_out_arg_limit), "out_preserve must be even" );
342
343 if (!RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1))) {
344 // the compiler cannot represent this method's calling sequence
345 C->record_method_not_compilable("must be able to represent all call arguments in reg mask");
346 }
347
348 if (C->failing()) return; // bailed out on incoming arg failure
349
350 // ---------------
351 // Collect roots of matcher trees. Every node for which
352 // _shared[_idx] is cleared is guaranteed to not be shared, and thus
353 // can be a valid interior of some tree.
354 find_shared( C->root() );
355 find_shared( C->top() );
356
357 C->print_method(PHASE_BEFORE_MATCHING, 1);
358
359 // Create new ideal node ConP #NULL even if it does exist in old space
360 // to avoid false sharing if the corresponding mach node is not used.
361 // The corresponding mach node is only used in rare cases for derived
362 // pointers.
363 Node* new_ideal_null = ConNode::make(TypePtr::NULL_PTR);
364
365 // Swap out to old-space; emptying new-space
366 Arena *old = C->node_arena()->move_contents(C->old_arena());
367
368 // Save debug and profile information for nodes in old space:
369 _old_node_note_array = C->node_note_array();
370 if (_old_node_note_array != NULL) {
371 C->set_node_note_array(new(C->comp_arena()) GrowableArray<Node_Notes*>
372 (C->comp_arena(), _old_node_note_array->length(),
373 0, NULL));
374 }
375
376 // Pre-size the new_node table to avoid the need for range checks.
377 grow_new_node_array(C->unique());
378
379 // Reset node counter so MachNodes start with _idx at 0
380 int live_nodes = C->live_nodes();
381 C->set_unique(0);
382 C->reset_dead_node_list();
383
384 // Recursively match trees from old space into new space.
385 // Correct leaves of new-space Nodes; they point to old-space.
386 _visited.clear();
387 C->set_cached_top_node(xform( C->top(), live_nodes ));
388 if (!C->failing()) {
389 Node* xroot = xform( C->root(), 1 );
390 if (xroot == NULL) {
391 Matcher::soft_match_failure(); // recursive matching process failed
392 C->record_method_not_compilable("instruction match failed");
393 } else {
394 // During matching shared constants were attached to C->root()
395 // because xroot wasn't available yet, so transfer the uses to
396 // the xroot.
397 for( DUIterator_Fast jmax, j = C->root()->fast_outs(jmax); j < jmax; j++ ) {
398 Node* n = C->root()->fast_out(j);
399 if (C->node_arena()->contains(n)) {
400 assert(n->in(0) == C->root(), "should be control user");
401 n->set_req(0, xroot);
402 --j;
403 --jmax;
404 }
405 }
406
407 // Generate new mach node for ConP #NULL
408 assert(new_ideal_null != NULL, "sanity");
409 _mach_null = match_tree(new_ideal_null);
410 // Don't set control, it will confuse GCM since there are no uses.
411 // The control will be set when this node is used first time
412 // in find_base_for_derived().
413 assert(_mach_null != NULL, "");
414
415 C->set_root(xroot->is_Root() ? xroot->as_Root() : NULL);
416
417 #ifdef ASSERT
418 verify_new_nodes_only(xroot);
419 #endif
420 }
421 }
422 if (C->top() == NULL || C->root() == NULL) {
423 C->record_method_not_compilable("graph lost"); // %%% cannot happen?
424 }
425 if (C->failing()) {
426 // delete old;
427 old->destruct_contents();
428 return;
429 }
430 assert( C->top(), "" );
431 assert( C->root(), "" );
432 validate_null_checks();
433
434 // Now smoke old-space
435 NOT_DEBUG( old->destruct_contents() );
436
437 // ------------------------
438 // Set up save-on-entry registers.
439 Fixup_Save_On_Entry( );
440
441 { // Cleanup mach IR after selection phase is over.
442 Compile::TracePhase tp("postselect_cleanup", &timers[_t_postselect_cleanup]);
443 do_postselect_cleanup();
444 if (C->failing()) return;
445 assert(verify_after_postselect_cleanup(), "");
446 }
447 }
448
449 //------------------------------Fixup_Save_On_Entry----------------------------
450 // The stated purpose of this routine is to take care of save-on-entry
451 // registers. However, the overall goal of the Match phase is to convert into
452 // machine-specific instructions which have RegMasks to guide allocation.
453 // So what this procedure really does is put a valid RegMask on each input
454 // to the machine-specific variations of all Return, TailCall and Halt
455 // instructions. It also adds edgs to define the save-on-entry values (and of
456 // course gives them a mask).
457
458 static RegMask *init_input_masks( uint size, RegMask &ret_adr, RegMask &fp ) {
459 RegMask *rms = NEW_RESOURCE_ARRAY( RegMask, size );
460 // Do all the pre-defined register masks
461 rms[TypeFunc::Control ] = RegMask::Empty;
462 rms[TypeFunc::I_O ] = RegMask::Empty;
463 rms[TypeFunc::Memory ] = RegMask::Empty;
464 rms[TypeFunc::ReturnAdr] = ret_adr;
465 rms[TypeFunc::FramePtr ] = fp;
466 return rms;
467 }
468
469 const int Matcher::scalable_predicate_reg_slots() {
470 assert(Matcher::has_predicated_vectors() && Matcher::supports_scalable_vector(),
471 "scalable predicate vector should be supported");
472 int vector_reg_bit_size = Matcher::scalable_vector_reg_size(T_BYTE) << LogBitsPerByte;
473 // We assume each predicate register is one-eighth of the size of
474 // scalable vector register, one mask bit per vector byte.
475 int predicate_reg_bit_size = vector_reg_bit_size >> 3;
476 // Compute number of slots which is required when scalable predicate
477 // register is spilled. E.g. if scalable vector register is 640 bits,
478 // predicate register is 80 bits, which is 2.5 * slots.
479 // We will round up the slot number to power of 2, which is required
480 // by find_first_set().
481 int slots = predicate_reg_bit_size & (BitsPerInt - 1)
482 ? (predicate_reg_bit_size >> LogBitsPerInt) + 1
483 : predicate_reg_bit_size >> LogBitsPerInt;
484 return round_up_power_of_2(slots);
485 }
486
487 #define NOF_STACK_MASKS (3*13)
488
489 // Create the initial stack mask used by values spilling to the stack.
490 // Disallow any debug info in outgoing argument areas by setting the
491 // initial mask accordingly.
492 void Matcher::init_first_stack_mask() {
493
494 // Allocate storage for spill masks as masks for the appropriate load type.
495 RegMask *rms = (RegMask*)C->comp_arena()->AmallocWords(sizeof(RegMask) * NOF_STACK_MASKS);
496
497 // Initialize empty placeholder masks into the newly allocated arena
498 for (int i = 0; i < NOF_STACK_MASKS; i++) {
499 new (rms + i) RegMask();
500 }
501
502 idealreg2spillmask [Op_RegN] = &rms[0];
503 idealreg2spillmask [Op_RegI] = &rms[1];
504 idealreg2spillmask [Op_RegL] = &rms[2];
505 idealreg2spillmask [Op_RegF] = &rms[3];
506 idealreg2spillmask [Op_RegD] = &rms[4];
507 idealreg2spillmask [Op_RegP] = &rms[5];
508
509 idealreg2debugmask [Op_RegN] = &rms[6];
510 idealreg2debugmask [Op_RegI] = &rms[7];
511 idealreg2debugmask [Op_RegL] = &rms[8];
512 idealreg2debugmask [Op_RegF] = &rms[9];
513 idealreg2debugmask [Op_RegD] = &rms[10];
514 idealreg2debugmask [Op_RegP] = &rms[11];
515
516 idealreg2mhdebugmask[Op_RegN] = &rms[12];
517 idealreg2mhdebugmask[Op_RegI] = &rms[13];
518 idealreg2mhdebugmask[Op_RegL] = &rms[14];
519 idealreg2mhdebugmask[Op_RegF] = &rms[15];
520 idealreg2mhdebugmask[Op_RegD] = &rms[16];
521 idealreg2mhdebugmask[Op_RegP] = &rms[17];
522
523 idealreg2spillmask [Op_VecA] = &rms[18];
524 idealreg2spillmask [Op_VecS] = &rms[19];
525 idealreg2spillmask [Op_VecD] = &rms[20];
526 idealreg2spillmask [Op_VecX] = &rms[21];
527 idealreg2spillmask [Op_VecY] = &rms[22];
528 idealreg2spillmask [Op_VecZ] = &rms[23];
529
530 idealreg2debugmask [Op_VecA] = &rms[24];
531 idealreg2debugmask [Op_VecS] = &rms[25];
532 idealreg2debugmask [Op_VecD] = &rms[26];
533 idealreg2debugmask [Op_VecX] = &rms[27];
534 idealreg2debugmask [Op_VecY] = &rms[28];
535 idealreg2debugmask [Op_VecZ] = &rms[29];
536
537 idealreg2mhdebugmask[Op_VecA] = &rms[30];
538 idealreg2mhdebugmask[Op_VecS] = &rms[31];
539 idealreg2mhdebugmask[Op_VecD] = &rms[32];
540 idealreg2mhdebugmask[Op_VecX] = &rms[33];
541 idealreg2mhdebugmask[Op_VecY] = &rms[34];
542 idealreg2mhdebugmask[Op_VecZ] = &rms[35];
543
544 idealreg2spillmask [Op_RegVectMask] = &rms[36];
545 idealreg2debugmask [Op_RegVectMask] = &rms[37];
546 idealreg2mhdebugmask[Op_RegVectMask] = &rms[38];
547
548 OptoReg::Name i;
549
550 // At first, start with the empty mask
551 C->FIRST_STACK_mask().Clear();
552
553 // Add in the incoming argument area
554 OptoReg::Name init_in = OptoReg::add(_old_SP, C->out_preserve_stack_slots());
555 for (i = init_in; i < _in_arg_limit; i = OptoReg::add(i,1)) {
556 C->FIRST_STACK_mask().Insert(i);
557 }
558
559 // Add in all bits past the outgoing argument area
560 guarantee(RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1)),
561 "must be able to represent all call arguments in reg mask");
562 OptoReg::Name init = _out_arg_limit;
563 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) {
564 C->FIRST_STACK_mask().Insert(i);
565 }
566 // Finally, set the "infinite stack" bit.
567 C->FIRST_STACK_mask().set_AllStack();
568
569 // Make spill masks. Registers for their class, plus FIRST_STACK_mask.
570 RegMask aligned_stack_mask = C->FIRST_STACK_mask();
571 // Keep spill masks aligned.
572 aligned_stack_mask.clear_to_pairs();
573 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack");
574 RegMask scalable_stack_mask = aligned_stack_mask;
575
576 *idealreg2spillmask[Op_RegP] = *idealreg2regmask[Op_RegP];
577 #ifdef _LP64
578 *idealreg2spillmask[Op_RegN] = *idealreg2regmask[Op_RegN];
579 idealreg2spillmask[Op_RegN]->OR(C->FIRST_STACK_mask());
580 idealreg2spillmask[Op_RegP]->OR(aligned_stack_mask);
581 #else
582 idealreg2spillmask[Op_RegP]->OR(C->FIRST_STACK_mask());
583 #endif
584 *idealreg2spillmask[Op_RegI] = *idealreg2regmask[Op_RegI];
585 idealreg2spillmask[Op_RegI]->OR(C->FIRST_STACK_mask());
586 *idealreg2spillmask[Op_RegL] = *idealreg2regmask[Op_RegL];
587 idealreg2spillmask[Op_RegL]->OR(aligned_stack_mask);
588 *idealreg2spillmask[Op_RegF] = *idealreg2regmask[Op_RegF];
589 idealreg2spillmask[Op_RegF]->OR(C->FIRST_STACK_mask());
590 *idealreg2spillmask[Op_RegD] = *idealreg2regmask[Op_RegD];
591 idealreg2spillmask[Op_RegD]->OR(aligned_stack_mask);
592
593 if (Matcher::has_predicated_vectors()) {
594 *idealreg2spillmask[Op_RegVectMask] = *idealreg2regmask[Op_RegVectMask];
595 idealreg2spillmask[Op_RegVectMask]->OR(aligned_stack_mask);
596 } else {
597 *idealreg2spillmask[Op_RegVectMask] = RegMask::Empty;
598 }
599
600 if (Matcher::vector_size_supported(T_BYTE,4)) {
601 *idealreg2spillmask[Op_VecS] = *idealreg2regmask[Op_VecS];
602 idealreg2spillmask[Op_VecS]->OR(C->FIRST_STACK_mask());
603 } else {
604 *idealreg2spillmask[Op_VecS] = RegMask::Empty;
605 }
606
607 if (Matcher::vector_size_supported(T_FLOAT,2)) {
608 // For VecD we need dual alignment and 8 bytes (2 slots) for spills.
609 // RA guarantees such alignment since it is needed for Double and Long values.
610 *idealreg2spillmask[Op_VecD] = *idealreg2regmask[Op_VecD];
611 idealreg2spillmask[Op_VecD]->OR(aligned_stack_mask);
612 } else {
613 *idealreg2spillmask[Op_VecD] = RegMask::Empty;
614 }
615
616 if (Matcher::vector_size_supported(T_FLOAT,4)) {
617 // For VecX we need quadro alignment and 16 bytes (4 slots) for spills.
618 //
619 // RA can use input arguments stack slots for spills but until RA
620 // we don't know frame size and offset of input arg stack slots.
621 //
622 // Exclude last input arg stack slots to avoid spilling vectors there
623 // otherwise vector spills could stomp over stack slots in caller frame.
624 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1);
625 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecX); k++) {
626 aligned_stack_mask.Remove(in);
627 in = OptoReg::add(in, -1);
628 }
629 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecX);
630 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack");
631 *idealreg2spillmask[Op_VecX] = *idealreg2regmask[Op_VecX];
632 idealreg2spillmask[Op_VecX]->OR(aligned_stack_mask);
633 } else {
634 *idealreg2spillmask[Op_VecX] = RegMask::Empty;
635 }
636
637 if (Matcher::vector_size_supported(T_FLOAT,8)) {
638 // For VecY we need octo alignment and 32 bytes (8 slots) for spills.
639 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1);
640 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecY); k++) {
641 aligned_stack_mask.Remove(in);
642 in = OptoReg::add(in, -1);
643 }
644 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecY);
645 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack");
646 *idealreg2spillmask[Op_VecY] = *idealreg2regmask[Op_VecY];
647 idealreg2spillmask[Op_VecY]->OR(aligned_stack_mask);
648 } else {
649 *idealreg2spillmask[Op_VecY] = RegMask::Empty;
650 }
651
652 if (Matcher::vector_size_supported(T_FLOAT,16)) {
653 // For VecZ we need enough alignment and 64 bytes (16 slots) for spills.
654 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1);
655 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecZ); k++) {
656 aligned_stack_mask.Remove(in);
657 in = OptoReg::add(in, -1);
658 }
659 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecZ);
660 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack");
661 *idealreg2spillmask[Op_VecZ] = *idealreg2regmask[Op_VecZ];
662 idealreg2spillmask[Op_VecZ]->OR(aligned_stack_mask);
663 } else {
664 *idealreg2spillmask[Op_VecZ] = RegMask::Empty;
665 }
666
667 if (Matcher::supports_scalable_vector()) {
668 int k = 1;
669 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1);
670 if (Matcher::has_predicated_vectors()) {
671 // Exclude last input arg stack slots to avoid spilling vector register there,
672 // otherwise RegVectMask spills could stomp over stack slots in caller frame.
