Thu, 24 May 2018 19:26:50 +0800
#7046 C2 supports long branch
Contributed-by: fujie
1 /*
2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 /*
26 * This file has been modified by Loongson Technology in 2015. These
27 * modifications are Copyright (c) 2015 Loongson Technology, and are made
28 * available on the same license terms set forth above.
29 */
31 #include "precompiled.hpp"
32 #include "memory/allocation.inline.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/callnode.hpp"
35 #include "opto/connode.hpp"
36 #include "opto/idealGraphPrinter.hpp"
37 #include "opto/matcher.hpp"
38 #include "opto/memnode.hpp"
39 #include "opto/opcodes.hpp"
40 #include "opto/regmask.hpp"
41 #include "opto/rootnode.hpp"
42 #include "opto/runtime.hpp"
43 #include "opto/type.hpp"
44 #include "opto/vectornode.hpp"
45 #include "runtime/atomic.hpp"
46 #include "runtime/os.hpp"
47 #if defined AD_MD_HPP
48 # include AD_MD_HPP
49 #elif defined TARGET_ARCH_MODEL_x86_32
50 # include "adfiles/ad_x86_32.hpp"
51 #elif defined TARGET_ARCH_MODEL_x86_64
52 # include "adfiles/ad_x86_64.hpp"
53 #elif defined TARGET_ARCH_MODEL_sparc
54 # include "adfiles/ad_sparc.hpp"
55 #elif defined TARGET_ARCH_MODEL_zero
56 # include "adfiles/ad_zero.hpp"
57 #elif defined TARGET_ARCH_MODEL_ppc_64
58 # include "adfiles/ad_ppc_64.hpp"
59 #elif defined TARGET_ARCH_MODEL_mips_64
60 # include "adfiles/ad_mips_64.hpp"
61 #endif
63 OptoReg::Name OptoReg::c_frame_pointer;
65 const RegMask *Matcher::idealreg2regmask[_last_machine_leaf];
66 RegMask Matcher::mreg2regmask[_last_Mach_Reg];
67 RegMask Matcher::STACK_ONLY_mask;
68 RegMask Matcher::c_frame_ptr_mask;
69 const uint Matcher::_begin_rematerialize = _BEGIN_REMATERIALIZE;
70 const uint Matcher::_end_rematerialize = _END_REMATERIALIZE;
72 //---------------------------Matcher-------------------------------------------
73 Matcher::Matcher()
74 : PhaseTransform( Phase::Ins_Select ),
75 #ifdef ASSERT
76 _old2new_map(C->comp_arena()),
77 _new2old_map(C->comp_arena()),
78 #endif
79 _shared_nodes(C->comp_arena()),
80 _reduceOp(reduceOp), _leftOp(leftOp), _rightOp(rightOp),
81 _swallowed(swallowed),
82 _begin_inst_chain_rule(_BEGIN_INST_CHAIN_RULE),
83 _end_inst_chain_rule(_END_INST_CHAIN_RULE),
84 _must_clone(must_clone),
85 _register_save_policy(register_save_policy),
86 _c_reg_save_policy(c_reg_save_policy),
87 _register_save_type(register_save_type),
88 _ruleName(ruleName),
89 _allocation_started(false),
90 _states_arena(Chunk::medium_size),
91 _visited(&_states_arena),
92 _shared(&_states_arena),
93 _dontcare(&_states_arena) {
94 C->set_matcher(this);
96 idealreg2spillmask [Op_RegI] = NULL;
97 idealreg2spillmask [Op_RegN] = NULL;
98 idealreg2spillmask [Op_RegL] = NULL;
99 idealreg2spillmask [Op_RegF] = NULL;
100 idealreg2spillmask [Op_RegD] = NULL;
101 idealreg2spillmask [Op_RegP] = NULL;
102 idealreg2spillmask [Op_VecS] = NULL;
103 idealreg2spillmask [Op_VecD] = NULL;
104 idealreg2spillmask [Op_VecX] = NULL;
105 idealreg2spillmask [Op_VecY] = NULL;
107 idealreg2debugmask [Op_RegI] = NULL;
108 idealreg2debugmask [Op_RegN] = NULL;
109 idealreg2debugmask [Op_RegL] = NULL;
110 idealreg2debugmask [Op_RegF] = NULL;
111 idealreg2debugmask [Op_RegD] = NULL;
112 idealreg2debugmask [Op_RegP] = NULL;
113 idealreg2debugmask [Op_VecS] = NULL;
114 idealreg2debugmask [Op_VecD] = NULL;
115 idealreg2debugmask [Op_VecX] = NULL;
116 idealreg2debugmask [Op_VecY] = NULL;
118 idealreg2mhdebugmask[Op_RegI] = NULL;
119 idealreg2mhdebugmask[Op_RegN] = NULL;
120 idealreg2mhdebugmask[Op_RegL] = NULL;
121 idealreg2mhdebugmask[Op_RegF] = NULL;
122 idealreg2mhdebugmask[Op_RegD] = NULL;
123 idealreg2mhdebugmask[Op_RegP] = NULL;
124 idealreg2mhdebugmask[Op_VecS] = NULL;
125 idealreg2mhdebugmask[Op_VecD] = NULL;
126 idealreg2mhdebugmask[Op_VecX] = NULL;
127 idealreg2mhdebugmask[Op_VecY] = NULL;
129 debug_only(_mem_node = NULL;) // Ideal memory node consumed by mach node
130 }
132 //------------------------------warp_incoming_stk_arg------------------------
133 // This warps a VMReg into an OptoReg::Name
134 OptoReg::Name Matcher::warp_incoming_stk_arg( VMReg reg ) {
135 OptoReg::Name warped;
136 if( reg->is_stack() ) { // Stack slot argument?
137 warped = OptoReg::add(_old_SP, reg->reg2stack() );
138 warped = OptoReg::add(warped, C->out_preserve_stack_slots());
139 if( warped >= _in_arg_limit )
140 _in_arg_limit = OptoReg::add(warped, 1); // Bump max stack slot seen
141 if (!RegMask::can_represent_arg(warped)) {
142 // the compiler cannot represent this method's calling sequence
143 C->record_method_not_compilable_all_tiers("unsupported incoming calling sequence");
144 return OptoReg::Bad;
145 }
146 return warped;
147 }
148 return OptoReg::as_OptoReg(reg);
149 }
151 //---------------------------compute_old_SP------------------------------------
152 OptoReg::Name Compile::compute_old_SP() {
153 int fixed = fixed_slots();
154 int preserve = in_preserve_stack_slots();
155 return OptoReg::stack2reg(round_to(fixed + preserve, Matcher::stack_alignment_in_slots()));
156 }
160 #ifdef ASSERT
161 void Matcher::verify_new_nodes_only(Node* xroot) {
162 // Make sure that the new graph only references new nodes
163 ResourceMark rm;
164 Unique_Node_List worklist;
165 VectorSet visited(Thread::current()->resource_area());
166 worklist.push(xroot);
167 while (worklist.size() > 0) {
168 Node* n = worklist.pop();
169 visited <<= n->_idx;
170 assert(C->node_arena()->contains(n), "dead node");
171 for (uint j = 0; j < n->req(); j++) {
172 Node* in = n->in(j);
173 if (in != NULL) {
174 assert(C->node_arena()->contains(in), "dead node");
175 if (!visited.test(in->_idx)) {
176 worklist.push(in);
177 }
178 }
179 }
180 }
181 }
182 #endif
185 //---------------------------match---------------------------------------------
186 void Matcher::match( ) {
187 if( MaxLabelRootDepth < 100 ) { // Too small?
188 assert(false, "invalid MaxLabelRootDepth, increase it to 100 minimum");
189 MaxLabelRootDepth = 100;
190 }
191 // One-time initialization of some register masks.