673 for (; (in >= init_in) && (k < scalable_predicate_reg_slots()); k++) {
674 scalable_stack_mask.Remove(in);
675 in = OptoReg::add(in, -1);
676 }
677
678 // For RegVectMask
679 scalable_stack_mask.clear_to_sets(scalable_predicate_reg_slots());
680 assert(scalable_stack_mask.is_AllStack(), "should be infinite stack");
681 *idealreg2spillmask[Op_RegVectMask] = *idealreg2regmask[Op_RegVectMask];
682 idealreg2spillmask[Op_RegVectMask]->OR(scalable_stack_mask);
683 }
684
685 // Exclude last input arg stack slots to avoid spilling vector register there,
686 // otherwise vector spills could stomp over stack slots in caller frame.
687 for (; (in >= init_in) && (k < scalable_vector_reg_size(T_FLOAT)); k++) {
688 scalable_stack_mask.Remove(in);
689 in = OptoReg::add(in, -1);
690 }
691
692 // For VecA
693 scalable_stack_mask.clear_to_sets(RegMask::SlotsPerVecA);
694 assert(scalable_stack_mask.is_AllStack(), "should be infinite stack");
695 *idealreg2spillmask[Op_VecA] = *idealreg2regmask[Op_VecA];
696 idealreg2spillmask[Op_VecA]->OR(scalable_stack_mask);
697 } else {
698 *idealreg2spillmask[Op_VecA] = RegMask::Empty;
699 }
700
701 if (UseFPUForSpilling) {
702 // This mask logic assumes that the spill operations are
703 // symmetric and that the registers involved are the same size.
704 // On sparc for instance we may have to use 64 bit moves will
705 // kill 2 registers when used with F0-F31.
706 idealreg2spillmask[Op_RegI]->OR(*idealreg2regmask[Op_RegF]);
707 idealreg2spillmask[Op_RegF]->OR(*idealreg2regmask[Op_RegI]);
708 #ifdef _LP64
709 idealreg2spillmask[Op_RegN]->OR(*idealreg2regmask[Op_RegF]);
710 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]);
711 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]);
712 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegD]);
713 #else
714 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegF]);
715 #ifdef ARM
716 // ARM has support for moving 64bit values between a pair of
717 // integer registers and a double register
718 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]);
719 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]);
720 #endif
721 #endif
722 }
723
724 // Make up debug masks. Any spill slot plus callee-save (SOE) registers.
725 // Caller-save (SOC, AS) registers are assumed to be trashable by the various
726 // inline-cache fixup routines.
727 *idealreg2debugmask [Op_RegN] = *idealreg2spillmask[Op_RegN];
728 *idealreg2debugmask [Op_RegI] = *idealreg2spillmask[Op_RegI];
729 *idealreg2debugmask [Op_RegL] = *idealreg2spillmask[Op_RegL];
730 *idealreg2debugmask [Op_RegF] = *idealreg2spillmask[Op_RegF];
731 *idealreg2debugmask [Op_RegD] = *idealreg2spillmask[Op_RegD];
732 *idealreg2debugmask [Op_RegP] = *idealreg2spillmask[Op_RegP];
733 *idealreg2debugmask [Op_RegVectMask] = *idealreg2spillmask[Op_RegVectMask];
734
735 *idealreg2debugmask [Op_VecA] = *idealreg2spillmask[Op_VecA];
736 *idealreg2debugmask [Op_VecS] = *idealreg2spillmask[Op_VecS];
737 *idealreg2debugmask [Op_VecD] = *idealreg2spillmask[Op_VecD];
738 *idealreg2debugmask [Op_VecX] = *idealreg2spillmask[Op_VecX];
739 *idealreg2debugmask [Op_VecY] = *idealreg2spillmask[Op_VecY];
740 *idealreg2debugmask [Op_VecZ] = *idealreg2spillmask[Op_VecZ];
741
742 *idealreg2mhdebugmask[Op_RegN] = *idealreg2spillmask[Op_RegN];
743 *idealreg2mhdebugmask[Op_RegI] = *idealreg2spillmask[Op_RegI];
744 *idealreg2mhdebugmask[Op_RegL] = *idealreg2spillmask[Op_RegL];
745 *idealreg2mhdebugmask[Op_RegF] = *idealreg2spillmask[Op_RegF];
746 *idealreg2mhdebugmask[Op_RegD] = *idealreg2spillmask[Op_RegD];
747 *idealreg2mhdebugmask[Op_RegP] = *idealreg2spillmask[Op_RegP];
748 *idealreg2mhdebugmask[Op_RegVectMask] = *idealreg2spillmask[Op_RegVectMask];
749
750 *idealreg2mhdebugmask[Op_VecA] = *idealreg2spillmask[Op_VecA];
751 *idealreg2mhdebugmask[Op_VecS] = *idealreg2spillmask[Op_VecS];
752 *idealreg2mhdebugmask[Op_VecD] = *idealreg2spillmask[Op_VecD];
753 *idealreg2mhdebugmask[Op_VecX] = *idealreg2spillmask[Op_VecX];
754 *idealreg2mhdebugmask[Op_VecY] = *idealreg2spillmask[Op_VecY];
755 *idealreg2mhdebugmask[Op_VecZ] = *idealreg2spillmask[Op_VecZ];
756
757 // Prevent stub compilations from attempting to reference
758 // callee-saved (SOE) registers from debug info
759 bool exclude_soe = !Compile::current()->is_method_compilation();
760 RegMask* caller_save_mask = exclude_soe ? &caller_save_regmask_exclude_soe : &caller_save_regmask;
761 RegMask* mh_caller_save_mask = exclude_soe ? &mh_caller_save_regmask_exclude_soe : &mh_caller_save_regmask;
762
763 idealreg2debugmask[Op_RegN]->SUBTRACT(*caller_save_mask);
764 idealreg2debugmask[Op_RegI]->SUBTRACT(*caller_save_mask);
765 idealreg2debugmask[Op_RegL]->SUBTRACT(*caller_save_mask);
766 idealreg2debugmask[Op_RegF]->SUBTRACT(*caller_save_mask);
767 idealreg2debugmask[Op_RegD]->SUBTRACT(*caller_save_mask);
768 idealreg2debugmask[Op_RegP]->SUBTRACT(*caller_save_mask);
769 idealreg2debugmask[Op_RegVectMask]->SUBTRACT(*caller_save_mask);
770
771 idealreg2debugmask[Op_VecA]->SUBTRACT(*caller_save_mask);
772 idealreg2debugmask[Op_VecS]->SUBTRACT(*caller_save_mask);
773 idealreg2debugmask[Op_VecD]->SUBTRACT(*caller_save_mask);
774 idealreg2debugmask[Op_VecX]->SUBTRACT(*caller_save_mask);
775 idealreg2debugmask[Op_VecY]->SUBTRACT(*caller_save_mask);
776 idealreg2debugmask[Op_VecZ]->SUBTRACT(*caller_save_mask);
777
778 idealreg2mhdebugmask[Op_RegN]->SUBTRACT(*mh_caller_save_mask);
779 idealreg2mhdebugmask[Op_RegI]->SUBTRACT(*mh_caller_save_mask);
780 idealreg2mhdebugmask[Op_RegL]->SUBTRACT(*mh_caller_save_mask);
781 idealreg2mhdebugmask[Op_RegF]->SUBTRACT(*mh_caller_save_mask);
782 idealreg2mhdebugmask[Op_RegD]->SUBTRACT(*mh_caller_save_mask);
783 idealreg2mhdebugmask[Op_RegP]->SUBTRACT(*mh_caller_save_mask);
784 idealreg2mhdebugmask[Op_RegVectMask]->SUBTRACT(*mh_caller_save_mask);
785
786 idealreg2mhdebugmask[Op_VecA]->SUBTRACT(*mh_caller_save_mask);
787 idealreg2mhdebugmask[Op_VecS]->SUBTRACT(*mh_caller_save_mask);
788 idealreg2mhdebugmask[Op_VecD]->SUBTRACT(*mh_caller_save_mask);
789 idealreg2mhdebugmask[Op_VecX]->SUBTRACT(*mh_caller_save_mask);
790 idealreg2mhdebugmask[Op_VecY]->SUBTRACT(*mh_caller_save_mask);
791 idealreg2mhdebugmask[Op_VecZ]->SUBTRACT(*mh_caller_save_mask);
792 }
793
794 //---------------------------is_save_on_entry----------------------------------
795 bool Matcher::is_save_on_entry(int reg) {
796 return
797 _register_save_policy[reg] == 'E' ||
798 _register_save_policy[reg] == 'A'; // Save-on-entry register?
799 }
800
801 //---------------------------Fixup_Save_On_Entry-------------------------------
802 void Matcher::Fixup_Save_On_Entry( ) {
803 init_first_stack_mask();
804
805 Node *root = C->root(); // Short name for root
806 // Count number of save-on-entry registers.
807 uint soe_cnt = number_of_saved_registers();
808 uint i;
809
810 // Find the procedure Start Node
811 StartNode *start = C->start();
812 assert( start, "Expect a start node" );
813
814 // Input RegMask array shared by all Returns.
815 // The type for doubles and longs has a count of 2, but
816 // there is only 1 returned value
817 uint ret_edge_cnt = C->tf()->range_cc()->cnt();
818 RegMask *ret_rms = init_input_masks( ret_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
819 for (i = TypeFunc::Parms; i < ret_edge_cnt; i++) {
820 ret_rms[i] = _return_values_mask[i-TypeFunc::Parms];
821 }
822
823 // Input RegMask array shared by all Rethrows.
824 uint reth_edge_cnt = TypeFunc::Parms+1;
825 RegMask *reth_rms = init_input_masks( reth_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
826 // Rethrow takes exception oop only, but in the argument 0 slot.
827 OptoReg::Name reg = find_receiver();
828 if (reg >= 0) {
829 reth_rms[TypeFunc::Parms] = mreg2regmask[reg];
830 #ifdef _LP64
831 // Need two slots for ptrs in 64-bit land
832 reth_rms[TypeFunc::Parms].Insert(OptoReg::add(OptoReg::Name(reg), 1));
833 #endif
834 }
835
836 // Input RegMask array shared by all TailCalls
837 uint tail_call_edge_cnt = TypeFunc::Parms+2;
838 RegMask *tail_call_rms = init_input_masks( tail_call_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
839
840 // Input RegMask array shared by all TailJumps
841 uint tail_jump_edge_cnt = TypeFunc::Parms+2;
842 RegMask *tail_jump_rms = init_input_masks( tail_jump_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
843
844 // TailCalls have 2 returned values (target & moop), whose masks come
845 // from the usual MachNode/MachOper mechanism. Find a sample
846 // TailCall to extract these masks and put the correct masks into
847 // the tail_call_rms array.
848 for( i=1; i < root->req(); i++ ) {
849 MachReturnNode *m = root->in(i)->as_MachReturn();
850 if( m->ideal_Opcode() == Op_TailCall ) {
851 tail_call_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0);
852 tail_call_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1);
853 break;
854 }
855 }
856
857 // TailJumps have 2 returned values (target & ex_oop), whose masks come
858 // from the usual MachNode/MachOper mechanism. Find a sample
859 // TailJump to extract these masks and put the correct masks into
860 // the tail_jump_rms array.
861 for( i=1; i < root->req(); i++ ) {
862 MachReturnNode *m = root->in(i)->as_MachReturn();
863 if( m->ideal_Opcode() == Op_TailJump ) {
864 tail_jump_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0);
865 tail_jump_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1);
866 break;
867 }
868 }
869
870 // Input RegMask array shared by all Halts
871 uint halt_edge_cnt = TypeFunc::Parms;
872 RegMask *halt_rms = init_input_masks( halt_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
873
874 // Capture the return input masks into each exit flavor
875 for( i=1; i < root->req(); i++ ) {
876 MachReturnNode *exit = root->in(i)->as_MachReturn();
877 switch( exit->ideal_Opcode() ) {
878 case Op_Return : exit->_in_rms = ret_rms; break;
879 case Op_Rethrow : exit->_in_rms = reth_rms; break;
880 case Op_TailCall : exit->_in_rms = tail_call_rms; break;
881 case Op_TailJump : exit->_in_rms = tail_jump_rms; break;
882 case Op_Halt : exit->_in_rms = halt_rms; break;
883 default : ShouldNotReachHere();
884 }
885 }
886
887 // Next unused projection number from Start.
888 int proj_cnt = C->tf()->domain_cc()->cnt();
889
890 // Do all the save-on-entry registers. Make projections from Start for
891 // them, and give them a use at the exit points. To the allocator, they
892 // look like incoming register arguments.
893 for( i = 0; i < _last_Mach_Reg; i++ ) {
894 if( is_save_on_entry(i) ) {
895
896 // Add the save-on-entry to the mask array
897 ret_rms [ ret_edge_cnt] = mreg2regmask[i];
898 reth_rms [ reth_edge_cnt] = mreg2regmask[i];
899 tail_call_rms[tail_call_edge_cnt] = mreg2regmask[i];
900 tail_jump_rms[tail_jump_edge_cnt] = mreg2regmask[i];
901 // Halts need the SOE registers, but only in the stack as debug info.
902 // A just-prior uncommon-trap or deoptimization will use the SOE regs.
903 halt_rms [ halt_edge_cnt] = *idealreg2spillmask[_register_save_type[i]];
904
905 Node *mproj;
906
907 // Is this a RegF low half of a RegD? Double up 2 adjacent RegF's
908 // into a single RegD.
909 if( (i&1) == 0 &&
910 _register_save_type[i ] == Op_RegF &&
911 _register_save_type[i+1] == Op_RegF &&
912 is_save_on_entry(i+1) ) {
913 // Add other bit for double
914 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1));
915 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1));
916 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1));
917 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1));
918 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1));
919 mproj = new MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegD );
920 proj_cnt += 2; // Skip 2 for doubles
921 }
922 else if( (i&1) == 1 && // Else check for high half of double
923 _register_save_type[i-1] == Op_RegF &&
924 _register_save_type[i ] == Op_RegF &&
925 is_save_on_entry(i-1) ) {
926 ret_rms [ ret_edge_cnt] = RegMask::Empty;
927 reth_rms [ reth_edge_cnt] = RegMask::Empty;
928 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty;
929 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty;
930 halt_rms [ halt_edge_cnt] = RegMask::Empty;
931 mproj = C->top();
932 }
933 // Is this a RegI low half of a RegL? Double up 2 adjacent RegI's
934 // into a single RegL.
935 else if( (i&1) == 0 &&
936 _register_save_type[i ] == Op_RegI &&
937 _register_save_type[i+1] == Op_RegI &&
938 is_save_on_entry(i+1) ) {
939 // Add other bit for long
940 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1));
941 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1));
942 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1));
943 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1));
944 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1));
945 mproj = new MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegL );
946 proj_cnt += 2; // Skip 2 for longs
947 }
948 else if( (i&1) == 1 && // Else check for high half of long
949 _register_save_type[i-1] == Op_RegI &&
950 _register_save_type[i ] == Op_RegI &&
951 is_save_on_entry(i-1) ) {
952 ret_rms [ ret_edge_cnt] = RegMask::Empty;
953 reth_rms [ reth_edge_cnt] = RegMask::Empty;
954 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty;
955 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty;
956 halt_rms [ halt_edge_cnt] = RegMask::Empty;
957 mproj = C->top();
958 } else {
959 // Make a projection for it off the Start
960 mproj = new MachProjNode( start, proj_cnt++, ret_rms[ret_edge_cnt], _register_save_type[i] );
961 }
962
963 ret_edge_cnt ++;
964 reth_edge_cnt ++;
965 tail_call_edge_cnt ++;
966 tail_jump_edge_cnt ++;
967 halt_edge_cnt ++;
968
969 // Add a use of the SOE register to all exit paths
970 for( uint j=1; j < root->req(); j++ )
971 root->in(j)->add_req(mproj);
972 } // End of if a save-on-entry register
973 } // End of for all machine registers
974 }
975
976 //------------------------------init_spill_mask--------------------------------
977 void Matcher::init_spill_mask( Node *ret ) {
978 if( idealreg2regmask[Op_RegI] ) return; // One time only init
979
980 OptoReg::c_frame_pointer = c_frame_pointer();
981 c_frame_ptr_mask = c_frame_pointer();
982 #ifdef _LP64
983 // pointers are twice as big
984 c_frame_ptr_mask.Insert(OptoReg::add(c_frame_pointer(),1));
985 #endif
986
987 // Start at OptoReg::stack0()
988 STACK_ONLY_mask.Clear();
989 OptoReg::Name init = OptoReg::stack2reg(0);
990 // STACK_ONLY_mask is all stack bits
991 OptoReg::Name i;
992 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1))
993 STACK_ONLY_mask.Insert(i);
994 // Also set the "infinite stack" bit.
995 STACK_ONLY_mask.set_AllStack();
996
997 for (i = OptoReg::Name(0); i < OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i, 1)) {
998 // Copy the register names over into the shared world.
999 // SharedInfo::regName[i] = regName[i];
1000 // Handy RegMasks per machine register
1001 mreg2regmask[i].Insert(i);
1002
1003 // Set up regmasks used to exclude save-on-call (and always-save) registers from debug masks.
1004 if (_register_save_policy[i] == 'C' ||
1005 _register_save_policy[i] == 'A') {
1006 caller_save_regmask.Insert(i);
1007 mh_caller_save_regmask.Insert(i);
1008 }
1009 // Exclude save-on-entry registers from debug masks for stub compilations.
1010 if (_register_save_policy[i] == 'C' ||
1011 _register_save_policy[i] == 'A' ||
1012 _register_save_policy[i] == 'E') {
1013 caller_save_regmask_exclude_soe.Insert(i);
1014 mh_caller_save_regmask_exclude_soe.Insert(i);
1015 }
1016 }
1017
1018 // Also exclude the register we use to save the SP for MethodHandle
1019 // invokes to from the corresponding MH debug masks
1020 const RegMask sp_save_mask = method_handle_invoke_SP_save_mask();
1021 mh_caller_save_regmask.OR(sp_save_mask);
1022 mh_caller_save_regmask_exclude_soe.OR(sp_save_mask);
1023
1024 // Grab the Frame Pointer
1025 Node *fp = ret->in(TypeFunc::FramePtr);
1026 // Share frame pointer while making spill ops
1027 set_shared(fp);
1028
1029 // Get the ADLC notion of the right regmask, for each basic type.
1030 #ifdef _LP64
1031 idealreg2regmask[Op_RegN] = regmask_for_ideal_register(Op_RegN, ret);
1032 #endif
1033 idealreg2regmask[Op_RegI] = regmask_for_ideal_register(Op_RegI, ret);
1034 idealreg2regmask[Op_RegP] = regmask_for_ideal_register(Op_RegP, ret);
1035 idealreg2regmask[Op_RegF] = regmask_for_ideal_register(Op_RegF, ret);
1036 idealreg2regmask[Op_RegD] = regmask_for_ideal_register(Op_RegD, ret);
1037 idealreg2regmask[Op_RegL] = regmask_for_ideal_register(Op_RegL, ret);
1038 idealreg2regmask[Op_VecA] = regmask_for_ideal_register(Op_VecA, ret);
1039 idealreg2regmask[Op_VecS] = regmask_for_ideal_register(Op_VecS, ret);
1040 idealreg2regmask[Op_VecD] = regmask_for_ideal_register(Op_VecD, ret);
1041 idealreg2regmask[Op_VecX] = regmask_for_ideal_register(Op_VecX, ret);
1042 idealreg2regmask[Op_VecY] = regmask_for_ideal_register(Op_VecY, ret);
1043 idealreg2regmask[Op_VecZ] = regmask_for_ideal_register(Op_VecZ, ret);
1044 idealreg2regmask[Op_RegVectMask] = regmask_for_ideal_register(Op_RegVectMask, ret);
1045 }
1046
1047 #ifdef ASSERT
1048 static void match_alias_type(Compile* C, Node* n, Node* m) {
1049 if (!VerifyAliases) return; // do not go looking for trouble by default
1050 const TypePtr* nat = n->adr_type();
1051 const TypePtr* mat = m->adr_type();
1052 int nidx = C->get_alias_index(nat);
1053 int midx = C->get_alias_index(mat);
1054 // Detune the assert for cases like (AndI 0xFF (LoadB p)).
1055 if (nidx == Compile::AliasIdxTop && midx >= Compile::AliasIdxRaw) {
1056 for (uint i = 1; i < n->req(); i++) {
1057 Node* n1 = n->in(i);
1058 const TypePtr* n1at = n1->adr_type();
1059 if (n1at != NULL) {
1060 nat = n1at;
1061 nidx = C->get_alias_index(n1at);
1062 }
1063 }
1064 }
1065 // %%% Kludgery. Instead, fix ideal adr_type methods for all these cases:
1066 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxRaw) {
1067 switch (n->Opcode()) {
1068 case Op_PrefetchAllocation:
1069 nidx = Compile::AliasIdxRaw;
1070 nat = TypeRawPtr::BOTTOM;
1071 break;
1072 }
1073 }
1074 if (nidx == Compile::AliasIdxRaw && midx == Compile::AliasIdxTop) {
1075 switch (n->Opcode()) {
1076 case Op_ClearArray:
1077 midx = Compile::AliasIdxRaw;
1078 mat = TypeRawPtr::BOTTOM;
1079 break;
1080 }
1081 }
1082 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxBot) {
1083 switch (n->Opcode()) {
1084 case Op_Return:
1085 case Op_Rethrow:
1086 case Op_Halt:
1087 case Op_TailCall:
1088 case Op_TailJump:
1089 nidx = Compile::AliasIdxBot;
1090 nat = TypePtr::BOTTOM;
1091 break;
1092 }
1093 }
1094 if (nidx == Compile::AliasIdxBot && midx == Compile::AliasIdxTop) {
1095 switch (n->Opcode()) {
1096 case Op_StrComp:
1097 case Op_StrEquals:
1098 case Op_StrIndexOf:
1099 case Op_StrIndexOfChar:
1100 case Op_AryEq:
1101 case Op_CountPositives:
1102 case Op_MemBarVolatile:
1103 case Op_MemBarCPUOrder: // %%% these ideals should have narrower adr_type?
1104 case Op_StrInflatedCopy:
1105 case Op_StrCompressedCopy:
1106 case Op_OnSpinWait:
1107 case Op_EncodeISOArray:
1108 nidx = Compile::AliasIdxTop;
1109 nat = NULL;
1110 break;
1111 }
1112 }
1113 if (nidx != midx) {
1114 if (PrintOpto || (PrintMiscellaneous && (WizardMode || Verbose))) {
1115 tty->print_cr("==== Matcher alias shift %d => %d", nidx, midx);
1116 n->dump();
1117 m->dump();
1118 }
1119 assert(C->subsume_loads() && C->must_alias(nat, midx),
1120 "must not lose alias info when matching");
1121 }
1122 }
1123 #endif
1124
1125 //------------------------------xform------------------------------------------
1126 // Given a Node in old-space, Match him (Label/Reduce) to produce a machine
1127 // Node in new-space. Given a new-space Node, recursively walk his children.
1128 Node *Matcher::transform( Node *n ) { ShouldNotCallThis(); return n; }
1129 Node *Matcher::xform( Node *n, int max_stack ) {
1130 // Use one stack to keep both: child's node/state and parent's node/index
1131 MStack mstack(max_stack * 2 * 2); // usually: C->live_nodes() * 2 * 2
1132 mstack.push(n, Visit, NULL, -1); // set NULL as parent to indicate root
1133 while (mstack.is_nonempty()) {
1134 C->check_node_count(NodeLimitFudgeFactor, "too many nodes matching instructions");
1135 if (C->failing()) return NULL;
1136 n = mstack.node(); // Leave node on stack
1137 Node_State nstate = mstack.state();
1138 if (nstate == Visit) {
1139 mstack.set_state(Post_Visit);
1140 Node *oldn = n;
1141 // Old-space or new-space check
1142 if (!C->node_arena()->contains(n)) {
1143 // Old space!
1144 Node* m;
1145 if (has_new_node(n)) { // Not yet Label/Reduced
1146 m = new_node(n);
1147 } else {
1148 if (!is_dontcare(n)) { // Matcher can match this guy
1149 // Calls match special. They match alone with no children.
1150 // Their children, the incoming arguments, match normally.
1151 m = n->is_SafePoint() ? match_sfpt(n->as_SafePoint()):match_tree(n);
1152 if (C->failing()) return NULL;
1153 if (m == NULL) { Matcher::soft_match_failure(); return NULL; }
1154 if (n->is_MemBar()) {
1155 m->as_MachMemBar()->set_adr_type(n->adr_type());
1156 }
1157 } else { // Nothing the matcher cares about
1158 if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Multi()) { // Projections?
1159 // Convert to machine-dependent projection
1160 RegMask* mask = NULL;
1161 if (n->in(0)->is_Call() && n->in(0)->as_Call()->tf()->returns_inline_type_as_fields()) {
1162 mask = return_values_mask(n->in(0)->as_Call()->tf());
1163 }
1164 m = n->in(0)->as_Multi()->match(n->as_Proj(), this, mask);
1165 NOT_PRODUCT(record_new2old(m, n);)
1166 if (m->in(0) != NULL) // m might be top
1167 collect_null_checks(m, n);
1168 } else { // Else just a regular 'ol guy
1169 m = n->clone(); // So just clone into new-space
1170 NOT_PRODUCT(record_new2old(m, n);)
1171 // Def-Use edges will be added incrementally as Uses
1172 // of this node are matched.
1173 assert(m->outcnt() == 0, "no Uses of this clone yet");
1174 }
1175 }
1176
1177 set_new_node(n, m); // Map old to new
1178 if (_old_node_note_array != NULL) {
1179 Node_Notes* nn = C->locate_node_notes(_old_node_note_array,
1180 n->_idx);
1181 C->set_node_notes_at(m->_idx, nn);
1182 }
1183 debug_only(match_alias_type(C, n, m));
1184 }
1185 n = m; // n is now a new-space node
1186 mstack.set_node(n);
1187 }
1188
1189 // New space!
1190 if (_visited.test_set(n->_idx)) continue; // while(mstack.is_nonempty())
1191
1192 int i;
1193 // Put precedence edges on stack first (match them last).
1194 for (i = oldn->req(); (uint)i < oldn->len(); i++) {
1195 Node *m = oldn->in(i);
1196 if (m == NULL) break;
1197 // set -1 to call add_prec() instead of set_req() during Step1
1198 mstack.push(m, Visit, n, -1);
1199 }
1200
1201 // Handle precedence edges for interior nodes
1202 for (i = n->len()-1; (uint)i >= n->req(); i--) {
1203 Node *m = n->in(i);
1204 if (m == NULL || C->node_arena()->contains(m)) continue;
1205 n->rm_prec(i);
1206 // set -1 to call add_prec() instead of set_req() during Step1
1207 mstack.push(m, Visit, n, -1);
1208 }
1209
1210 // For constant debug info, I'd rather have unmatched constants.
1211 int cnt = n->req();
1212 JVMState* jvms = n->jvms();
1213 int debug_cnt = jvms ? jvms->debug_start() : cnt;
1214
1215 // Now do only debug info. Clone constants rather than matching.
1216 // Constants are represented directly in the debug info without
1217 // the need for executable machine instructions.
1218 // Monitor boxes are also represented directly.
1219 for (i = cnt - 1; i >= debug_cnt; --i) { // For all debug inputs do
1220 Node *m = n->in(i); // Get input
1221 int op = m->Opcode();
1222 assert((op == Op_BoxLock) == jvms->is_monitor_use(i), "boxes only at monitor sites");
1223 if( op == Op_ConI || op == Op_ConP || op == Op_ConN || op == Op_ConNKlass ||
1224 op == Op_ConF || op == Op_ConD || op == Op_ConL
1225 // || op == Op_BoxLock // %%%% enable this and remove (+++) in chaitin.cpp
1226 ) {
1227 m = m->clone();
1228 NOT_PRODUCT(record_new2old(m, n));
1229 mstack.push(m, Post_Visit, n, i); // Don't need to visit
1230 mstack.push(m->in(0), Visit, m, 0);
1231 } else {
1232 mstack.push(m, Visit, n, i);
1233 }
1234 }
1235
1236 // And now walk his children, and convert his inputs to new-space.
1237 for( ; i >= 0; --i ) { // For all normal inputs do
1238 Node *m = n->in(i); // Get input
1239 if(m != NULL)
1240 mstack.push(m, Visit, n, i);
1241 }
1242
1243 }
1244 else if (nstate == Post_Visit) {
1245 // Set xformed input
1246 Node *p = mstack.parent();
1247 if (p != NULL) { // root doesn't have parent
1248 int i = (int)mstack.index();
1249 if (i >= 0)
1250 p->set_req(i, n); // required input
1251 else if (i == -1)
1252 p->add_prec(n); // precedence input
1253 else
1254 ShouldNotReachHere();
1255 }
1256 mstack.pop(); // remove processed node from stack
1257 }
1258 else {
1259 ShouldNotReachHere();
1260 }
1261 } // while (mstack.is_nonempty())
1262 return n; // Return new-space Node
1263 }
1264
1265 //------------------------------warp_outgoing_stk_arg------------------------
1266 OptoReg::Name Matcher::warp_outgoing_stk_arg( VMReg reg, OptoReg::Name begin_out_arg_area, OptoReg::Name &out_arg_limit_per_call ) {
1267 // Convert outgoing argument location to a pre-biased stack offset
1268 if (reg->is_stack()) {
1269 OptoReg::Name warped = reg->reg2stack();
1270 // Adjust the stack slot offset to be the register number used
1271 // by the allocator.
1272 warped = OptoReg::add(begin_out_arg_area, warped);
1273 // Keep track of the largest numbered stack slot used for an arg.
1274 // Largest used slot per call-site indicates the amount of stack
1275 // that is killed by the call.
1276 if( warped >= out_arg_limit_per_call )
1277 out_arg_limit_per_call = OptoReg::add(warped,1);
1278 if (!RegMask::can_represent_arg(warped)) {
1279 C->record_method_not_compilable("unsupported calling sequence");
1280 return OptoReg::Bad;
1281 }
1282 return warped;
1283 }
1284 return OptoReg::as_OptoReg(reg);
1285 }
1286
1287
1288 //------------------------------match_sfpt-------------------------------------
1289 // Helper function to match call instructions. Calls match special.
1290 // They match alone with no children. Their children, the incoming
1291 // arguments, match normally.