192 init_spill_mask( C->root()->in(1) );
193 _return_addr_mask = return_addr();
194 #ifdef _LP64
195 // Pointers take 2 slots in 64-bit land
196 _return_addr_mask.Insert(OptoReg::add(return_addr(),1));
197 #endif
199 // Map a Java-signature return type into return register-value
200 // machine registers for 0, 1 and 2 returned values.
201 const TypeTuple *range = C->tf()->range();
202 if( range->cnt() > TypeFunc::Parms ) { // If not a void function
203 // Get ideal-register return type
204 int ireg = range->field_at(TypeFunc::Parms)->ideal_reg();
205 // Get machine return register
206 uint sop = C->start()->Opcode();
207 OptoRegPair regs = return_value(ireg, false);
209 // And mask for same
210 _return_value_mask = RegMask(regs.first());
211 if( OptoReg::is_valid(regs.second()) )
212 _return_value_mask.Insert(regs.second());
213 }
215 // ---------------
216 // Frame Layout
218 // Need the method signature to determine the incoming argument types,
219 // because the types determine which registers the incoming arguments are
220 // in, and this affects the matched code.
221 const TypeTuple *domain = C->tf()->domain();
222 uint argcnt = domain->cnt() - TypeFunc::Parms;
223 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt );
224 VMRegPair *vm_parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt );
225 _parm_regs = NEW_RESOURCE_ARRAY( OptoRegPair, argcnt );
226 _calling_convention_mask = NEW_RESOURCE_ARRAY( RegMask, argcnt );
227 uint i;
228 for( i = 0; i<argcnt; i++ ) {
229 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type();
230 }
232 // Pass array of ideal registers and length to USER code (from the AD file)
233 // that will convert this to an array of register numbers.
234 const StartNode *start = C->start();
235 start->calling_convention( sig_bt, vm_parm_regs, argcnt );
236 #ifdef ASSERT
237 // Sanity check users' calling convention. Real handy while trying to
238 // get the initial port correct.
239 { for (uint i = 0; i<argcnt; i++) {
240 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) {
241 assert(domain->field_at(i+TypeFunc::Parms)==Type::HALF, "only allowed on halve" );
242 _parm_regs[i].set_bad();
243 continue;
244 }
245 VMReg parm_reg = vm_parm_regs[i].first();
246 assert(parm_reg->is_valid(), "invalid arg?");
247 if (parm_reg->is_reg()) {
248 OptoReg::Name opto_parm_reg = OptoReg::as_OptoReg(parm_reg);
249 assert(can_be_java_arg(opto_parm_reg) ||
250 C->stub_function() == CAST_FROM_FN_PTR(address, OptoRuntime::rethrow_C) ||
251 opto_parm_reg == inline_cache_reg(),
252 "parameters in register must be preserved by runtime stubs");
253 }
254 for (uint j = 0; j < i; j++) {
255 assert(parm_reg != vm_parm_regs[j].first(),
256 "calling conv. must produce distinct regs");
257 }
258 }
259 }
260 #endif
262 // Do some initial frame layout.
264 // Compute the old incoming SP (may be called FP) as
265 // OptoReg::stack0() + locks + in_preserve_stack_slots + pad2.
266 _old_SP = C->compute_old_SP();
267 assert( is_even(_old_SP), "must be even" );
269 // Compute highest incoming stack argument as
270 // _old_SP + out_preserve_stack_slots + incoming argument size.
271 _in_arg_limit = OptoReg::add(_old_SP, C->out_preserve_stack_slots());
272 assert( is_even(_in_arg_limit), "out_preserve must be even" );
273 for( i = 0; i < argcnt; i++ ) {
274 // Permit args to have no register
275 _calling_convention_mask[i].Clear();
276 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) {
277 continue;
278 }
279 // calling_convention returns stack arguments as a count of
280 // slots beyond OptoReg::stack0()/VMRegImpl::stack0. We need to convert this to
281 // the allocators point of view, taking into account all the
282 // preserve area, locks & pad2.
284 OptoReg::Name reg1 = warp_incoming_stk_arg(vm_parm_regs[i].first());
285 if( OptoReg::is_valid(reg1))
286 _calling_convention_mask[i].Insert(reg1);
288 OptoReg::Name reg2 = warp_incoming_stk_arg(vm_parm_regs[i].second());
289 if( OptoReg::is_valid(reg2))
290 _calling_convention_mask[i].Insert(reg2);
292 // Saved biased stack-slot register number
293 _parm_regs[i].set_pair(reg2, reg1);
294 }
296 // Finally, make sure the incoming arguments take up an even number of
297 // words, in case the arguments or locals need to contain doubleword stack
298 // slots. The rest of the system assumes that stack slot pairs (in
299 // particular, in the spill area) which look aligned will in fact be
300 // aligned relative to the stack pointer in the target machine. Double
301 // stack slots will always be allocated aligned.
302 _new_SP = OptoReg::Name(round_to(_in_arg_limit, RegMask::SlotsPerLong));
304 // Compute highest outgoing stack argument as
305 // _new_SP + out_preserve_stack_slots + max(outgoing argument size).
306 _out_arg_limit = OptoReg::add(_new_SP, C->out_preserve_stack_slots());
307 assert( is_even(_out_arg_limit), "out_preserve must be even" );
309 if (!RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1))) {
310 // the compiler cannot represent this method's calling sequence
311 C->record_method_not_compilable("must be able to represent all call arguments in reg mask");
312 }
314 if (C->failing()) return; // bailed out on incoming arg failure
316 // ---------------
317 // Collect roots of matcher trees. Every node for which
318 // _shared[_idx] is cleared is guaranteed to not be shared, and thus
319 // can be a valid interior of some tree.
320 find_shared( C->root() );
321 find_shared( C->top() );
323 C->print_method(PHASE_BEFORE_MATCHING);
325 // Create new ideal node ConP #NULL even if it does exist in old space
326 // to avoid false sharing if the corresponding mach node is not used.
327 // The corresponding mach node is only used in rare cases for derived
328 // pointers.
329 Node* new_ideal_null = ConNode::make(C, TypePtr::NULL_PTR);
331 // Swap out to old-space; emptying new-space
332 Arena *old = C->node_arena()->move_contents(C->old_arena());
334 // Save debug and profile information for nodes in old space:
335 _old_node_note_array = C->node_note_array();
336 if (_old_node_note_array != NULL) {
337 C->set_node_note_array(new(C->comp_arena()) GrowableArray<Node_Notes*>
338 (C->comp_arena(), _old_node_note_array->length(),
339 0, NULL));
340 }
342 // Pre-size the new_node table to avoid the need for range checks.
343 grow_new_node_array(C->unique());
345 // Reset node counter so MachNodes start with _idx at 0
346 int live_nodes = C->live_nodes();
347 C->set_unique(0);
348 C->reset_dead_node_list();
350 // Recursively match trees from old space into new space.
351 // Correct leaves of new-space Nodes; they point to old-space.
352 _visited.Clear(); // Clear visit bits for xform call
353 C->set_cached_top_node(xform( C->top(), live_nodes));
354 if (!C->failing()) {
355 Node* xroot = xform( C->root(), 1 );
356 if (xroot == NULL) {
357 Matcher::soft_match_failure(); // recursive matching process failed
358 C->record_method_not_compilable("instruction match failed");
359 } else {
360 // During matching shared constants were attached to C->root()
361 // because xroot wasn't available yet, so transfer the uses to
362 // the xroot.