1292 MachNode *Matcher::match_sfpt( SafePointNode *sfpt ) {
1293 MachSafePointNode *msfpt = NULL;
1294 MachCallNode *mcall = NULL;
1295 uint cnt;
1296 // Split out case for SafePoint vs Call
1297 CallNode *call;
1298 const TypeTuple *domain;
1299 ciMethod* method = NULL;
1300 bool is_method_handle_invoke = false; // for special kill effects
1301 if( sfpt->is_Call() ) {
1302 call = sfpt->as_Call();
1303 domain = call->tf()->domain_cc();
1304 cnt = domain->cnt();
1305
1306 // Match just the call, nothing else
1307 MachNode *m = match_tree(call);
1308 if (C->failing()) return NULL;
1309 if( m == NULL ) { Matcher::soft_match_failure(); return NULL; }
1310
1311 // Copy data from the Ideal SafePoint to the machine version
1312 mcall = m->as_MachCall();
1313
1314 mcall->set_tf( call->tf());
1315 mcall->set_entry_point( call->entry_point());
1316 mcall->set_cnt( call->cnt());
1317 mcall->set_guaranteed_safepoint(call->guaranteed_safepoint());
1318
1319 if( mcall->is_MachCallJava() ) {
1320 MachCallJavaNode *mcall_java = mcall->as_MachCallJava();
1321 const CallJavaNode *call_java = call->as_CallJava();
1322 assert(call_java->validate_symbolic_info(), "inconsistent info");
1323 method = call_java->method();
1324 mcall_java->_method = method;
1325 mcall_java->_optimized_virtual = call_java->is_optimized_virtual();
1326 is_method_handle_invoke = call_java->is_method_handle_invoke();
1327 mcall_java->_method_handle_invoke = is_method_handle_invoke;
1328 mcall_java->_override_symbolic_info = call_java->override_symbolic_info();
1329 mcall_java->_arg_escape = call_java->arg_escape();
1330 if (is_method_handle_invoke) {
1331 C->set_has_method_handle_invokes(true);
1332 }
1333 if( mcall_java->is_MachCallStaticJava() )
1334 mcall_java->as_MachCallStaticJava()->_name =
1335 call_java->as_CallStaticJava()->_name;
1336 if( mcall_java->is_MachCallDynamicJava() )
1337 mcall_java->as_MachCallDynamicJava()->_vtable_index =
1338 call_java->as_CallDynamicJava()->_vtable_index;
1339 }
1340 else if( mcall->is_MachCallRuntime() ) {
1341 MachCallRuntimeNode* mach_call_rt = mcall->as_MachCallRuntime();
1342 mach_call_rt->_name = call->as_CallRuntime()->_name;
1343 mach_call_rt->_leaf_no_fp = call->is_CallLeafNoFP();
1344 }
1345 msfpt = mcall;
1346 }
1347 // This is a non-call safepoint
1348 else {
1349 call = NULL;
1350 domain = NULL;
1351 MachNode *mn = match_tree(sfpt);
1352 if (C->failing()) return NULL;
1353 msfpt = mn->as_MachSafePoint();
1354 cnt = TypeFunc::Parms;
1355 }
1356 msfpt->_has_ea_local_in_scope = sfpt->has_ea_local_in_scope();
1357
1358 // Advertise the correct memory effects (for anti-dependence computation).
1359 msfpt->set_adr_type(sfpt->adr_type());
1360
1361 // Allocate a private array of RegMasks. These RegMasks are not shared.
1362 msfpt->_in_rms = NEW_RESOURCE_ARRAY( RegMask, cnt );
1363 // Empty them all.
1364 for (uint i = 0; i < cnt; i++) ::new (&(msfpt->_in_rms[i])) RegMask();
1365
1366 // Do all the pre-defined non-Empty register masks
1367 msfpt->_in_rms[TypeFunc::ReturnAdr] = _return_addr_mask;
1368 msfpt->_in_rms[TypeFunc::FramePtr ] = c_frame_ptr_mask;
1369
1370 // Place first outgoing argument can possibly be put.
1371 OptoReg::Name begin_out_arg_area = OptoReg::add(_new_SP, C->out_preserve_stack_slots());
1372 assert( is_even(begin_out_arg_area), "" );
1373 // Compute max outgoing register number per call site.
1374 OptoReg::Name out_arg_limit_per_call = begin_out_arg_area;
1375 // Calls to C may hammer extra stack slots above and beyond any arguments.
1376 // These are usually backing store for register arguments for varargs.
1377 if( call != NULL && call->is_CallRuntime() )
1378 out_arg_limit_per_call = OptoReg::add(out_arg_limit_per_call,C->varargs_C_out_slots_killed());
1379
1380
1381 // Do the normal argument list (parameters) register masks
1382 // Null entry point is a special cast where the target of the call
1383 // is in a register.
1384 int adj = (call != NULL && call->entry_point() == NULL) ? 1 : 0;
1385 int argcnt = cnt - TypeFunc::Parms - adj;
1386 if( argcnt > 0 ) { // Skip it all if we have no args
1387 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt );
1388 VMRegPair *parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt );
1389 int i;
1390 for( i = 0; i < argcnt; i++ ) {
1391 sig_bt[i] = domain->field_at(i+TypeFunc::Parms+adj)->basic_type();
1392 }
1393 // V-call to pick proper calling convention
1394 call->calling_convention( sig_bt, parm_regs, argcnt );
1395
1396 #ifdef ASSERT
1397 // Sanity check users' calling convention. Really handy during
1398 // the initial porting effort. Fairly expensive otherwise.
1399 { for (int i = 0; i<argcnt; i++) {
1400 if( !parm_regs[i].first()->is_valid() &&
1401 !parm_regs[i].second()->is_valid() ) continue;
1402 VMReg reg1 = parm_regs[i].first();
1403 VMReg reg2 = parm_regs[i].second();
1404 for (int j = 0; j < i; j++) {
1405 if( !parm_regs[j].first()->is_valid() &&
1406 !parm_regs[j].second()->is_valid() ) continue;
1407 VMReg reg3 = parm_regs[j].first();
1408 VMReg reg4 = parm_regs[j].second();
1409 if( !reg1->is_valid() ) {
1410 assert( !reg2->is_valid(), "valid halvsies" );
1411 } else if( !reg3->is_valid() ) {
1412 assert( !reg4->is_valid(), "valid halvsies" );
1413 } else {
1414 assert( reg1 != reg2, "calling conv. must produce distinct regs");
1415 assert( reg1 != reg3, "calling conv. must produce distinct regs");
1416 assert( reg1 != reg4, "calling conv. must produce distinct regs");
1417 assert( reg2 != reg3, "calling conv. must produce distinct regs");
1418 assert( reg2 != reg4 || !reg2->is_valid(), "calling conv. must produce distinct regs");
1419 assert( reg3 != reg4, "calling conv. must produce distinct regs");
1420 }
1421 }
1422 }
1423 }
1424 #endif
1425
1426 // Visit each argument. Compute its outgoing register mask.
1427 // Return results now can have 2 bits returned.
1428 // Compute max over all outgoing arguments both per call-site
1429 // and over the entire method.
1430 for( i = 0; i < argcnt; i++ ) {
1431 // Address of incoming argument mask to fill in
1432 RegMask *rm = &mcall->_in_rms[i+TypeFunc::Parms+adj];
1433 VMReg first = parm_regs[i].first();
1434 VMReg second = parm_regs[i].second();
1435 if(!first->is_valid() &&
1436 !second->is_valid()) {
1437 continue; // Avoid Halves
1438 }
1439 // Handle case where arguments are in vector registers.
1440 if(call->in(TypeFunc::Parms + i)->bottom_type()->isa_vect()) {
1441 OptoReg::Name reg_fst = OptoReg::as_OptoReg(first);
1442 OptoReg::Name reg_snd = OptoReg::as_OptoReg(second);
1443 assert (reg_fst <= reg_snd, "fst=%d snd=%d", reg_fst, reg_snd);
1444 for (OptoReg::Name r = reg_fst; r <= reg_snd; r++) {
1445 rm->Insert(r);
1446 }
1447 }
1448 // Grab first register, adjust stack slots and insert in mask.
1449 OptoReg::Name reg1 = warp_outgoing_stk_arg(first, begin_out_arg_area, out_arg_limit_per_call );
1450 if (OptoReg::is_valid(reg1)) {
1451 rm->Insert( reg1 );
1452 }
1453 // Grab second register (if any), adjust stack slots and insert in mask.
1454 OptoReg::Name reg2 = warp_outgoing_stk_arg(second, begin_out_arg_area, out_arg_limit_per_call );
1455 if (OptoReg::is_valid(reg2)) {
1456 rm->Insert( reg2 );
1457 }
1458 } // End of for all arguments
1459 }
1460
1461 // Compute the max stack slot killed by any call. These will not be
1462 // available for debug info, and will be used to adjust FIRST_STACK_mask
1463 // after all call sites have been visited.
1464 if( _out_arg_limit < out_arg_limit_per_call)
1465 _out_arg_limit = out_arg_limit_per_call;
1466
1467 if (mcall) {
1468 // Kill the outgoing argument area, including any non-argument holes and
1469 // any legacy C-killed slots. Use Fat-Projections to do the killing.
1470 // Since the max-per-method covers the max-per-call-site and debug info
1471 // is excluded on the max-per-method basis, debug info cannot land in
1472 // this killed area.
1473 uint r_cnt = mcall->tf()->range_sig()->cnt();
1474 MachProjNode *proj = new MachProjNode( mcall, r_cnt+10000, RegMask::Empty, MachProjNode::fat_proj );
1475 if (!RegMask::can_represent_arg(OptoReg::Name(out_arg_limit_per_call-1))) {
1476 C->record_method_not_compilable("unsupported outgoing calling sequence");
1477 } else {
1478 for (int i = begin_out_arg_area; i < out_arg_limit_per_call; i++)
1479 proj->_rout.Insert(OptoReg::Name(i));
1480 }
1481 if (proj->_rout.is_NotEmpty()) {
1482 push_projection(proj);
1483 }
1484 }
1485 // Transfer the safepoint information from the call to the mcall
1486 // Move the JVMState list
1487 msfpt->set_jvms(sfpt->jvms());
1488 for (JVMState* jvms = msfpt->jvms(); jvms; jvms = jvms->caller()) {
1489 jvms->set_map(sfpt);
1490 }
1491
1492 // Debug inputs begin just after the last incoming parameter
1493 assert((mcall == NULL) || (mcall->jvms() == NULL) ||
1494 (mcall->jvms()->debug_start() + mcall->_jvmadj == mcall->tf()->domain_cc()->cnt()), "");
1495
1496 // Add additional edges.
1497 if (msfpt->mach_constant_base_node_input() != (uint)-1 && !msfpt->is_MachCallLeaf()) {
1498 // For these calls we can not add MachConstantBase in expand(), as the
1499 // ins are not complete then.
1500 msfpt->ins_req(msfpt->mach_constant_base_node_input(), C->mach_constant_base_node());
1501 if (msfpt->jvms() &&
1502 msfpt->mach_constant_base_node_input() <= msfpt->jvms()->debug_start() + msfpt->_jvmadj) {
1503 // We added an edge before jvms, so we must adapt the position of the ins.
1504 msfpt->jvms()->adapt_position(+1);
1505 }
1506 }
1507
1508 // Registers killed by the call are set in the local scheduling pass
1509 // of Global Code Motion.
1510 return msfpt;
1511 }
1512
1513 //---------------------------match_tree----------------------------------------
1514 // Match a Ideal Node DAG - turn it into a tree; Label & Reduce. Used as part
1515 // of the whole-sale conversion from Ideal to Mach Nodes. Also used for
1516 // making GotoNodes while building the CFG and in init_spill_mask() to identify
1517 // a Load's result RegMask for memoization in idealreg2regmask[]
1518 MachNode *Matcher::match_tree( const Node *n ) {
1519 assert( n->Opcode() != Op_Phi, "cannot match" );
1520 assert( !n->is_block_start(), "cannot match" );
1521 // Set the mark for all locally allocated State objects.
1522 // When this call returns, the _states_arena arena will be reset
1523 // freeing all State objects.
1524 ResourceMark rm( &_states_arena );
1525
1526 LabelRootDepth = 0;
1527
1528 // StoreNodes require their Memory input to match any LoadNodes
1529 Node *mem = n->is_Store() ? n->in(MemNode::Memory) : (Node*)1 ;
1530 #ifdef ASSERT
1531 Node* save_mem_node = _mem_node;
1532 _mem_node = n->is_Store() ? (Node*)n : NULL;
1533 #endif
1534 // State object for root node of match tree
1535 // Allocate it on _states_arena - stack allocation can cause stack overflow.
1536 State *s = new (&_states_arena) State;
1537 s->_kids[0] = NULL;
1538 s->_kids[1] = NULL;
1539 s->_leaf = (Node*)n;
1540 // Label the input tree, allocating labels from top-level arena
1541 Node* root_mem = mem;
1542 Label_Root(n, s, n->in(0), root_mem);
1543 if (C->failing()) return NULL;
1544
1545 // The minimum cost match for the whole tree is found at the root State
1546 uint mincost = max_juint;
1547 uint cost = max_juint;
1548 uint i;
1549 for (i = 0; i < NUM_OPERANDS; i++) {
1550 if (s->valid(i) && // valid entry and
1551 s->cost(i) < cost && // low cost and
1552 s->rule(i) >= NUM_OPERANDS) {// not an operand
1553 mincost = i;
1554 cost = s->cost(i);
1555 }
1556 }
1557 if (mincost == max_juint) {
1558 #ifndef PRODUCT
1559 tty->print("No matching rule for:");
1560 s->dump();
1561 #endif
1562 Matcher::soft_match_failure();
1563 return NULL;
1564 }
1565 // Reduce input tree based upon the state labels to machine Nodes
1566 MachNode *m = ReduceInst(s, s->rule(mincost), mem);
1567 // New-to-old mapping is done in ReduceInst, to cover complex instructions.
1568 NOT_PRODUCT(_old2new_map.map(n->_idx, m);)
1569
1570 // Add any Matcher-ignored edges
1571 uint cnt = n->req();
1572 uint start = 1;
1573 if( mem != (Node*)1 ) start = MemNode::Memory+1;
1574 if( n->is_AddP() ) {
1575 assert( mem == (Node*)1, "" );
1576 start = AddPNode::Base+1;
1577 }
1578 for( i = start; i < cnt; i++ ) {
1579 if( !n->match_edge(i) ) {
1580 if( i < m->req() )
1581 m->ins_req( i, n->in(i) );
1582 else
1583 m->add_req( n->in(i) );
1584 }
1585 }
1586
1587 debug_only( _mem_node = save_mem_node; )
1588 return m;
1589 }
1590
1591
1592 //------------------------------match_into_reg---------------------------------
1593 // Choose to either match this Node in a register or part of the current
1594 // match tree. Return true for requiring a register and false for matching
1595 // as part of the current match tree.
1596 static bool match_into_reg( const Node *n, Node *m, Node *control, int i, bool shared ) {
1597
1598 const Type *t = m->bottom_type();
1599
1600 if (t->singleton()) {
1601 // Never force constants into registers. Allow them to match as
1602 // constants or registers. Copies of the same value will share
1603 // the same register. See find_shared_node.
1604 return false;
1605 } else { // Not a constant
1606 // Stop recursion if they have different Controls.
1607 Node* m_control = m->in(0);
1608 // Control of load's memory can post-dominates load's control.
1609 // So use it since load can't float above its memory.
1610 Node* mem_control = (m->is_Load()) ? m->in(MemNode::Memory)->in(0) : NULL;
1611 if (control && m_control && control != m_control && control != mem_control) {
1612
1613 // Actually, we can live with the most conservative control we
1614 // find, if it post-dominates the others. This allows us to
1615 // pick up load/op/store trees where the load can float a little
1616 // above the store.
1617 Node *x = control;
1618 const uint max_scan = 6; // Arbitrary scan cutoff
1619 uint j;
1620 for (j=0; j<max_scan; j++) {
1621 if (x->is_Region()) // Bail out at merge points
1622 return true;
1623 x = x->in(0);
1624 if (x == m_control) // Does 'control' post-dominate
1625 break; // m->in(0)? If so, we can use it
1626 if (x == mem_control) // Does 'control' post-dominate
1627 break; // mem_control? If so, we can use it
1628 }
1629 if (j == max_scan) // No post-domination before scan end?
1630 return true; // Then break the match tree up
1631 }
1632 if ((m->is_DecodeN() && Matcher::narrow_oop_use_complex_address()) ||
1633 (m->is_DecodeNKlass() && Matcher::narrow_klass_use_complex_address())) {
1634 // These are commonly used in address expressions and can
1635 // efficiently fold into them on X64 in some cases.
1636 return false;
1637 }
1638 }
1639
1640 // Not forceable cloning. If shared, put it into a register.
1641 return shared;
1642 }
1643
1644
1645 //------------------------------Instruction Selection--------------------------
1646 // Label method walks a "tree" of nodes, using the ADLC generated DFA to match
1647 // ideal nodes to machine instructions. Trees are delimited by shared Nodes,
1648 // things the Matcher does not match (e.g., Memory), and things with different
1649 // Controls (hence forced into different blocks). We pass in the Control
1650 // selected for this entire State tree.
1651
1652 // The Matcher works on Trees, but an Intel add-to-memory requires a DAG: the
1653 // Store and the Load must have identical Memories (as well as identical
1654 // pointers). Since the Matcher does not have anything for Memory (and
1655 // does not handle DAGs), I have to match the Memory input myself. If the
1656 // Tree root is a Store or if there are multiple Loads in the tree, I require
1657 // all Loads to have the identical memory.
1658 Node* Matcher::Label_Root(const Node* n, State* svec, Node* control, Node*& mem) {
1659 // Since Label_Root is a recursive function, its possible that we might run
1660 // out of stack space. See bugs 6272980 & 6227033 for more info.
1661 LabelRootDepth++;
1662 if (LabelRootDepth > MaxLabelRootDepth) {
1663 C->record_method_not_compilable("Out of stack space, increase MaxLabelRootDepth");
1664 return NULL;
1665 }
1666 uint care = 0; // Edges matcher cares about
1667 uint cnt = n->req();
1668 uint i = 0;
1669
1670 // Examine children for memory state
1671 // Can only subsume a child into your match-tree if that child's memory state
1672 // is not modified along the path to another input.
1673 // It is unsafe even if the other inputs are separate roots.
1674 Node *input_mem = NULL;
1675 for( i = 1; i < cnt; i++ ) {
1676 if( !n->match_edge(i) ) continue;
1677 Node *m = n->in(i); // Get ith input
1678 assert( m, "expect non-null children" );
1679 if( m->is_Load() ) {
1680 if( input_mem == NULL ) {
1681 input_mem = m->in(MemNode::Memory);
1682 if (mem == (Node*)1) {
1683 // Save this memory to bail out if there's another memory access
1684 // to a different memory location in the same tree.
1685 mem = input_mem;
1686 }
1687 } else if( input_mem != m->in(MemNode::Memory) ) {
1688 input_mem = NodeSentinel;
1689 }
1690 }
1691 }
1692
1693 for( i = 1; i < cnt; i++ ){// For my children
1694 if( !n->match_edge(i) ) continue;
1695 Node *m = n->in(i); // Get ith input
1696 // Allocate states out of a private arena
1697 State *s = new (&_states_arena) State;
1698 svec->_kids[care++] = s;
1699 assert( care <= 2, "binary only for now" );
1700
1701 // Recursively label the State tree.
1702 s->_kids[0] = NULL;
1703 s->_kids[1] = NULL;
1704 s->_leaf = m;
1705
1706 // Check for leaves of the State Tree; things that cannot be a part of
1707 // the current tree. If it finds any, that value is matched as a
1708 // register operand. If not, then the normal matching is used.
1709 if( match_into_reg(n, m, control, i, is_shared(m)) ||
1710 // Stop recursion if this is a LoadNode and there is another memory access
1711 // to a different memory location in the same tree (for example, a StoreNode
1712 // at the root of this tree or another LoadNode in one of the children).
1713 ((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) ||
1714 // Can NOT include the match of a subtree when its memory state
1715 // is used by any of the other subtrees
1716 (input_mem == NodeSentinel) ) {
1717 // Print when we exclude matching due to different memory states at input-loads
1718 if (PrintOpto && (Verbose && WizardMode) && (input_mem == NodeSentinel)
1719 && !((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem)) {
1720 tty->print_cr("invalid input_mem");
1721 }
1722 // Switch to a register-only opcode; this value must be in a register
1723 // and cannot be subsumed as part of a larger instruction.
1724 s->DFA( m->ideal_reg(), m );
1725
1726 } else {
1727 // If match tree has no control and we do, adopt it for entire tree
1728 if( control == NULL && m->in(0) != NULL && m->req() > 1 )
1729 control = m->in(0); // Pick up control
1730 // Else match as a normal part of the match tree.
1731 control = Label_Root(m, s, control, mem);
1732 if (C->failing()) return NULL;
1733 }
1734 }
1735
1736 // Call DFA to match this node, and return
1737 svec->DFA( n->Opcode(), n );
1738
1739 #ifdef ASSERT
1740 uint x;
1741 for( x = 0; x < _LAST_MACH_OPER; x++ )
1742 if( svec->valid(x) )
1743 break;
1744
1745 if (x >= _LAST_MACH_OPER) {
1746 n->dump();
1747 svec->dump();
1748 assert( false, "bad AD file" );
1749 }
1750 #endif
1751 return control;
1752 }
1753
1754
1755 // Con nodes reduced using the same rule can share their MachNode
1756 // which reduces the number of copies of a constant in the final
1757 // program. The register allocator is free to split uses later to
1758 // split live ranges.
1759 MachNode* Matcher::find_shared_node(Node* leaf, uint rule) {
1760 if (!leaf->is_Con() && !leaf->is_DecodeNarrowPtr()) return NULL;
1761
1762 // See if this Con has already been reduced using this rule.
1763 if (_shared_nodes.Size() <= leaf->_idx) return NULL;
1764 MachNode* last = (MachNode*)_shared_nodes.at(leaf->_idx);
1765 if (last != NULL && rule == last->rule()) {
1766 // Don't expect control change for DecodeN
1767 if (leaf->is_DecodeNarrowPtr())
1768 return last;
1769 // Get the new space root.
1770 Node* xroot = new_node(C->root());
1771 if (xroot == NULL) {
1772 // This shouldn't happen give the order of matching.
1773 return NULL;
1774 }
1775
1776 // Shared constants need to have their control be root so they
1777 // can be scheduled properly.
1778 Node* control = last->in(0);
1779 if (control != xroot) {
1780 if (control == NULL || control == C->root()) {
1781 last->set_req(0, xroot);
1782 } else {
1783 assert(false, "unexpected control");
1784 return NULL;
1785 }
1786 }
1787 return last;
1788 }
1789 return NULL;
1790 }
1791
1792
1793 //------------------------------ReduceInst-------------------------------------
1794 // Reduce a State tree (with given Control) into a tree of MachNodes.
1795 // This routine (and it's cohort ReduceOper) convert Ideal Nodes into
1796 // complicated machine Nodes. Each MachNode covers some tree of Ideal Nodes.
1797 // Each MachNode has a number of complicated MachOper operands; each
1798 // MachOper also covers a further tree of Ideal Nodes.
1799
1800 // The root of the Ideal match tree is always an instruction, so we enter
1801 // the recursion here. After building the MachNode, we need to recurse
1802 // the tree checking for these cases:
1803 // (1) Child is an instruction -
1804 // Build the instruction (recursively), add it as an edge.
1805 // Build a simple operand (register) to hold the result of the instruction.
1806 // (2) Child is an interior part of an instruction -
1807 // Skip over it (do nothing)
1808 // (3) Child is the start of a operand -
1809 // Build the operand, place it inside the instruction
1810 // Call ReduceOper.
1811 MachNode *Matcher::ReduceInst( State *s, int rule, Node *&mem ) {
1812 assert( rule >= NUM_OPERANDS, "called with operand rule" );
1813
1814 MachNode* shared_node = find_shared_node(s->_leaf, rule);
1815 if (shared_node != NULL) {
1816 return shared_node;
1817 }
1818
1819 // Build the object to represent this state & prepare for recursive calls
1820 MachNode *mach = s->MachNodeGenerator(rule);
1821 guarantee(mach != NULL, "Missing MachNode");
1822 mach->_opnds[0] = s->MachOperGenerator(_reduceOp[rule]);
1823 assert( mach->_opnds[0] != NULL, "Missing result operand" );
1824 Node *leaf = s->_leaf;
1825 NOT_PRODUCT(record_new2old(mach, leaf);)
1826 // Check for instruction or instruction chain rule
1827 if( rule >= _END_INST_CHAIN_RULE || rule < _BEGIN_INST_CHAIN_RULE ) {
1828 assert(C->node_arena()->contains(s->_leaf) || !has_new_node(s->_leaf),
1829 "duplicating node that's already been matched");
1830 // Instruction
1831 mach->add_req( leaf->in(0) ); // Set initial control
1832 // Reduce interior of complex instruction
1833 ReduceInst_Interior( s, rule, mem, mach, 1 );
1834 } else {
1835 // Instruction chain rules are data-dependent on their inputs
1836 mach->add_req(0); // Set initial control to none
1837 ReduceInst_Chain_Rule( s, rule, mem, mach );
1838 }
1839
1840 // If a Memory was used, insert a Memory edge
1841 if( mem != (Node*)1 ) {
1842 mach->ins_req(MemNode::Memory,mem);
1843 #ifdef ASSERT
1844 // Verify adr type after matching memory operation
1845 const MachOper* oper = mach->memory_operand();
1846 if (oper != NULL && oper != (MachOper*)-1) {
1847 // It has a unique memory operand. Find corresponding ideal mem node.
1848 Node* m = NULL;
1849 if (leaf->is_Mem()) {
1850 m = leaf;
1851 } else {
1852 m = _mem_node;
1853 assert(m != NULL && m->is_Mem(), "expecting memory node");
1854 }
1855 const Type* mach_at = mach->adr_type();
1856 // DecodeN node consumed by an address may have different type
1857 // than its input. Don't compare types for such case.
1858 if (m->adr_type() != mach_at &&
1859 (m->in(MemNode::Address)->is_DecodeNarrowPtr() ||
1860 (m->in(MemNode::Address)->is_AddP() &&
1861 m->in(MemNode::Address)->in(AddPNode::Address)->is_DecodeNarrowPtr()) ||
1862 (m->in(MemNode::Address)->is_AddP() &&
1863 m->in(MemNode::Address)->in(AddPNode::Address)->is_AddP() &&
1864 m->in(MemNode::Address)->in(AddPNode::Address)->in(AddPNode::Address)->is_DecodeNarrowPtr()))) {
1865 mach_at = m->adr_type();
1866 }
1867 if (m->adr_type() != mach_at) {
1868 m->dump();
1869 tty->print_cr("mach:");
1870 mach->dump(1);
1871 }
1872 assert(m->adr_type() == mach_at, "matcher should not change adr type");
1873 }
1874 #endif
1875 }
1876
1877 // If the _leaf is an AddP, insert the base edge
1878 if (leaf->is_AddP()) {
1879 mach->ins_req(AddPNode::Base,leaf->in(AddPNode::Base));
1880 }
1881
1882 uint number_of_projections_prior = number_of_projections();
1883
1884 // Perform any 1-to-many expansions required
1885 MachNode *ex = mach->Expand(s, _projection_list, mem);
1886 if (ex != mach) {
1887 assert(ex->ideal_reg() == mach->ideal_reg(), "ideal types should match");
1888 if( ex->in(1)->is_Con() )
1889 ex->in(1)->set_req(0, C->root());
1890 // Remove old node from the graph
1891 for( uint i=0; i<mach->req(); i++ ) {
1892 mach->set_req(i,NULL);
1893 }
1894 NOT_PRODUCT(record_new2old(ex, s->_leaf);)
1895 }
1896
1897 // PhaseChaitin::fixup_spills will sometimes generate spill code
1898 // via the matcher. By the time, nodes have been wired into the CFG,
1899 // and any further nodes generated by expand rules will be left hanging
1900 // in space, and will not get emitted as output code. Catch this.
1901 // Also, catch any new register allocation constraints ("projections")
1902 // generated belatedly during spill code generation.
1903 if (_allocation_started) {
1904 guarantee(ex == mach, "no expand rules during spill generation");
1905 guarantee(number_of_projections_prior == number_of_projections(), "no allocation during spill generation");
1906 }
1907
1908 if (leaf->is_Con() || leaf->is_DecodeNarrowPtr()) {
1909 // Record the con for sharing
1910 _shared_nodes.map(leaf->_idx, ex);
1911 }
1912
1913 // Have mach nodes inherit GC barrier data
1914 if (leaf->is_LoadStore()) {
1915 mach->set_barrier_data(leaf->as_LoadStore()->barrier_data());
1916 } else if (leaf->is_Mem()) {
1917 mach->set_barrier_data(leaf->as_Mem()->barrier_data());
1918 }
1919
1920 return ex;
1921 }
1922
1923 void Matcher::handle_precedence_edges(Node* n, MachNode *mach) {
1924 for (uint i = n->req(); i < n->len(); i++) {
1925 if (n->in(i) != NULL) {
1926 mach->add_prec(n->in(i));
1927 }
1928 }
1929 }
1930
1931 void Matcher::ReduceInst_Chain_Rule(State* s, int rule, Node* &mem, MachNode* mach) {
1932 // 'op' is what I am expecting to receive
1933 int op = _leftOp[rule];
1934 // Operand type to catch childs result
1935 // This is what my child will give me.
1936 unsigned int opnd_class_instance = s->rule(op);
1937 // Choose between operand class or not.
1938 // This is what I will receive.
1939 int catch_op = (FIRST_OPERAND_CLASS <= op && op < NUM_OPERANDS) ? opnd_class_instance : op;
1940 // New rule for child. Chase operand classes to get the actual rule.
1941 unsigned int newrule = s->rule(catch_op);
1942
1943 if (newrule < NUM_OPERANDS) {
1944 // Chain from operand or operand class, may be output of shared node
1945 assert(opnd_class_instance < NUM_OPERANDS, "Bad AD file: Instruction chain rule must chain from operand");
1946 // Insert operand into array of operands for this instruction
1947 mach->_opnds[1] = s->MachOperGenerator(opnd_class_instance);
1948
1949 ReduceOper(s, newrule, mem, mach);
1950 } else {
1951 // Chain from the result of an instruction
1952 assert(newrule >= _LAST_MACH_OPER, "Do NOT chain from internal operand");
1953 mach->_opnds[1] = s->MachOperGenerator(_reduceOp[catch_op]);
1954 Node *mem1 = (Node*)1;
1955 debug_only(Node *save_mem_node = _mem_node;)
1956 mach->add_req( ReduceInst(s, newrule, mem1) );
1957 debug_only(_mem_node = save_mem_node;)
1958 }
1959 return;
1960 }
1961
1962
1963 uint Matcher::ReduceInst_Interior( State *s, int rule, Node *&mem, MachNode *mach, uint num_opnds ) {
1964 handle_precedence_edges(s->_leaf, mach);
1965
1966 if( s->_leaf->is_Load() ) {
1967 Node *mem2 = s->_leaf->in(MemNode::Memory);
1968 assert( mem == (Node*)1 || mem == mem2, "multiple Memories being matched at once?" );
1969 debug_only( if( mem == (Node*)1 ) _mem_node = s->_leaf;)
1970 mem = mem2;
1971 }
1972 if( s->_leaf->in(0) != NULL && s->_leaf->req() > 1) {
1973 if( mach->in(0) == NULL )
1974 mach->set_req(0, s->_leaf->in(0));
1975 }
1976
1977 // Now recursively walk the state tree & add operand list.