363 for( DUIterator_Fast jmax, j = C->root()->fast_outs(jmax); j < jmax; j++ ) {
364 Node* n = C->root()->fast_out(j);
365 if (C->node_arena()->contains(n)) {
366 assert(n->in(0) == C->root(), "should be control user");
367 n->set_req(0, xroot);
368 --j;
369 --jmax;
370 }
371 }
373 // Generate new mach node for ConP #NULL
374 assert(new_ideal_null != NULL, "sanity");
375 _mach_null = match_tree(new_ideal_null);
376 // Don't set control, it will confuse GCM since there are no uses.
377 // The control will be set when this node is used first time
378 // in find_base_for_derived().
379 assert(_mach_null != NULL, "");
381 C->set_root(xroot->is_Root() ? xroot->as_Root() : NULL);
383 #ifdef ASSERT
384 verify_new_nodes_only(xroot);
385 #endif
386 }
387 }
388 if (C->top() == NULL || C->root() == NULL) {
389 C->record_method_not_compilable("graph lost"); // %%% cannot happen?
390 }
391 if (C->failing()) {
392 // delete old;
393 old->destruct_contents();
394 return;
395 }
396 assert( C->top(), "" );
397 assert( C->root(), "" );
398 validate_null_checks();
400 // Now smoke old-space
401 NOT_DEBUG( old->destruct_contents() );
403 // ------------------------
404 // Set up save-on-entry registers
405 Fixup_Save_On_Entry( );
406 }
409 //------------------------------Fixup_Save_On_Entry----------------------------
410 // The stated purpose of this routine is to take care of save-on-entry
411 // registers. However, the overall goal of the Match phase is to convert into
412 // machine-specific instructions which have RegMasks to guide allocation.
413 // So what this procedure really does is put a valid RegMask on each input
414 // to the machine-specific variations of all Return, TailCall and Halt
415 // instructions. It also adds edgs to define the save-on-entry values (and of
416 // course gives them a mask).
418 static RegMask *init_input_masks( uint size, RegMask &ret_adr, RegMask &fp ) {
419 RegMask *rms = NEW_RESOURCE_ARRAY( RegMask, size );
420 // Do all the pre-defined register masks
421 rms[TypeFunc::Control ] = RegMask::Empty;
422 rms[TypeFunc::I_O ] = RegMask::Empty;
423 rms[TypeFunc::Memory ] = RegMask::Empty;
424 rms[TypeFunc::ReturnAdr] = ret_adr;
425 rms[TypeFunc::FramePtr ] = fp;
426 return rms;
427 }
429 //---------------------------init_first_stack_mask-----------------------------
430 // Create the initial stack mask used by values spilling to the stack.
431 // Disallow any debug info in outgoing argument areas by setting the
432 // initial mask accordingly.
433 void Matcher::init_first_stack_mask() {
435 // Allocate storage for spill masks as masks for the appropriate load type.
436 RegMask *rms = (RegMask*)C->comp_arena()->Amalloc_D(sizeof(RegMask) * (3*6+4));
438 idealreg2spillmask [Op_RegN] = &rms[0];
439 idealreg2spillmask [Op_RegI] = &rms[1];
440 idealreg2spillmask [Op_RegL] = &rms[2];
441 idealreg2spillmask [Op_RegF] = &rms[3];
442 idealreg2spillmask [Op_RegD] = &rms[4];
443 idealreg2spillmask [Op_RegP] = &rms[5];
445 idealreg2debugmask [Op_RegN] = &rms[6];
446 idealreg2debugmask [Op_RegI] = &rms[7];
447 idealreg2debugmask [Op_RegL] = &rms[8];
448 idealreg2debugmask [Op_RegF] = &rms[9];
449 idealreg2debugmask [Op_RegD] = &rms[10];
450 idealreg2debugmask [Op_RegP] = &rms[11];
452 idealreg2mhdebugmask[Op_RegN] = &rms[12];
453 idealreg2mhdebugmask[Op_RegI] = &rms[13];
454 idealreg2mhdebugmask[Op_RegL] = &rms[14];
455 idealreg2mhdebugmask[Op_RegF] = &rms[15];
456 idealreg2mhdebugmask[Op_RegD] = &rms[16];
457 idealreg2mhdebugmask[Op_RegP] = &rms[17];
459 idealreg2spillmask [Op_VecS] = &rms[18];
460 idealreg2spillmask [Op_VecD] = &rms[19];
461 idealreg2spillmask [Op_VecX] = &rms[20];
462 idealreg2spillmask [Op_VecY] = &rms[21];
464 OptoReg::Name i;
466 // At first, start with the empty mask
467 C->FIRST_STACK_mask().Clear();
469 // Add in the incoming argument area
470 OptoReg::Name init_in = OptoReg::add(_old_SP, C->out_preserve_stack_slots());
471 for (i = init_in; i < _in_arg_limit; i = OptoReg::add(i,1)) {
472 C->FIRST_STACK_mask().Insert(i);
473 }
474 // Add in all bits past the outgoing argument area
475 guarantee(RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1)),
476 "must be able to represent all call arguments in reg mask");
477 OptoReg::Name init = _out_arg_limit;
478 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) {
479 C->FIRST_STACK_mask().Insert(i);
480 }
481 // Finally, set the "infinite stack" bit.
482 C->FIRST_STACK_mask().set_AllStack();
484 // Make spill masks. Registers for their class, plus FIRST_STACK_mask.
485 RegMask aligned_stack_mask = C->FIRST_STACK_mask();
486 // Keep spill masks aligned.
487 aligned_stack_mask.clear_to_pairs();
488 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack");
490 *idealreg2spillmask[Op_RegP] = *idealreg2regmask[Op_RegP];
491 #ifdef _LP64
492 *idealreg2spillmask[Op_RegN] = *idealreg2regmask[Op_RegN];
493 idealreg2spillmask[Op_RegN]->OR(C->FIRST_STACK_mask());
494 idealreg2spillmask[Op_RegP]->OR(aligned_stack_mask);
495 #else
496 idealreg2spillmask[Op_RegP]->OR(C->FIRST_STACK_mask());
497 #endif
498 *idealreg2spillmask[Op_RegI] = *idealreg2regmask[Op_RegI];
499 idealreg2spillmask[Op_RegI]->OR(C->FIRST_STACK_mask());
500 *idealreg2spillmask[Op_RegL] = *idealreg2regmask[Op_RegL];
501 idealreg2spillmask[Op_RegL]->OR(aligned_stack_mask);
502 *idealreg2spillmask[Op_RegF] = *idealreg2regmask[Op_RegF];
503 idealreg2spillmask[Op_RegF]->OR(C->FIRST_STACK_mask());
504 *idealreg2spillmask[Op_RegD] = *idealreg2regmask[Op_RegD];
505 idealreg2spillmask[Op_RegD]->OR(aligned_stack_mask);
507 if (Matcher::vector_size_supported(T_BYTE,4)) {
508 *idealreg2spillmask[Op_VecS] = *idealreg2regmask[Op_VecS];
509 idealreg2spillmask[Op_VecS]->OR(C->FIRST_STACK_mask());
510 }
511 if (Matcher::vector_size_supported(T_FLOAT,2)) {
512 // For VecD we need dual alignment and 8 bytes (2 slots) for spills.
513 // RA guarantees such alignment since it is needed for Double and Long values.
514 *idealreg2spillmask[Op_VecD] = *idealreg2regmask[Op_VecD];
515 idealreg2spillmask[Op_VecD]->OR(aligned_stack_mask);
516 }
517 if (Matcher::vector_size_supported(T_FLOAT,4)) {
518 // For VecX we need quadro alignment and 16 bytes (4 slots) for spills.
519 //
520 // RA can use input arguments stack slots for spills but until RA
521 // we don't know frame size and offset of input arg stack slots.