1978 for( uint i=0; i<2; i++ ) { // binary tree
1979 State *newstate = s->_kids[i];
1980 if( newstate == NULL ) break; // Might only have 1 child
1981 // 'op' is what I am expecting to receive
1982 int op;
1983 if( i == 0 ) {
1984 op = _leftOp[rule];
1985 } else {
1986 op = _rightOp[rule];
1987 }
1988 // Operand type to catch childs result
1989 // This is what my child will give me.
1990 int opnd_class_instance = newstate->rule(op);
1991 // Choose between operand class or not.
1992 // This is what I will receive.
1993 int catch_op = (op >= FIRST_OPERAND_CLASS && op < NUM_OPERANDS) ? opnd_class_instance : op;
1994 // New rule for child. Chase operand classes to get the actual rule.
1995 int newrule = newstate->rule(catch_op);
1996
1997 if (newrule < NUM_OPERANDS) { // Operand/operandClass or internalOp/instruction?
1998 // Operand/operandClass
1999 // Insert operand into array of operands for this instruction
2000 mach->_opnds[num_opnds++] = newstate->MachOperGenerator(opnd_class_instance);
2001 ReduceOper(newstate, newrule, mem, mach);
2002
2003 } else { // Child is internal operand or new instruction
2004 if (newrule < _LAST_MACH_OPER) { // internal operand or instruction?
2005 // internal operand --> call ReduceInst_Interior
2006 // Interior of complex instruction. Do nothing but recurse.
2007 num_opnds = ReduceInst_Interior(newstate, newrule, mem, mach, num_opnds);
2008 } else {
2009 // instruction --> call build operand( ) to catch result
2010 // --> ReduceInst( newrule )
2011 mach->_opnds[num_opnds++] = s->MachOperGenerator(_reduceOp[catch_op]);
2012 Node *mem1 = (Node*)1;
2013 debug_only(Node *save_mem_node = _mem_node;)
2014 mach->add_req( ReduceInst( newstate, newrule, mem1 ) );
2015 debug_only(_mem_node = save_mem_node;)
2016 }
2017 }
2018 assert( mach->_opnds[num_opnds-1], "" );
2019 }
2020 return num_opnds;
2021 }
2022
2023 // This routine walks the interior of possible complex operands.
2024 // At each point we check our children in the match tree:
2025 // (1) No children -
2026 // We are a leaf; add _leaf field as an input to the MachNode
2027 // (2) Child is an internal operand -
2028 // Skip over it ( do nothing )
2029 // (3) Child is an instruction -
2030 // Call ReduceInst recursively and
2031 // and instruction as an input to the MachNode
2032 void Matcher::ReduceOper( State *s, int rule, Node *&mem, MachNode *mach ) {
2033 assert( rule < _LAST_MACH_OPER, "called with operand rule" );
2034 State *kid = s->_kids[0];
2035 assert( kid == NULL || s->_leaf->in(0) == NULL, "internal operands have no control" );
2036
2037 // Leaf? And not subsumed?
2038 if( kid == NULL && !_swallowed[rule] ) {
2039 mach->add_req( s->_leaf ); // Add leaf pointer
2040 return; // Bail out
2041 }
2042
2043 if( s->_leaf->is_Load() ) {
2044 assert( mem == (Node*)1, "multiple Memories being matched at once?" );
2045 mem = s->_leaf->in(MemNode::Memory);
2046 debug_only(_mem_node = s->_leaf;)
2047 }
2048
2049 handle_precedence_edges(s->_leaf, mach);
2050
2051 if( s->_leaf->in(0) && s->_leaf->req() > 1) {
2052 if( !mach->in(0) )
2053 mach->set_req(0,s->_leaf->in(0));
2054 else {
2055 assert( s->_leaf->in(0) == mach->in(0), "same instruction, differing controls?" );
2056 }
2057 }
2058
2059 for (uint i = 0; kid != NULL && i < 2; kid = s->_kids[1], i++) { // binary tree
2060 int newrule;
2061 if( i == 0) {
2062 newrule = kid->rule(_leftOp[rule]);
2063 } else {
2064 newrule = kid->rule(_rightOp[rule]);
2065 }
2066
2067 if (newrule < _LAST_MACH_OPER) { // Operand or instruction?
2068 // Internal operand; recurse but do nothing else
2069 ReduceOper(kid, newrule, mem, mach);
2070
2071 } else { // Child is a new instruction
2072 // Reduce the instruction, and add a direct pointer from this
2073 // machine instruction to the newly reduced one.
2074 Node *mem1 = (Node*)1;
2075 debug_only(Node *save_mem_node = _mem_node;)
2076 mach->add_req( ReduceInst( kid, newrule, mem1 ) );
2077 debug_only(_mem_node = save_mem_node;)
2078 }
2079 }
2080 }
2081
2082
2083 // -------------------------------------------------------------------------
2084 // Java-Java calling convention
2085 // (what you use when Java calls Java)
2086
2087 //------------------------------find_receiver----------------------------------
2088 // For a given signature, return the OptoReg for parameter 0.
2089 OptoReg::Name Matcher::find_receiver() {
2090 VMRegPair regs;
2091 BasicType sig_bt = T_OBJECT;
2092 SharedRuntime::java_calling_convention(&sig_bt, ®s, 1);
2093 // Return argument 0 register. In the LP64 build pointers
2094 // take 2 registers, but the VM wants only the 'main' name.
2095 return OptoReg::as_OptoReg(regs.first());
2096 }
2097
2098 bool Matcher::is_vshift_con_pattern(Node* n, Node* m) {
2099 if (n != NULL && m != NULL) {
2100 return VectorNode::is_vector_shift(n) &&
2101 VectorNode::is_vector_shift_count(m) && m->in(1)->is_Con();
2102 }
2103 return false;
2104 }
2105
2106 bool Matcher::clone_node(Node* n, Node* m, Matcher::MStack& mstack) {
2107 // Must clone all producers of flags, or we will not match correctly.
2108 // Suppose a compare setting int-flags is shared (e.g., a switch-tree)
2109 // then it will match into an ideal Op_RegFlags. Alas, the fp-flags
2110 // are also there, so we may match a float-branch to int-flags and
2111 // expect the allocator to haul the flags from the int-side to the
2112 // fp-side. No can do.
2113 if (_must_clone[m->Opcode()]) {
2114 mstack.push(m, Visit);
2115 return true;
2116 }
2117 return pd_clone_node(n, m, mstack);
2118 }
2119
2120 bool Matcher::clone_base_plus_offset_address(AddPNode* m, Matcher::MStack& mstack, VectorSet& address_visited) {
2121 Node *off = m->in(AddPNode::Offset);
2122 if (off->is_Con()) {
2123 address_visited.test_set(m->_idx); // Flag as address_visited
2124 mstack.push(m->in(AddPNode::Address), Pre_Visit);
2125 // Clone X+offset as it also folds into most addressing expressions
2126 mstack.push(off, Visit);
2127 mstack.push(m->in(AddPNode::Base), Pre_Visit);
2128 return true;
2129 }
2130 return false;
2131 }
2132
2133 // A method-klass-holder may be passed in the inline_cache_reg
2134 // and then expanded into the inline_cache_reg and a method_ptr register
2135 // defined in ad_<arch>.cpp
2136
2137 //------------------------------find_shared------------------------------------
2138 // Set bits if Node is shared or otherwise a root
2139 void Matcher::find_shared(Node* n) {
2140 // Allocate stack of size C->live_nodes() * 2 to avoid frequent realloc
2141 MStack mstack(C->live_nodes() * 2);
2142 // Mark nodes as address_visited if they are inputs to an address expression
2143 VectorSet address_visited;
2144 mstack.push(n, Visit); // Don't need to pre-visit root node
2145 while (mstack.is_nonempty()) {
2146 n = mstack.node(); // Leave node on stack
2147 Node_State nstate = mstack.state();
2148 uint nop = n->Opcode();
2149 if (nstate == Pre_Visit) {
2150 if (address_visited.test(n->_idx)) { // Visited in address already?
2151 // Flag as visited and shared now.
2152 set_visited(n);
2153 }
2154 if (is_visited(n)) { // Visited already?
2155 // Node is shared and has no reason to clone. Flag it as shared.
2156 // This causes it to match into a register for the sharing.
2157 set_shared(n); // Flag as shared and
2158 if (n->is_DecodeNarrowPtr()) {
2159 // Oop field/array element loads must be shared but since
2160 // they are shared through a DecodeN they may appear to have
2161 // a single use so force sharing here.
2162 set_shared(n->in(1));
2163 }
2164 mstack.pop(); // remove node from stack
2165 continue;
2166 }
2167 nstate = Visit; // Not already visited; so visit now
2168 }
2169 if (nstate == Visit) {
2170 mstack.set_state(Post_Visit);
2171 set_visited(n); // Flag as visited now
2172 bool mem_op = false;
2173 int mem_addr_idx = MemNode::Address;
2174 if (find_shared_visit(mstack, n, nop, mem_op, mem_addr_idx)) {
2175 continue;
2176 }
2177 for (int i = n->len() - 1; i >= 0; --i) { // For my children
2178 Node* m = n->in(i); // Get ith input
2179 if (m == NULL) {
2180 continue; // Ignore NULLs
2181 }
2182 if (clone_node(n, m, mstack)) {
2183 continue;
2184 }
2185
2186 // Clone addressing expressions as they are "free" in memory access instructions
2187 if (mem_op && i == mem_addr_idx && m->is_AddP() &&
2188 // When there are other uses besides address expressions
2189 // put it on stack and mark as shared.
2190 !is_visited(m)) {
2191 // Some inputs for address expression are not put on stack
2192 // to avoid marking them as shared and forcing them into register
2193 // if they are used only in address expressions.
2194 // But they should be marked as shared if there are other uses
2195 // besides address expressions.
2196
2197 if (pd_clone_address_expressions(m->as_AddP(), mstack, address_visited)) {
2198 continue;
2199 }
2200 } // if( mem_op &&
2201 mstack.push(m, Pre_Visit);
2202 } // for(int i = ...)
2203 }
2204 else if (nstate == Alt_Post_Visit) {
2205 mstack.pop(); // Remove node from stack
2206 // We cannot remove the Cmp input from the Bool here, as the Bool may be
2207 // shared and all users of the Bool need to move the Cmp in parallel.
2208 // This leaves both the Bool and the If pointing at the Cmp. To
2209 // prevent the Matcher from trying to Match the Cmp along both paths
2210 // BoolNode::match_edge always returns a zero.
2211
2212 // We reorder the Op_If in a pre-order manner, so we can visit without
2213 // accidentally sharing the Cmp (the Bool and the If make 2 users).
2214 n->add_req( n->in(1)->in(1) ); // Add the Cmp next to the Bool
2215 }
2216 else if (nstate == Post_Visit) {
2217 mstack.pop(); // Remove node from stack
2218
2219 // Now hack a few special opcodes
2220 uint opcode = n->Opcode();
2221 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->matcher_find_shared_post_visit(this, n, opcode);
2222 if (!gc_handled) {
2223 find_shared_post_visit(n, opcode);
2224 }
2225 }
2226 else {
2227 ShouldNotReachHere();
2228 }
2229 } // end of while (mstack.is_nonempty())
2230 }
2231
2232 bool Matcher::find_shared_visit(MStack& mstack, Node* n, uint opcode, bool& mem_op, int& mem_addr_idx) {
2233 switch(opcode) { // Handle some opcodes special
2234 case Op_Phi: // Treat Phis as shared roots
2235 case Op_Parm:
2236 case Op_Proj: // All handled specially during matching
2237 case Op_SafePointScalarObject:
2238 set_shared(n);
2239 set_dontcare(n);
2240 break;
2241 case Op_If:
2242 case Op_CountedLoopEnd:
2243 mstack.set_state(Alt_Post_Visit); // Alternative way
2244 // Convert (If (Bool (CmpX A B))) into (If (Bool) (CmpX A B)). Helps
2245 // with matching cmp/branch in 1 instruction. The Matcher needs the
2246 // Bool and CmpX side-by-side, because it can only get at constants
2247 // that are at the leaves of Match trees, and the Bool's condition acts
2248 // as a constant here.
2249 mstack.push(n->in(1), Visit); // Clone the Bool
2250 mstack.push(n->in(0), Pre_Visit); // Visit control input
2251 return true; // while (mstack.is_nonempty())
2252 case Op_ConvI2D: // These forms efficiently match with a prior
2253 case Op_ConvI2F: // Load but not a following Store
2254 if( n->in(1)->is_Load() && // Prior load
2255 n->outcnt() == 1 && // Not already shared
2256 n->unique_out()->is_Store() ) // Following store
2257 set_shared(n); // Force it to be a root
2258 break;
2259 case Op_ReverseBytesI:
2260 case Op_ReverseBytesL:
2261 if( n->in(1)->is_Load() && // Prior load
2262 n->outcnt() == 1 ) // Not already shared
2263 set_shared(n); // Force it to be a root
2264 break;
2265 case Op_BoxLock: // Can't match until we get stack-regs in ADLC
2266 case Op_IfFalse:
2267 case Op_IfTrue:
2268 case Op_MachProj:
2269 case Op_MergeMem:
2270 case Op_Catch:
2271 case Op_CatchProj:
2272 case Op_CProj:
2273 case Op_JumpProj:
2274 case Op_JProj:
2275 case Op_NeverBranch:
2276 set_dontcare(n);
2277 break;
2278 case Op_Jump:
2279 mstack.push(n->in(1), Pre_Visit); // Switch Value (could be shared)
2280 mstack.push(n->in(0), Pre_Visit); // Visit Control input
2281 return true; // while (mstack.is_nonempty())
2282 case Op_StrComp:
2283 case Op_StrEquals:
2284 case Op_StrIndexOf:
2285 case Op_StrIndexOfChar:
2286 case Op_AryEq:
2287 case Op_CountPositives:
2288 case Op_StrInflatedCopy:
2289 case Op_StrCompressedCopy:
2290 case Op_EncodeISOArray:
2291 case Op_FmaD:
2292 case Op_FmaF:
2293 case Op_FmaVD:
2294 case Op_FmaVF:
2295 case Op_MacroLogicV:
2296 case Op_VectorCmpMasked:
2297 case Op_CompressV:
2298 case Op_CompressM:
2299 case Op_ExpandV:
2300 case Op_VectorLoadMask:
2301 set_shared(n); // Force result into register (it will be anyways)
2302 break;
2303 case Op_ConP: { // Convert pointers above the centerline to NUL
2304 TypeNode *tn = n->as_Type(); // Constants derive from type nodes
2305 const TypePtr* tp = tn->type()->is_ptr();
2306 if (tp->_ptr == TypePtr::AnyNull) {
2307 tn->set_type(TypePtr::NULL_PTR);
2308 }
2309 break;
2310 }
2311 case Op_ConN: { // Convert narrow pointers above the centerline to NUL
2312 TypeNode *tn = n->as_Type(); // Constants derive from type nodes
2313 const TypePtr* tp = tn->type()->make_ptr();
2314 if (tp && tp->_ptr == TypePtr::AnyNull) {
2315 tn->set_type(TypeNarrowOop::NULL_PTR);
2316 }
2317 break;
2318 }
2319 case Op_Binary: // These are introduced in the Post_Visit state.