522 //
523 // Exclude last input arg stack slots to avoid spilling vectors there
524 // otherwise vector spills could stomp over stack slots in caller frame.
525 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1);
526 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecX); k++) {
527 aligned_stack_mask.Remove(in);
528 in = OptoReg::add(in, -1);
529 }
530 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecX);
531 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack");
532 *idealreg2spillmask[Op_VecX] = *idealreg2regmask[Op_VecX];
533 idealreg2spillmask[Op_VecX]->OR(aligned_stack_mask);
534 }
535 if (Matcher::vector_size_supported(T_FLOAT,8)) {
536 // For VecY we need octo alignment and 32 bytes (8 slots) for spills.
537 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1);
538 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecY); k++) {
539 aligned_stack_mask.Remove(in);
540 in = OptoReg::add(in, -1);
541 }
542 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecY);
543 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack");
544 *idealreg2spillmask[Op_VecY] = *idealreg2regmask[Op_VecY];
545 idealreg2spillmask[Op_VecY]->OR(aligned_stack_mask);
546 }
547 if (UseFPUForSpilling) {
548 // This mask logic assumes that the spill operations are
549 // symmetric and that the registers involved are the same size.
550 // On sparc for instance we may have to use 64 bit moves will
551 // kill 2 registers when used with F0-F31.
552 idealreg2spillmask[Op_RegI]->OR(*idealreg2regmask[Op_RegF]);
553 idealreg2spillmask[Op_RegF]->OR(*idealreg2regmask[Op_RegI]);
554 #ifdef _LP64
555 idealreg2spillmask[Op_RegN]->OR(*idealreg2regmask[Op_RegF]);
556 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]);
557 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]);
558 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegD]);
559 #else
560 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegF]);
561 #ifdef ARM
562 // ARM has support for moving 64bit values between a pair of
563 // integer registers and a double register
564 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]);
565 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]);
566 #endif
567 #endif
568 }
570 // Make up debug masks. Any spill slot plus callee-save registers.
571 // Caller-save registers are assumed to be trashable by the various
572 // inline-cache fixup routines.
573 *idealreg2debugmask [Op_RegN]= *idealreg2spillmask[Op_RegN];
574 *idealreg2debugmask [Op_RegI]= *idealreg2spillmask[Op_RegI];
575 *idealreg2debugmask [Op_RegL]= *idealreg2spillmask[Op_RegL];
576 *idealreg2debugmask [Op_RegF]= *idealreg2spillmask[Op_RegF];
577 *idealreg2debugmask [Op_RegD]= *idealreg2spillmask[Op_RegD];
578 *idealreg2debugmask [Op_RegP]= *idealreg2spillmask[Op_RegP];
580 *idealreg2mhdebugmask[Op_RegN]= *idealreg2spillmask[Op_RegN];
581 *idealreg2mhdebugmask[Op_RegI]= *idealreg2spillmask[Op_RegI];
582 *idealreg2mhdebugmask[Op_RegL]= *idealreg2spillmask[Op_RegL];
583 *idealreg2mhdebugmask[Op_RegF]= *idealreg2spillmask[Op_RegF];
584 *idealreg2mhdebugmask[Op_RegD]= *idealreg2spillmask[Op_RegD];
585 *idealreg2mhdebugmask[Op_RegP]= *idealreg2spillmask[Op_RegP];
587 // Prevent stub compilations from attempting to reference
588 // callee-saved registers from debug info
589 bool exclude_soe = !Compile::current()->is_method_compilation();
591 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) {
592 // registers the caller has to save do not work
593 if( _register_save_policy[i] == 'C' ||
594 _register_save_policy[i] == 'A' ||
595 (_register_save_policy[i] == 'E' && exclude_soe) ) {
596 idealreg2debugmask [Op_RegN]->Remove(i);
597 idealreg2debugmask [Op_RegI]->Remove(i); // Exclude save-on-call
598 idealreg2debugmask [Op_RegL]->Remove(i); // registers from debug
599 idealreg2debugmask [Op_RegF]->Remove(i); // masks
600 idealreg2debugmask [Op_RegD]->Remove(i);
601 idealreg2debugmask [Op_RegP]->Remove(i);
603 idealreg2mhdebugmask[Op_RegN]->Remove(i);
604 idealreg2mhdebugmask[Op_RegI]->Remove(i);
605 idealreg2mhdebugmask[Op_RegL]->Remove(i);
606 idealreg2mhdebugmask[Op_RegF]->Remove(i);
607 idealreg2mhdebugmask[Op_RegD]->Remove(i);
608 idealreg2mhdebugmask[Op_RegP]->Remove(i);
609 }
610 }
612 // Subtract the register we use to save the SP for MethodHandle
613 // invokes to from the debug mask.
614 const RegMask save_mask = method_handle_invoke_SP_save_mask();
615 idealreg2mhdebugmask[Op_RegN]->SUBTRACT(save_mask);
616 idealreg2mhdebugmask[Op_RegI]->SUBTRACT(save_mask);
617 idealreg2mhdebugmask[Op_RegL]->SUBTRACT(save_mask);
618 idealreg2mhdebugmask[Op_RegF]->SUBTRACT(save_mask);
619 idealreg2mhdebugmask[Op_RegD]->SUBTRACT(save_mask);
620 idealreg2mhdebugmask[Op_RegP]->SUBTRACT(save_mask);
621 }
623 //---------------------------is_save_on_entry----------------------------------
624 bool Matcher::is_save_on_entry( int reg ) {
625 return
626 _register_save_policy[reg] == 'E' ||
627 _register_save_policy[reg] == 'A' || // Save-on-entry register?
628 // Also save argument registers in the trampolining stubs
629 (C->save_argument_registers() && is_spillable_arg(reg));
630 }
632 //---------------------------Fixup_Save_On_Entry-------------------------------
633 void Matcher::Fixup_Save_On_Entry( ) {
634 init_first_stack_mask();
636 Node *root = C->root(); // Short name for root
637 // Count number of save-on-entry registers.
638 uint soe_cnt = number_of_saved_registers();
639 uint i;
641 // Find the procedure Start Node
642 StartNode *start = C->start();
643 assert( start, "Expect a start node" );
645 // Save argument registers in the trampolining stubs
646 if( C->save_argument_registers() )
647 for( i = 0; i < _last_Mach_Reg; i++ )
648 if( is_spillable_arg(i) )
649 soe_cnt++;
651 // Input RegMask array shared by all Returns.
652 // The type for doubles and longs has a count of 2, but
653 // there is only 1 returned value
654 uint ret_edge_cnt = TypeFunc::Parms + ((C->tf()->range()->cnt() == TypeFunc::Parms) ? 0 : 1);
655 RegMask *ret_rms = init_input_masks( ret_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
656 // Returns have 0 or 1 returned values depending on call signature.
657 // Return register is specified by return_value in the AD file.
658 if (ret_edge_cnt > TypeFunc::Parms)
659 ret_rms[TypeFunc::Parms+0] = _return_value_mask;
661 // Input RegMask array shared by all Rethrows.
662 uint reth_edge_cnt = TypeFunc::Parms+1;
663 RegMask *reth_rms = init_input_masks( reth_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
664 // Rethrow takes exception oop only, but in the argument 0 slot.
665 reth_rms[TypeFunc::Parms] = mreg2regmask[find_receiver(false)];
666 #ifdef _LP64
667 // Need two slots for ptrs in 64-bit land
668 reth_rms[TypeFunc::Parms].Insert(OptoReg::add(OptoReg::Name(find_receiver(false)),1));
669 #endif
671 // Input RegMask array shared by all TailCalls
672 uint tail_call_edge_cnt = TypeFunc::Parms+2;
673 RegMask *tail_call_rms = init_input_masks( tail_call_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
675 // Input RegMask array shared by all TailJumps
676 uint tail_jump_edge_cnt = TypeFunc::Parms+2;
677 RegMask *tail_jump_rms = init_input_masks( tail_jump_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
679 // TailCalls have 2 returned values (target & moop), whose masks come
680 // from the usual MachNode/MachOper mechanism. Find a sample
681 // TailCall to extract these masks and put the correct masks into
682 // the tail_call_rms array.
683 for( i=1; i < root->req(); i++ ) {
684 MachReturnNode *m = root->in(i)->as_MachReturn();
685 if( m->ideal_Opcode() == Op_TailCall ) {
686 tail_call_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0);
687 tail_call_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1);
688 break;
689 }
690 }
692 // TailJumps have 2 returned values (target & ex_oop), whose masks come
693 // from the usual MachNode/MachOper mechanism. Find a sample
694 // TailJump to extract these masks and put the correct masks into
695 // the tail_jump_rms array.
696 for( i=1; i < root->req(); i++ ) {
697 MachReturnNode *m = root->in(i)->as_MachReturn();
698 if( m->ideal_Opcode() == Op_TailJump ) {
699 tail_jump_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0);
700 tail_jump_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1);
701 break;
702 }
703 }
705 // Input RegMask array shared by all Halts
706 uint halt_edge_cnt = TypeFunc::Parms;
707 RegMask *halt_rms = init_input_masks( halt_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
709 // Capture the return input masks into each exit flavor
710 for( i=1; i < root->req(); i++ ) {
711 MachReturnNode *exit = root->in(i)->as_MachReturn();
712 switch( exit->ideal_Opcode() ) {
713 case Op_Return : exit->_in_rms = ret_rms; break;
714 case Op_Rethrow : exit->_in_rms = reth_rms; break;
715 case Op_TailCall : exit->_in_rms = tail_call_rms; break;
716 case Op_TailJump : exit->_in_rms = tail_jump_rms; break;
717 case Op_Halt : exit->_in_rms = halt_rms; break;
718 default : ShouldNotReachHere();
719 }
720 }
722 // Next unused projection number from Start.
723 int proj_cnt = C->tf()->domain()->cnt();
725 // Do all the save-on-entry registers. Make projections from Start for
726 // them, and give them a use at the exit points. To the allocator, they
727 // look like incoming register arguments.
728 for( i = 0; i < _last_Mach_Reg; i++ ) {
729 if( is_save_on_entry(i) ) {
731 // Add the save-on-entry to the mask array
732 ret_rms [ ret_edge_cnt] = mreg2regmask[i];
733 reth_rms [ reth_edge_cnt] = mreg2regmask[i];
734 tail_call_rms[tail_call_edge_cnt] = mreg2regmask[i];
735 tail_jump_rms[tail_jump_edge_cnt] = mreg2regmask[i];
736 // Halts need the SOE registers, but only in the stack as debug info.
737 // A just-prior uncommon-trap or deoptimization will use the SOE regs.
738 halt_rms [ halt_edge_cnt] = *idealreg2spillmask[_register_save_type[i]];
740 Node *mproj;
742 // Is this a RegF low half of a RegD? Double up 2 adjacent RegF's
743 // into a single RegD.
744 if( (i&1) == 0 &&
745 _register_save_type[i ] == Op_RegF &&
746 _register_save_type[i+1] == Op_RegF &&
747 is_save_on_entry(i+1) ) {
748 // Add other bit for double
749 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1));
750 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1));
751 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1));
752 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1));
753 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1));
754 mproj = new (C) MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegD );
755 proj_cnt += 2; // Skip 2 for doubles
756 }
757 else if( (i&1) == 1 && // Else check for high half of double
758 _register_save_type[i-1] == Op_RegF &&
759 _register_save_type[i ] == Op_RegF &&
760 is_save_on_entry(i-1) ) {
761 ret_rms [ ret_edge_cnt] = RegMask::Empty;
762 reth_rms [ reth_edge_cnt] = RegMask::Empty;
763 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty;
764 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty;
765 halt_rms [ halt_edge_cnt] = RegMask::Empty;
766 mproj = C->top();
767 }
768 // Is this a RegI low half of a RegL? Double up 2 adjacent RegI's
769 // into a single RegL.
770 else if( (i&1) == 0 &&
771 _register_save_type[i ] == Op_RegI &&
772 _register_save_type[i+1] == Op_RegI &&
773 is_save_on_entry(i+1) ) {
774 // Add other bit for long
775 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1));
776 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1));
777 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1));
778 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1));
779 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1));
780 mproj = new (C) MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegL );
781 proj_cnt += 2; // Skip 2 for longs
782 }
783 else if( (i&1) == 1 && // Else check for high half of long
784 _register_save_type[i-1] == Op_RegI &&
785 _register_save_type[i ] == Op_RegI &&
786 is_save_on_entry(i-1) ) {
787 ret_rms [ ret_edge_cnt] = RegMask::Empty;
788 reth_rms [ reth_edge_cnt] = RegMask::Empty;
789 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty;
790 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty;
791 halt_rms [ halt_edge_cnt] = RegMask::Empty;
792 mproj = C->top();
793 } else {
794 // Make a projection for it off the Start
795 mproj = new (C) MachProjNode( start, proj_cnt++, ret_rms[ret_edge_cnt], _register_save_type[i] );
796 }
798 ret_edge_cnt ++;
799 reth_edge_cnt ++;
800 tail_call_edge_cnt ++;
801 tail_jump_edge_cnt ++;
802 halt_edge_cnt ++;
804 // Add a use of the SOE register to all exit paths
805 for( uint j=1; j < root->req(); j++ )
806 root->in(j)->add_req(mproj);
807 } // End of if a save-on-entry register
808 } // End of for all machine registers
809 }
811 //------------------------------init_spill_mask--------------------------------
812 void Matcher::init_spill_mask( Node *ret ) {
813 if( idealreg2regmask[Op_RegI] ) return; // One time only init
815 OptoReg::c_frame_pointer = c_frame_pointer();
816 c_frame_ptr_mask = c_frame_pointer();
817 #ifdef _LP64
818 // pointers are twice as big
819 c_frame_ptr_mask.Insert(OptoReg::add(c_frame_pointer(),1));
820 #endif
822 // Start at OptoReg::stack0()
823 STACK_ONLY_mask.Clear();
824 OptoReg::Name init = OptoReg::stack2reg(0);
825 // STACK_ONLY_mask is all stack bits
826 OptoReg::Name i;
827 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1))
828 STACK_ONLY_mask.Insert(i);
829 // Also set the "infinite stack" bit.
830 STACK_ONLY_mask.set_AllStack();
832 // Copy the register names over into the shared world
833 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) {
834 // SharedInfo::regName[i] = regName[i];
835 // Handy RegMasks per machine register
836 mreg2regmask[i].Insert(i);
837 }
839 // Grab the Frame Pointer
840 Node *fp = ret->in(TypeFunc::FramePtr);
841 Node *mem = ret->in(TypeFunc::Memory);
842 const TypePtr* atp = TypePtr::BOTTOM;
843 // Share frame pointer while making spill ops
844 set_shared(fp);
846 // Compute generic short-offset Loads
847 #ifdef _LP64
848 MachNode *spillCP = match_tree(new (C) LoadNNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM,MemNode::unordered));
849 #endif
850 MachNode *spillI = match_tree(new (C) LoadINode(NULL,mem,fp,atp,TypeInt::INT,MemNode::unordered));
851 MachNode *spillL = match_tree(new (C) LoadLNode(NULL,mem,fp,atp,TypeLong::LONG,MemNode::unordered, LoadNode::DependsOnlyOnTest,false));
852 MachNode *spillF = match_tree(new (C) LoadFNode(NULL,mem,fp,atp,Type::FLOAT,MemNode::unordered));
853 MachNode *spillD = match_tree(new (C) LoadDNode(NULL,mem,fp,atp,Type::DOUBLE,MemNode::unordered));
854 MachNode *spillP = match_tree(new (C) LoadPNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM,MemNode::unordered));
855 assert(spillI != NULL && spillL != NULL && spillF != NULL &&
856 spillD != NULL && spillP != NULL, "");
857 // Get the ADLC notion of the right regmask, for each basic type.