2320 ShouldNotReachHere();
2321 break;
2322 case Op_ClearArray:
2323 case Op_SafePoint:
2324 mem_op = true;
2325 break;
2326 default:
2327 if( n->is_Store() ) {
2328 // Do match stores, despite no ideal reg
2329 mem_op = true;
2330 break;
2331 }
2332 if( n->is_Mem() ) { // Loads and LoadStores
2333 mem_op = true;
2334 // Loads must be root of match tree due to prior load conflict
2335 if( C->subsume_loads() == false )
2336 set_shared(n);
2337 }
2338 // Fall into default case
2339 if( !n->ideal_reg() )
2340 set_dontcare(n); // Unmatchable Nodes
2341 } // end_switch
2342 return false;
2343 }
2344
2345 void Matcher::find_shared_post_visit(Node* n, uint opcode) {
2346 if (n->is_predicated_vector()) {
2347 // Restructure into binary trees for Matching.
2348 if (n->req() == 4) {
2349 n->set_req(1, new BinaryNode(n->in(1), n->in(2)));
2350 n->set_req(2, n->in(3));
2351 n->del_req(3);
2352 } else if (n->req() == 5) {
2353 n->set_req(1, new BinaryNode(n->in(1), n->in(2)));
2354 n->set_req(2, new BinaryNode(n->in(3), n->in(4)));
2355 n->del_req(4);
2356 n->del_req(3);
2357 } else if (n->req() == 6) {
2358 Node* b3 = new BinaryNode(n->in(4), n->in(5));
2359 Node* b2 = new BinaryNode(n->in(3), b3);
2360 Node* b1 = new BinaryNode(n->in(2), b2);
2361 n->set_req(2, b1);
2362 n->del_req(5);
2363 n->del_req(4);
2364 n->del_req(3);
2365 }
2366 return;
2367 }
2368
2369 switch(opcode) { // Handle some opcodes special
2370 case Op_CompareAndExchangeB:
2371 case Op_CompareAndExchangeS:
2372 case Op_CompareAndExchangeI:
2373 case Op_CompareAndExchangeL:
2374 case Op_CompareAndExchangeP:
2375 case Op_CompareAndExchangeN:
2376 case Op_WeakCompareAndSwapB:
2377 case Op_WeakCompareAndSwapS:
2378 case Op_WeakCompareAndSwapI:
2379 case Op_WeakCompareAndSwapL:
2380 case Op_WeakCompareAndSwapP:
2381 case Op_WeakCompareAndSwapN:
2382 case Op_CompareAndSwapB:
2383 case Op_CompareAndSwapS:
2384 case Op_CompareAndSwapI:
2385 case Op_CompareAndSwapL:
2386 case Op_CompareAndSwapP:
2387 case Op_CompareAndSwapN: { // Convert trinary to binary-tree
2388 Node* newval = n->in(MemNode::ValueIn);
2389 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
2390 Node* pair = new BinaryNode(oldval, newval);
2391 n->set_req(MemNode::ValueIn, pair);
2392 n->del_req(LoadStoreConditionalNode::ExpectedIn);
2393 break;
2394 }
2395 case Op_CMoveD: // Convert trinary to binary-tree
2396 case Op_CMoveF:
2397 case Op_CMoveI:
2398 case Op_CMoveL:
2399 case Op_CMoveN:
2400 case Op_CMoveP: {
2401 // Restructure into a binary tree for Matching. It's possible that
2402 // we could move this code up next to the graph reshaping for IfNodes
2403 // or vice-versa, but I do not want to debug this for Ladybird.
2404 // 10/2/2000 CNC.
2405 Node* pair1 = new BinaryNode(n->in(1), n->in(1)->in(1));
2406 n->set_req(1, pair1);
2407 Node* pair2 = new BinaryNode(n->in(2), n->in(3));
2408 n->set_req(2, pair2);
2409 n->del_req(3);
2410 break;
2411 }
2412 case Op_CMoveVF:
2413 case Op_CMoveVD: {
2414 // Restructure into a binary tree for Matching:
2415 // CMoveVF (Binary bool mask) (Binary src1 src2)
2416 Node* in_cc = n->in(1);
2417 assert(in_cc->is_Con(), "The condition input of cmove vector node must be a constant.");
2418 Node* bol = new BoolNode(in_cc, (BoolTest::mask)in_cc->get_int());
2419 Node* pair1 = new BinaryNode(bol, in_cc);
2420 n->set_req(1, pair1);
2421 Node* pair2 = new BinaryNode(n->in(2), n->in(3));
2422 n->set_req(2, pair2);
2423 n->del_req(3);
2424 break;
2425 }
2426 case Op_VectorCmpMasked: {
2427 Node* pair1 = new BinaryNode(n->in(2), n->in(3));
2428 n->set_req(2, pair1);
2429 n->del_req(3);
2430 break;
2431 }
2432 case Op_MacroLogicV: {
2433 Node* pair1 = new BinaryNode(n->in(1), n->in(2));
2434 Node* pair2 = new BinaryNode(n->in(3), n->in(4));
2435 n->set_req(1, pair1);
2436 n->set_req(2, pair2);
2437 n->del_req(4);
2438 n->del_req(3);
2439 break;
2440 }
2441 case Op_StoreVectorMasked: {
2442 Node* pair = new BinaryNode(n->in(3), n->in(4));
2443 n->set_req(3, pair);
2444 n->del_req(4);
2445 break;
2446 }
2447 case Op_LoopLimit: {
2448 Node* pair1 = new BinaryNode(n->in(1), n->in(2));
2449 n->set_req(1, pair1);
2450 n->set_req(2, n->in(3));
2451 n->del_req(3);
2452 break;
2453 }
2454 case Op_StrEquals:
2455 case Op_StrIndexOfChar: {
2456 Node* pair1 = new BinaryNode(n->in(2), n->in(3));
2457 n->set_req(2, pair1);
2458 n->set_req(3, n->in(4));
2459 n->del_req(4);
2460 break;
2461 }
2462 case Op_StrComp:
2463 case Op_StrIndexOf: {
2464 Node* pair1 = new BinaryNode(n->in(2), n->in(3));
2465 n->set_req(2, pair1);
2466 Node* pair2 = new BinaryNode(n->in(4),n->in(5));
2467 n->set_req(3, pair2);
2468 n->del_req(5);
2469 n->del_req(4);
2470 break;
2471 }
2472 case Op_StrCompressedCopy:
2473 case Op_StrInflatedCopy:
2474 case Op_EncodeISOArray: {
2475 // Restructure into a binary tree for Matching.
2476 Node* pair = new BinaryNode(n->in(3), n->in(4));
2477 n->set_req(3, pair);
2478 n->del_req(4);
2479 break;
2480 }
2481 case Op_FmaD:
2482 case Op_FmaF:
2483 case Op_FmaVD:
2484 case Op_FmaVF: {
2485 // Restructure into a binary tree for Matching.
2486 Node* pair = new BinaryNode(n->in(1), n->in(2));
2487 n->set_req(2, pair);
2488 n->set_req(1, n->in(3));
2489 n->del_req(3);
2490 break;
2491 }
2492 case Op_MulAddS2I: {
2493 Node* pair1 = new BinaryNode(n->in(1), n->in(2));
2494 Node* pair2 = new BinaryNode(n->in(3), n->in(4));
2495 n->set_req(1, pair1);
2496 n->set_req(2, pair2);
2497 n->del_req(4);
2498 n->del_req(3);
2499 break;
2500 }
2501 case Op_ClearArray: {
2502 Node* pair = new BinaryNode(n->in(2), n->in(3));
2503 n->set_req(2, pair);
2504 n->set_req(3, n->in(4));
2505 n->del_req(4);
2506 break;
2507 }
2508 case Op_CopySignD:
2509 case Op_SignumVF:
2510 case Op_SignumVD:
2511 case Op_SignumF:
2512 case Op_SignumD: {
2513 Node* pair = new BinaryNode(n->in(2), n->in(3));
2514 n->set_req(2, pair);
2515 n->del_req(3);
2516 break;
2517 }
2518 case Op_VectorBlend:
2519 case Op_VectorInsert: {
2520 Node* pair = new BinaryNode(n->in(1), n->in(2));
2521 n->set_req(1, pair);
2522 n->set_req(2, n->in(3));
2523 n->del_req(3);
2524 break;
2525 }
2526 case Op_LoadVectorGatherMasked:
2527 case Op_StoreVectorScatter: {
2528 Node* pair = new BinaryNode(n->in(MemNode::ValueIn), n->in(MemNode::ValueIn+1));
2529 n->set_req(MemNode::ValueIn, pair);
2530 n->del_req(MemNode::ValueIn+1);
2531 break;
2532 }
2533 case Op_StoreVectorScatterMasked: {
2534 Node* pair = new BinaryNode(n->in(MemNode::ValueIn+1), n->in(MemNode::ValueIn+2));
2535 n->set_req(MemNode::ValueIn+1, pair);
2536 n->del_req(MemNode::ValueIn+2);
2537 pair = new BinaryNode(n->in(MemNode::ValueIn), n->in(MemNode::ValueIn+1));
2538 n->set_req(MemNode::ValueIn, pair);
2539 n->del_req(MemNode::ValueIn+1);
2540 break;
2541 }
2542 case Op_VectorMaskCmp: {
2543 n->set_req(1, new BinaryNode(n->in(1), n->in(2)));
2544 n->set_req(2, n->in(3));
2545 n->del_req(3);
2546 break;
2547 }
2548 default:
2549 break;
2550 }
2551 }
2552
2553 #ifndef PRODUCT
2554 void Matcher::record_new2old(Node* newn, Node* old) {
2555 _new2old_map.map(newn->_idx, old);
2556 if (!_reused.test_set(old->_igv_idx)) {
2557 // Reuse the Ideal-level IGV identifier so that the node can be tracked
2558 // across matching. If there are multiple machine nodes expanded from the
2559 // same Ideal node, only one will reuse its IGV identifier.
2560 newn->_igv_idx = old->_igv_idx;
2561 }
2562 }
2563
2564 // machine-independent root to machine-dependent root
2565 void Matcher::dump_old2new_map() {
2566 _old2new_map.dump();
2567 }
2568 #endif // !PRODUCT
2569
2570 //---------------------------collect_null_checks-------------------------------
2571 // Find null checks in the ideal graph; write a machine-specific node for
2572 // it. Used by later implicit-null-check handling. Actually collects
2573 // either an IfTrue or IfFalse for the common NOT-null path, AND the ideal
2574 // value being tested.
2575 void Matcher::collect_null_checks( Node *proj, Node *orig_proj ) {
2576 Node *iff = proj->in(0);
2577 if( iff->Opcode() == Op_If ) {
2578 // During matching If's have Bool & Cmp side-by-side
2579 BoolNode *b = iff->in(1)->as_Bool();
2580 Node *cmp = iff->in(2);
2581 int opc = cmp->Opcode();
2582 if (opc != Op_CmpP && opc != Op_CmpN) return;
2583
2584 const Type* ct = cmp->in(2)->bottom_type();
2585 if (ct == TypePtr::NULL_PTR ||
2586 (opc == Op_CmpN && ct == TypeNarrowOop::NULL_PTR)) {
2587
2588 bool push_it = false;
2589 if( proj->Opcode() == Op_IfTrue ) {
2590 #ifndef PRODUCT
2591 extern int all_null_checks_found;
2592 all_null_checks_found++;
2593 #endif
2594 if( b->_test._test == BoolTest::ne ) {
2595 push_it = true;
2596 }
2597 } else {
2598 assert( proj->Opcode() == Op_IfFalse, "" );
2599 if( b->_test._test == BoolTest::eq ) {
2600 push_it = true;
2601 }
2602 }
2603 if( push_it ) {
2604 _null_check_tests.push(proj);
2605 Node* val = cmp->in(1);
2606 #ifdef _LP64
2607 if (val->bottom_type()->isa_narrowoop() &&
2608 !Matcher::narrow_oop_use_complex_address()) {
2609 //
2610 // Look for DecodeN node which should be pinned to orig_proj.
2611 // On platforms (Sparc) which can not handle 2 adds
2612 // in addressing mode we have to keep a DecodeN node and
2613 // use it to do implicit NULL check in address.
2614 //
2615 // DecodeN node was pinned to non-null path (orig_proj) during
2616 // CastPP transformation in final_graph_reshaping_impl().
2617 //
2618 uint cnt = orig_proj->outcnt();
2619 for (uint i = 0; i < orig_proj->outcnt(); i++) {
2620 Node* d = orig_proj->raw_out(i);
2621 if (d->is_DecodeN() && d->in(1) == val) {
2622 val = d;
2623 val->set_req(0, NULL); // Unpin now.
2624 // Mark this as special case to distinguish from
2625 // a regular case: CmpP(DecodeN, NULL).
2626 val = (Node*)(((intptr_t)val) | 1);
2627 break;
2628 }
2629 }
2630 }
2631 #endif
2632 _null_check_tests.push(val);
2633 }
2634 }
2635 }
2636 }
2637
2638 //---------------------------validate_null_checks------------------------------
2639 // Its possible that the value being NULL checked is not the root of a match
2640 // tree. If so, I cannot use the value in an implicit null check.
2641 void Matcher::validate_null_checks( ) {
2642 uint cnt = _null_check_tests.size();
2643 for( uint i=0; i < cnt; i+=2 ) {
2644 Node *test = _null_check_tests[i];
2645 Node *val = _null_check_tests[i+1];
2646 bool is_decoden = ((intptr_t)val) & 1;
2647 val = (Node*)(((intptr_t)val) & ~1);
2648 if (has_new_node(val)) {
2649 Node* new_val = new_node(val);
2650 if (is_decoden) {
2651 assert(val->is_DecodeNarrowPtr() && val->in(0) == NULL, "sanity");
2652 // Note: new_val may have a control edge if
2653 // the original ideal node DecodeN was matched before
2654 // it was unpinned in Matcher::collect_null_checks().
2655 // Unpin the mach node and mark it.
2656 new_val->set_req(0, NULL);
2657 new_val = (Node*)(((intptr_t)new_val) | 1);
2658 }
2659 // Is a match-tree root, so replace with the matched value
2660 _null_check_tests.map(i+1, new_val);
2661 } else {
2662 // Yank from candidate list
2663 _null_check_tests.map(i+1,_null_check_tests[--cnt]);
2664 _null_check_tests.map(i,_null_check_tests[--cnt]);
2665 _null_check_tests.pop();
2666 _null_check_tests.pop();
2667 i-=2;
2668 }
2669 }
2670 }
2671
2672 bool Matcher::gen_narrow_oop_implicit_null_checks() {
2673 // Advice matcher to perform null checks on the narrow oop side.
2674 // Implicit checks are not possible on the uncompressed oop side anyway
2675 // (at least not for read accesses).
2676 // Performs significantly better (especially on Power 6).
2677 if (!os::zero_page_read_protected()) {
2678 return true;
2679 }
2680 return CompressedOops::use_implicit_null_checks() &&
2681 (narrow_oop_use_complex_address() ||
2682 CompressedOops::base() != NULL);
2683 }
2684
2685 // Compute RegMask for an ideal register.