858 #ifdef _LP64
859 idealreg2regmask[Op_RegN] = &spillCP->out_RegMask();
860 #endif
861 idealreg2regmask[Op_RegI] = &spillI->out_RegMask();
862 idealreg2regmask[Op_RegL] = &spillL->out_RegMask();
863 idealreg2regmask[Op_RegF] = &spillF->out_RegMask();
864 idealreg2regmask[Op_RegD] = &spillD->out_RegMask();
865 idealreg2regmask[Op_RegP] = &spillP->out_RegMask();
867 // Vector regmasks.
868 if (Matcher::vector_size_supported(T_BYTE,4)) {
869 TypeVect::VECTS = TypeVect::make(T_BYTE, 4);
870 MachNode *spillVectS = match_tree(new (C) LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTS));
871 idealreg2regmask[Op_VecS] = &spillVectS->out_RegMask();
872 }
873 if (Matcher::vector_size_supported(T_FLOAT,2)) {
874 MachNode *spillVectD = match_tree(new (C) LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTD));
875 idealreg2regmask[Op_VecD] = &spillVectD->out_RegMask();
876 }
877 if (Matcher::vector_size_supported(T_FLOAT,4)) {
878 MachNode *spillVectX = match_tree(new (C) LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTX));
879 idealreg2regmask[Op_VecX] = &spillVectX->out_RegMask();
880 }
881 if (Matcher::vector_size_supported(T_FLOAT,8)) {
882 MachNode *spillVectY = match_tree(new (C) LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTY));
883 idealreg2regmask[Op_VecY] = &spillVectY->out_RegMask();
884 }
885 }
887 #ifdef ASSERT
888 static void match_alias_type(Compile* C, Node* n, Node* m) {
889 if (!VerifyAliases) return; // do not go looking for trouble by default
890 const TypePtr* nat = n->adr_type();
891 const TypePtr* mat = m->adr_type();
892 int nidx = C->get_alias_index(nat);
893 int midx = C->get_alias_index(mat);
894 // Detune the assert for cases like (AndI 0xFF (LoadB p)).
895 if (nidx == Compile::AliasIdxTop && midx >= Compile::AliasIdxRaw) {
896 for (uint i = 1; i < n->req(); i++) {
897 Node* n1 = n->in(i);
898 const TypePtr* n1at = n1->adr_type();
899 if (n1at != NULL) {
900 nat = n1at;
901 nidx = C->get_alias_index(n1at);
902 }
903 }
904 }
905 // %%% Kludgery. Instead, fix ideal adr_type methods for all these cases:
906 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxRaw) {
907 switch (n->Opcode()) {
908 case Op_PrefetchRead:
909 case Op_PrefetchWrite:
910 case Op_PrefetchAllocation:
911 nidx = Compile::AliasIdxRaw;
912 nat = TypeRawPtr::BOTTOM;
913 break;
914 }
915 }
916 if (nidx == Compile::AliasIdxRaw && midx == Compile::AliasIdxTop) {
917 switch (n->Opcode()) {
918 case Op_ClearArray:
919 midx = Compile::AliasIdxRaw;
920 mat = TypeRawPtr::BOTTOM;
921 break;
922 }
923 }
924 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxBot) {
925 switch (n->Opcode()) {
926 case Op_Return:
927 case Op_Rethrow:
928 case Op_Halt:
929 case Op_TailCall:
930 case Op_TailJump:
931 nidx = Compile::AliasIdxBot;
932 nat = TypePtr::BOTTOM;
933 break;
934 }
935 }
936 if (nidx == Compile::AliasIdxBot && midx == Compile::AliasIdxTop) {
937 switch (n->Opcode()) {
938 case Op_StrComp:
939 case Op_StrEquals:
940 case Op_StrIndexOf:
941 case Op_AryEq:
942 case Op_MemBarVolatile:
943 case Op_MemBarCPUOrder: // %%% these ideals should have narrower adr_type?
944 case Op_EncodeISOArray:
945 nidx = Compile::AliasIdxTop;
946 nat = NULL;
947 break;
948 }
949 }
950 if (nidx != midx) {
951 if (PrintOpto || (PrintMiscellaneous && (WizardMode || Verbose))) {
952 tty->print_cr("==== Matcher alias shift %d => %d", nidx, midx);
953 n->dump();
954 m->dump();
955 }
956 assert(C->subsume_loads() && C->must_alias(nat, midx),
957 "must not lose alias info when matching");
958 }
959 }
960 #endif
963 //------------------------------MStack-----------------------------------------
964 // State and MStack class used in xform() and find_shared() iterative methods.
965 enum Node_State { Pre_Visit, // node has to be pre-visited
966 Visit, // visit node
967 Post_Visit, // post-visit node
968 Alt_Post_Visit // alternative post-visit path
969 };
971 class MStack: public Node_Stack {
972 public:
973 MStack(int size) : Node_Stack(size) { }
975 void push(Node *n, Node_State ns) {
976 Node_Stack::push(n, (uint)ns);
977 }
978 void push(Node *n, Node_State ns, Node *parent, int indx) {
979 ++_inode_top;
980 if ((_inode_top + 1) >= _inode_max) grow();
981 _inode_top->node = parent;
982 _inode_top->indx = (uint)indx;
983 ++_inode_top;
984 _inode_top->node = n;
985 _inode_top->indx = (uint)ns;
986 }
987 Node *parent() {
988 pop();
989 return node();
990 }
991 Node_State state() const {
992 return (Node_State)index();
993 }
994 void set_state(Node_State ns) {
995 set_index((uint)ns);
996 }
997 };
1000 //------------------------------xform------------------------------------------
1001 // Given a Node in old-space, Match him (Label/Reduce) to produce a machine
1002 // Node in new-space. Given a new-space Node, recursively walk his children.
1003 Node *Matcher::transform( Node *n ) { ShouldNotCallThis(); return n; }
1004 Node *Matcher::xform( Node *n, int max_stack ) {
1005 // Use one stack to keep both: child's node/state and parent's node/index
1006 MStack mstack(max_stack * 2 * 2); // usually: C->live_nodes() * 2 * 2
1007 mstack.push(n, Visit, NULL, -1); // set NULL as parent to indicate root
1009 while (mstack.is_nonempty()) {
1010 C->check_node_count(NodeLimitFudgeFactor, "too many nodes matching instructions");
1011 if (C->failing()) return NULL;
1012 n = mstack.node(); // Leave node on stack
1013 Node_State nstate = mstack.state();
1014 if (nstate == Visit) {
1015 mstack.set_state(Post_Visit);
1016 Node *oldn = n;
1017 // Old-space or new-space check
1018 if (!C->node_arena()->contains(n)) {
1019 // Old space!
1020 Node* m;
1021 if (has_new_node(n)) { // Not yet Label/Reduced
1022 m = new_node(n);
1023 } else {
1024 if (!is_dontcare(n)) { // Matcher can match this guy
1025 // Calls match special. They match alone with no children.
1026 // Their children, the incoming arguments, match normally.
1027 m = n->is_SafePoint() ? match_sfpt(n->as_SafePoint()):match_tree(n);
1028 if (C->failing()) return NULL;
1029 if (m == NULL) { Matcher::soft_match_failure(); return NULL; }
1030 } else { // Nothing the matcher cares about
1031 if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Multi()) { // Projections?
1032 // Convert to machine-dependent projection
1033 m = n->in(0)->as_Multi()->match( n->as_Proj(), this );
1034 #ifdef ASSERT
1035 _new2old_map.map(m->_idx, n);
1036 #endif
1037 if (m->in(0) != NULL) // m might be top
1038 collect_null_checks(m, n);
1039 } else { // Else just a regular 'ol guy
1040 m = n->clone(); // So just clone into new-space
1041 #ifdef ASSERT
1042 _new2old_map.map(m->_idx, n);
1043 #endif
1044 // Def-Use edges will be added incrementally as Uses
1045 // of this node are matched.
1046 assert(m->outcnt() == 0, "no Uses of this clone yet");
1047 }
1048 }
1050 set_new_node(n, m); // Map old to new
1051 if (_old_node_note_array != NULL) {
1052 Node_Notes* nn = C->locate_node_notes(_old_node_note_array,
1053 n->_idx);
1054 C->set_node_notes_at(m->_idx, nn);
1055 }
1056 debug_only(match_alias_type(C, n, m));
1057 }
1058 n = m; // n is now a new-space node
1059 mstack.set_node(n);
1060 }
1062 // New space!
1063 if (_visited.test_set(n->_idx)) continue; // while(mstack.is_nonempty())
1065 int i;
1066 // Put precedence edges on stack first (match them last).
1067 for (i = oldn->req(); (uint)i < oldn->len(); i++) {
1068 Node *m = oldn->in(i);
1069 if (m == NULL) break;
1070 // set -1 to call add_prec() instead of set_req() during Step1
1071 mstack.push(m, Visit, n, -1);
1072 }
1074 // For constant debug info, I'd rather have unmatched constants.
1075 int cnt = n->req();
1076 JVMState* jvms = n->jvms();
1077 int debug_cnt = jvms ? jvms->debug_start() : cnt;
1079 // Now do only debug info. Clone constants rather than matching.
1080 // Constants are represented directly in the debug info without
1081 // the need for executable machine instructions.
1082 // Monitor boxes are also represented directly.
1083 for (i = cnt - 1; i >= debug_cnt; --i) { // For all debug inputs do
1084 Node *m = n->in(i); // Get input
1085 int op = m->Opcode();
1086 assert((op == Op_BoxLock) == jvms->is_monitor_use(i), "boxes only at monitor sites");
1087 if( op == Op_ConI || op == Op_ConP || op == Op_ConN || op == Op_ConNKlass ||
1088 op == Op_ConF || op == Op_ConD || op == Op_ConL
1089 // || op == Op_BoxLock // %%%% enable this and remove (+++) in chaitin.cpp
1090 ) {
1091 m = m->clone();
1092 #ifdef ASSERT
1093 _new2old_map.map(m->_idx, n);
1094 #endif
1095 mstack.push(m, Post_Visit, n, i); // Don't need to visit
1096 mstack.push(m->in(0), Visit, m, 0);
1097 } else {
1098 mstack.push(m, Visit, n, i);
1099 }
1100 }
1102 // And now walk his children, and convert his inputs to new-space.
1103 for( ; i >= 0; --i ) { // For all normal inputs do
1104 Node *m = n->in(i); // Get input
1105 if(m != NULL)
1106 mstack.push(m, Visit, n, i);
1107 }
1109 }
1110 else if (nstate == Post_Visit) {
1111 // Set xformed input
1112 Node *p = mstack.parent();
1113 if (p != NULL) { // root doesn't have parent
1114 int i = (int)mstack.index();
1115 if (i >= 0)
1116 p->set_req(i, n); // required input
1117 else if (i == -1)
1118 p->add_prec(n); // precedence input
1119 else
1120 ShouldNotReachHere();
1121 }
1122 mstack.pop(); // remove processed node from stack
1123 }
1124 else {
1125 ShouldNotReachHere();
1126 }
1127 } // while (mstack.is_nonempty())
1128 return n; // Return new-space Node
1129 }
1131 //------------------------------warp_outgoing_stk_arg------------------------
1132 OptoReg::Name Matcher::warp_outgoing_stk_arg( VMReg reg, OptoReg::Name begin_out_arg_area, OptoReg::Name &out_arg_limit_per_call ) {
1133 // Convert outgoing argument location to a pre-biased stack offset
1134 if (reg->is_stack()) {
1135 OptoReg::Name warped = reg->reg2stack();
1136 // Adjust the stack slot offset to be the register number used
1137 // by the allocator.
1138 warped = OptoReg::add(begin_out_arg_area, warped);
1139 // Keep track of the largest numbered stack slot used for an arg.
1140 // Largest used slot per call-site indicates the amount of stack
1141 // that is killed by the call.
1142 if( warped >= out_arg_limit_per_call )
1143 out_arg_limit_per_call = OptoReg::add(warped,1);
1144 if (!RegMask::can_represent_arg(warped)) {
1145 C->record_method_not_compilable_all_tiers("unsupported calling sequence");
1146 return OptoReg::Bad;
1147 }
1148 return warped;
1149 }
1150 return OptoReg::as_OptoReg(reg);
1151 }
1154 //------------------------------match_sfpt-------------------------------------
1155 // Helper function to match call instructions. Calls match special.
1156 // They match alone with no children. Their children, the incoming
1157 // arguments, match normally.
1158 MachNode *Matcher::match_sfpt( SafePointNode *sfpt ) {
1159 MachSafePointNode *msfpt = NULL;
1160 MachCallNode *mcall = NULL;
1161 uint cnt;
1162 // Split out case for SafePoint vs Call
1163 CallNode *call;
1164 const TypeTuple *domain;
1165 ciMethod* method = NULL;
1166 bool is_method_handle_invoke = false; // for special kill effects
1167 if( sfpt->is_Call() ) {
1168 call = sfpt->as_Call();
1169 domain = call->tf()->domain();
1170 cnt = domain->cnt();
1172 // Match just the call, nothing else
1173 MachNode *m = match_tree(call);
1174 if (C->failing()) return NULL;
1175 if( m == NULL ) { Matcher::soft_match_failure(); return NULL; }
1177 // Copy data from the Ideal SafePoint to the machine version
1178 mcall = m->as_MachCall();
1180 mcall->set_tf( call->tf());
1181 mcall->set_entry_point(call->entry_point());
1182 mcall->set_cnt( call->cnt());
1184 if( mcall->is_MachCallJava() ) {
1185 MachCallJavaNode *mcall_java = mcall->as_MachCallJava();
1186 const CallJavaNode *call_java = call->as_CallJava();
1187 method = call_java->method();
1188 mcall_java->_method = method;
1189 mcall_java->_bci = call_java->_bci;
1190 mcall_java->_optimized_virtual = call_java->is_optimized_virtual();
1191 is_method_handle_invoke = call_java->is_method_handle_invoke();
1192 mcall_java->_method_handle_invoke = is_method_handle_invoke;
1193 if (is_method_handle_invoke) {
1194 C->set_has_method_handle_invokes(true);
1195 }
1196 if( mcall_java->is_MachCallStaticJava() )
1197 mcall_java->as_MachCallStaticJava()->_name =
1198 call_java->as_CallStaticJava()->_name;
1199 if( mcall_java->is_MachCallDynamicJava() )
1200 mcall_java->as_MachCallDynamicJava()->_vtable_index =
1201 call_java->as_CallDynamicJava()->_vtable_index;
1202 }
1203 else if( mcall->is_MachCallRuntime() ) {
1204 mcall->as_MachCallRuntime()->_name = call->as_CallRuntime()->_name;
1205 }
1206 msfpt = mcall;
1207 }
1208 // This is a non-call safepoint
1209 else {
1210 call = NULL;
1211 domain = NULL;
1212 MachNode *mn = match_tree(sfpt);
1213 if (C->failing()) return NULL;
1214 msfpt = mn->as_MachSafePoint();
1215 cnt = TypeFunc::Parms;
1216 }
1218 // Advertise the correct memory effects (for anti-dependence computation).