2686 const RegMask* Matcher::regmask_for_ideal_register(uint ideal_reg, Node* ret) {
2687 const Type* t = Type::mreg2type[ideal_reg];
2688 if (t == NULL) {
2689 assert(ideal_reg >= Op_VecA && ideal_reg <= Op_VecZ, "not a vector: %d", ideal_reg);
2690 return NULL; // not supported
2691 }
2692 Node* fp = ret->in(TypeFunc::FramePtr);
2693 Node* mem = ret->in(TypeFunc::Memory);
2694 const TypePtr* atp = TypePtr::BOTTOM;
2695 MemNode::MemOrd mo = MemNode::unordered;
2696
2697 Node* spill;
2698 switch (ideal_reg) {
2699 case Op_RegN: spill = new LoadNNode(NULL, mem, fp, atp, t->is_narrowoop(), mo); break;
2700 case Op_RegI: spill = new LoadINode(NULL, mem, fp, atp, t->is_int(), mo); break;
2701 case Op_RegP: spill = new LoadPNode(NULL, mem, fp, atp, t->is_ptr(), mo); break;
2702 case Op_RegF: spill = new LoadFNode(NULL, mem, fp, atp, t, mo); break;
2703 case Op_RegD: spill = new LoadDNode(NULL, mem, fp, atp, t, mo); break;
2704 case Op_RegL: spill = new LoadLNode(NULL, mem, fp, atp, t->is_long(), mo); break;
2705
2706 case Op_VecA: // fall-through
2707 case Op_VecS: // fall-through
2708 case Op_VecD: // fall-through
2709 case Op_VecX: // fall-through
2710 case Op_VecY: // fall-through
2711 case Op_VecZ: spill = new LoadVectorNode(NULL, mem, fp, atp, t->is_vect()); break;
2712 case Op_RegVectMask: return Matcher::predicate_reg_mask();
2713
2714 default: ShouldNotReachHere();
2715 }
2716 MachNode* mspill = match_tree(spill);
2717 assert(mspill != NULL, "matching failed: %d", ideal_reg);
2718 // Handle generic vector operand case
2719 if (Matcher::supports_generic_vector_operands && t->isa_vect()) {
2720 specialize_mach_node(mspill);
2721 }
2722 return &mspill->out_RegMask();
2723 }
2724
2725 // Process Mach IR right after selection phase is over.
2726 void Matcher::do_postselect_cleanup() {
2727 if (supports_generic_vector_operands) {
2728 specialize_generic_vector_operands();
2729 if (C->failing()) return;
2730 }
2731 }
2732
2733 //----------------------------------------------------------------------
2734 // Generic machine operands elision.
2735 //----------------------------------------------------------------------
2736
2737 // Compute concrete vector operand for a generic TEMP vector mach node based on its user info.
2738 void Matcher::specialize_temp_node(MachTempNode* tmp, MachNode* use, uint idx) {
2739 assert(use->in(idx) == tmp, "not a user");
2740 assert(!Matcher::is_generic_vector(use->_opnds[0]), "use not processed yet");
2741
2742 if ((uint)idx == use->two_adr()) { // DEF_TEMP case
2743 tmp->_opnds[0] = use->_opnds[0]->clone();
2744 } else {
2745 uint ideal_vreg = vector_ideal_reg(C->max_vector_size());
2746 tmp->_opnds[0] = Matcher::pd_specialize_generic_vector_operand(tmp->_opnds[0], ideal_vreg, true /*is_temp*/);
2747 }
2748 }
2749
2750 // Compute concrete vector operand for a generic DEF/USE vector operand (of mach node m at index idx).
2751 MachOper* Matcher::specialize_vector_operand(MachNode* m, uint opnd_idx) {
2752 assert(Matcher::is_generic_vector(m->_opnds[opnd_idx]), "repeated updates");
2753 Node* def = NULL;
2754 if (opnd_idx == 0) { // DEF
2755 def = m; // use mach node itself to compute vector operand type
2756 } else {
2757 int base_idx = m->operand_index(opnd_idx);
2758 def = m->in(base_idx);
2759 if (def->is_Mach()) {
2760 if (def->is_MachTemp() && Matcher::is_generic_vector(def->as_Mach()->_opnds[0])) {
2761 specialize_temp_node(def->as_MachTemp(), m, base_idx); // MachTemp node use site
2762 } else if (is_reg2reg_move(def->as_Mach())) {
2763 def = def->in(1); // skip over generic reg-to-reg moves
2764 }
2765 }
2766 }
2767 assert(def->bottom_type()->isa_vect(), "not a vector");
2768 uint ideal_vreg = def->bottom_type()->ideal_reg();
2769 return Matcher::pd_specialize_generic_vector_operand(m->_opnds[opnd_idx], ideal_vreg, false /*is_temp*/);
2770 }
2771
2772 void Matcher::specialize_mach_node(MachNode* m) {
2773 assert(!m->is_MachTemp(), "processed along with its user");
2774 // For generic use operands pull specific register class operands from
2775 // its def instruction's output operand (def operand).
2776 for (uint i = 0; i < m->num_opnds(); i++) {
2777 if (Matcher::is_generic_vector(m->_opnds[i])) {
2778 m->_opnds[i] = specialize_vector_operand(m, i);
2779 }
2780 }
2781 }
2782
2783 // Replace generic vector operands with concrete vector operands and eliminate generic reg-to-reg moves from the graph.
2784 void Matcher::specialize_generic_vector_operands() {
2785 assert(supports_generic_vector_operands, "sanity");
2786 ResourceMark rm;
2787
2788 // Replace generic vector operands (vec/legVec) with concrete ones (vec[SDXYZ]/legVec[SDXYZ])
2789 // and remove reg-to-reg vector moves (MoveVec2Leg and MoveLeg2Vec).
2790 Unique_Node_List live_nodes;
2791 C->identify_useful_nodes(live_nodes);
2792
2793 while (live_nodes.size() > 0) {
2794 MachNode* m = live_nodes.pop()->isa_Mach();
2795 if (m != NULL) {
2796 if (Matcher::is_reg2reg_move(m)) {
2797 // Register allocator properly handles vec <=> leg moves using register masks.
2798 int opnd_idx = m->operand_index(1);
2799 Node* def = m->in(opnd_idx);
2800 m->subsume_by(def, C);
2801 } else if (m->is_MachTemp()) {
2802 // process MachTemp nodes at use site (see Matcher::specialize_vector_operand)
2803 } else {
2804 specialize_mach_node(m);
2805 }
2806 }
2807 }
2808 }
2809
2810 uint Matcher::vector_length(const Node* n) {
2811 const TypeVect* vt = n->bottom_type()->is_vect();
2812 return vt->length();
2813 }
2814
2815 uint Matcher::vector_length(const MachNode* use, const MachOper* opnd) {
2816 int def_idx = use->operand_index(opnd);
2817 Node* def = use->in(def_idx);
2818 return def->bottom_type()->is_vect()->length();
2819 }
2820
2821 uint Matcher::vector_length_in_bytes(const Node* n) {
2822 const TypeVect* vt = n->bottom_type()->is_vect();
2823 return vt->length_in_bytes();
2824 }
2825
2826 uint Matcher::vector_length_in_bytes(const MachNode* use, const MachOper* opnd) {
2827 uint def_idx = use->operand_index(opnd);
2828 Node* def = use->in(def_idx);
2829 return def->bottom_type()->is_vect()->length_in_bytes();
2830 }
2831
2832 BasicType Matcher::vector_element_basic_type(const Node* n) {
2833 const TypeVect* vt = n->bottom_type()->is_vect();
2834 return vt->element_basic_type();
2835 }
2836
2837 BasicType Matcher::vector_element_basic_type(const MachNode* use, const MachOper* opnd) {
2838 int def_idx = use->operand_index(opnd);
2839 Node* def = use->in(def_idx);
2840 return def->bottom_type()->is_vect()->element_basic_type();
2841 }
2842
2843 #ifdef ASSERT
2844 bool Matcher::verify_after_postselect_cleanup() {
2845 assert(!C->failing(), "sanity");
2846 if (supports_generic_vector_operands) {
2847 Unique_Node_List useful;
2848 C->identify_useful_nodes(useful);
2849 for (uint i = 0; i < useful.size(); i++) {
2850 MachNode* m = useful.at(i)->isa_Mach();
2851 if (m != NULL) {
2852 assert(!Matcher::is_reg2reg_move(m), "no MoveVec nodes allowed");
2853 for (uint j = 0; j < m->num_opnds(); j++) {
2854 assert(!Matcher::is_generic_vector(m->_opnds[j]), "no generic vector operands allowed");
2855 }
2856 }
2857 }
2858 }
2859 return true;
2860 }
2861 #endif // ASSERT
2862
2863 // Used by the DFA in dfa_xxx.cpp. Check for a following barrier or
2864 // atomic instruction acting as a store_load barrier without any
2865 // intervening volatile load, and thus we don't need a barrier here.
2866 // We retain the Node to act as a compiler ordering barrier.
2867 bool Matcher::post_store_load_barrier(const Node* vmb) {
2868 Compile* C = Compile::current();
2869 assert(vmb->is_MemBar(), "");
2870 assert(vmb->Opcode() != Op_MemBarAcquire && vmb->Opcode() != Op_LoadFence, "");
2871 const MemBarNode* membar = vmb->as_MemBar();
2872
2873 // Get the Ideal Proj node, ctrl, that can be used to iterate forward
2874 Node* ctrl = NULL;
2875 for (DUIterator_Fast imax, i = membar->fast_outs(imax); i < imax; i++) {
2876 Node* p = membar->fast_out(i);
2877 assert(p->is_Proj(), "only projections here");
2878 if ((p->as_Proj()->_con == TypeFunc::Control) &&
2879 !C->node_arena()->contains(p)) { // Unmatched old-space only
2880 ctrl = p;
2881 break;
2882 }
2883 }
2884 assert((ctrl != NULL), "missing control projection");
2885
2886 for (DUIterator_Fast jmax, j = ctrl->fast_outs(jmax); j < jmax; j++) {
2887 Node *x = ctrl->fast_out(j);
2888 int xop = x->Opcode();
2889
2890 // We don't need current barrier if we see another or a lock
2891 // before seeing volatile load.
2892 //
2893 // Op_Fastunlock previously appeared in the Op_* list below.
2894 // With the advent of 1-0 lock operations we're no longer guaranteed
2895 // that a monitor exit operation contains a serializing instruction.
2896
2897 if (xop == Op_MemBarVolatile ||
2898 xop == Op_CompareAndExchangeB ||
2899 xop == Op_CompareAndExchangeS ||
2900 xop == Op_CompareAndExchangeI ||
2901 xop == Op_CompareAndExchangeL ||
2902 xop == Op_CompareAndExchangeP ||
2903 xop == Op_CompareAndExchangeN ||
2904 xop == Op_WeakCompareAndSwapB ||
2905 xop == Op_WeakCompareAndSwapS ||
2906 xop == Op_WeakCompareAndSwapL ||
2907 xop == Op_WeakCompareAndSwapP ||
2908 xop == Op_WeakCompareAndSwapN ||
2909 xop == Op_WeakCompareAndSwapI ||
2910 xop == Op_CompareAndSwapB ||
2911 xop == Op_CompareAndSwapS ||
2912 xop == Op_CompareAndSwapL ||
2913 xop == Op_CompareAndSwapP ||
2914 xop == Op_CompareAndSwapN ||
2915 xop == Op_CompareAndSwapI ||
2916 BarrierSet::barrier_set()->barrier_set_c2()->matcher_is_store_load_barrier(x, xop)) {
2917 return true;
2918 }
2919
2920 // Op_FastLock previously appeared in the Op_* list above.
2921 if (xop == Op_FastLock) {
2922 return true;
2923 }
2924
2925 if (x->is_MemBar()) {
2926 // We must retain this membar if there is an upcoming volatile
2927 // load, which will be followed by acquire membar.
2928 if (xop == Op_MemBarAcquire || xop == Op_LoadFence) {
2929 return false;
2930 } else {
2931 // For other kinds of barriers, check by pretending we
2932 // are them, and seeing if we can be removed.
2933 return post_store_load_barrier(x->as_MemBar());
2934 }
2935 }
2936
2937 // probably not necessary to check for these
2938 if (x->is_Call() || x->is_SafePoint() || x->is_block_proj()) {
2939 return false;
2940 }
2941 }
2942 return false;
2943 }
2944
2945 // Check whether node n is a branch to an uncommon trap that we could
2946 // optimize as test with very high branch costs in case of going to
2947 // the uncommon trap. The code must be able to be recompiled to use
2948 // a cheaper test.
2949 bool Matcher::branches_to_uncommon_trap(const Node *n) {
2950 // Don't do it for natives, adapters, or runtime stubs
2951 Compile *C = Compile::current();
2952 if (!C->is_method_compilation()) return false;
2953
2954 assert(n->is_If(), "You should only call this on if nodes.");
2955 IfNode *ifn = n->as_If();
2956
2957 Node *ifFalse = NULL;
2958 for (DUIterator_Fast imax, i = ifn->fast_outs(imax); i < imax; i++) {
2959 if (ifn->fast_out(i)->is_IfFalse()) {
2960 ifFalse = ifn->fast_out(i);
2961 break;
2962 }
2963 }
2964 assert(ifFalse, "An If should have an ifFalse. Graph is broken.");
2965
2966 Node *reg = ifFalse;
2967 int cnt = 4; // We must protect against cycles. Limit to 4 iterations.
2968 // Alternatively use visited set? Seems too expensive.
2969 while (reg != NULL && cnt > 0) {
2970 CallNode *call = NULL;
2971 RegionNode *nxt_reg = NULL;
2972 for (DUIterator_Fast imax, i = reg->fast_outs(imax); i < imax; i++) {
2973 Node *o = reg->fast_out(i);
2974 if (o->is_Call()) {
2975 call = o->as_Call();
2976 }
2977 if (o->is_Region()) {
2978 nxt_reg = o->as_Region();
2979 }
2980 }
2981
2982 if (call &&
2983 call->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) {
2984 const Type* trtype = call->in(TypeFunc::Parms)->bottom_type();
2985 if (trtype->isa_int() && trtype->is_int()->is_con()) {
2986 jint tr_con = trtype->is_int()->get_con();
2987 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
2988 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
2989 assert((int)reason < (int)BitsPerInt, "recode bit map");
2990
2991 if (is_set_nth_bit(C->allowed_deopt_reasons(), (int)reason)
2992 && action != Deoptimization::Action_none) {
2993 // This uncommon trap is sure to recompile, eventually.
2994 // When that happens, C->too_many_traps will prevent
2995 // this transformation from happening again.
2996 return true;
2997 }
2998 }
2999 }
3000
3001 reg = nxt_reg;
3002 cnt--;
3003 }
3004
3005 return false;
3006 }
3007
3008 //=============================================================================
3009 //---------------------------State---------------------------------------------
3010 State::State(void) : _rule() {
3011 #ifdef ASSERT
3012 _id = 0;
3013 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe);
3014 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d);
3015 #endif
3016 }
3017
3018 #ifdef ASSERT
3019 State::~State() {
3020 _id = 99;
3021 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe);
3022 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d);
3023 memset(_cost, -3, sizeof(_cost));
3024 memset(_rule, -3, sizeof(_rule));
3025 }
3026 #endif
3027
3028 #ifndef PRODUCT
3029 //---------------------------dump----------------------------------------------
3030 void State::dump() {
3031 tty->print("\n");
3032 dump(0);
3033 }
3034
3035 void State::dump(int depth) {
3036 for (int j = 0; j < depth; j++) {
3037 tty->print(" ");
3038 }
3039 tty->print("--N: ");
3040 _leaf->dump();
3041 uint i;
3042 for (i = 0; i < _LAST_MACH_OPER; i++) {
3043 // Check for valid entry
3044 if (valid(i)) {
3045 for (int j = 0; j < depth; j++) {
3046 tty->print(" ");
3047 }
3048 assert(cost(i) != max_juint, "cost must be a valid value");
3049 assert(rule(i) < _last_Mach_Node, "rule[i] must be valid rule");
3050 tty->print_cr("%s %d %s",
3051 ruleName[i], cost(i), ruleName[rule(i)] );
3052 }
3053 }
3054 tty->cr();
3055
3056 for (i = 0; i < 2; i++) {
3057 if (_kids[i]) {
3058 _kids[i]->dump(depth + 1);
3059 }
3060 }
3061 }
3062 #endif