1219 msfpt->set_adr_type(sfpt->adr_type());
1221 // Allocate a private array of RegMasks. These RegMasks are not shared.
1222 msfpt->_in_rms = NEW_RESOURCE_ARRAY( RegMask, cnt );
1223 // Empty them all.
1224 memset( msfpt->_in_rms, 0, sizeof(RegMask)*cnt );
1226 // Do all the pre-defined non-Empty register masks
1227 msfpt->_in_rms[TypeFunc::ReturnAdr] = _return_addr_mask;
1228 msfpt->_in_rms[TypeFunc::FramePtr ] = c_frame_ptr_mask;
1230 // Place first outgoing argument can possibly be put.
1231 OptoReg::Name begin_out_arg_area = OptoReg::add(_new_SP, C->out_preserve_stack_slots());
1232 assert( is_even(begin_out_arg_area), "" );
1233 // Compute max outgoing register number per call site.
1234 OptoReg::Name out_arg_limit_per_call = begin_out_arg_area;
1235 // Calls to C may hammer extra stack slots above and beyond any arguments.
1236 // These are usually backing store for register arguments for varargs.
1237 if( call != NULL && call->is_CallRuntime() )
1238 out_arg_limit_per_call = OptoReg::add(out_arg_limit_per_call,C->varargs_C_out_slots_killed());
1241 // Do the normal argument list (parameters) register masks
1242 int argcnt = cnt - TypeFunc::Parms;
1243 if( argcnt > 0 ) { // Skip it all if we have no args
1244 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt );
1245 VMRegPair *parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt );
1246 int i;
1247 for( i = 0; i < argcnt; i++ ) {
1248 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type();
1249 }
1250 // V-call to pick proper calling convention
1251 call->calling_convention( sig_bt, parm_regs, argcnt );
1253 #ifdef ASSERT
1254 // Sanity check users' calling convention. Really handy during
1255 // the initial porting effort. Fairly expensive otherwise.
1256 { for (int i = 0; i<argcnt; i++) {
1257 if( !parm_regs[i].first()->is_valid() &&
1258 !parm_regs[i].second()->is_valid() ) continue;
1259 VMReg reg1 = parm_regs[i].first();
1260 VMReg reg2 = parm_regs[i].second();
1261 for (int j = 0; j < i; j++) {
1262 if( !parm_regs[j].first()->is_valid() &&
1263 !parm_regs[j].second()->is_valid() ) continue;
1264 VMReg reg3 = parm_regs[j].first();
1265 VMReg reg4 = parm_regs[j].second();
1266 if( !reg1->is_valid() ) {
1267 assert( !reg2->is_valid(), "valid halvsies" );
1268 } else if( !reg3->is_valid() ) {
1269 assert( !reg4->is_valid(), "valid halvsies" );
1270 } else {
1271 assert( reg1 != reg2, "calling conv. must produce distinct regs");
1272 assert( reg1 != reg3, "calling conv. must produce distinct regs");
1273 assert( reg1 != reg4, "calling conv. must produce distinct regs");
1274 assert( reg2 != reg3, "calling conv. must produce distinct regs");
1275 assert( reg2 != reg4 || !reg2->is_valid(), "calling conv. must produce distinct regs");
1276 assert( reg3 != reg4, "calling conv. must produce distinct regs");
1277 }
1278 }
1279 }
1280 }
1281 #endif
1283 // Visit each argument. Compute its outgoing register mask.
1284 // Return results now can have 2 bits returned.
1285 // Compute max over all outgoing arguments both per call-site
1286 // and over the entire method.
1287 for( i = 0; i < argcnt; i++ ) {
1288 // Address of incoming argument mask to fill in
1289 RegMask *rm = &mcall->_in_rms[i+TypeFunc::Parms];
1290 if( !parm_regs[i].first()->is_valid() &&
1291 !parm_regs[i].second()->is_valid() ) {
1292 continue; // Avoid Halves
1293 }
1294 // Grab first register, adjust stack slots and insert in mask.
1295 OptoReg::Name reg1 = warp_outgoing_stk_arg(parm_regs[i].first(), begin_out_arg_area, out_arg_limit_per_call );
1296 if (OptoReg::is_valid(reg1))
1297 rm->Insert( reg1 );
1298 // Grab second register (if any), adjust stack slots and insert in mask.
1299 OptoReg::Name reg2 = warp_outgoing_stk_arg(parm_regs[i].second(), begin_out_arg_area, out_arg_limit_per_call );
1300 if (OptoReg::is_valid(reg2))
1301 rm->Insert( reg2 );
1302 } // End of for all arguments
1304 // Compute number of stack slots needed to restore stack in case of
1305 // Pascal-style argument popping.
1306 mcall->_argsize = out_arg_limit_per_call - begin_out_arg_area;
1307 }
1309 // Compute the max stack slot killed by any call. These will not be
1310 // available for debug info, and will be used to adjust FIRST_STACK_mask
1311 // after all call sites have been visited.
1312 if( _out_arg_limit < out_arg_limit_per_call)
1313 _out_arg_limit = out_arg_limit_per_call;
1315 if (mcall) {
1316 // Kill the outgoing argument area, including any non-argument holes and
1317 // any legacy C-killed slots. Use Fat-Projections to do the killing.
1318 // Since the max-per-method covers the max-per-call-site and debug info
1319 // is excluded on the max-per-method basis, debug info cannot land in
1320 // this killed area.
1321 uint r_cnt = mcall->tf()->range()->cnt();
1322 MachProjNode *proj = new (C) MachProjNode( mcall, r_cnt+10000, RegMask::Empty, MachProjNode::fat_proj );
1323 if (!RegMask::can_represent_arg(OptoReg::Name(out_arg_limit_per_call-1))) {
1324 C->record_method_not_compilable_all_tiers("unsupported outgoing calling sequence");
1325 } else {
1326 for (int i = begin_out_arg_area; i < out_arg_limit_per_call; i++)
1327 proj->_rout.Insert(OptoReg::Name(i));
1328 }
1329 if (proj->_rout.is_NotEmpty()) {
1330 push_projection(proj);
1331 }
1332 }
1333 // Transfer the safepoint information from the call to the mcall
1334 // Move the JVMState list
1335 msfpt->set_jvms(sfpt->jvms());
1336 for (JVMState* jvms = msfpt->jvms(); jvms; jvms = jvms->caller()) {
1337 jvms->set_map(sfpt);
1338 }
1340 // Debug inputs begin just after the last incoming parameter
1341 assert((mcall == NULL) || (mcall->jvms() == NULL) ||
1342 (mcall->jvms()->debug_start() + mcall->_jvmadj == mcall->tf()->domain()->cnt()), "");
1344 // Move the OopMap
1345 msfpt->_oop_map = sfpt->_oop_map;
1347 // Add additional edges.
1348 if (msfpt->mach_constant_base_node_input() != (uint)-1 && !msfpt->is_MachCallLeaf()) {
1349 // For these calls we can not add MachConstantBase in expand(), as the
1350 // ins are not complete then.
1351 msfpt->ins_req(msfpt->mach_constant_base_node_input(), C->mach_constant_base_node());
1352 if (msfpt->jvms() &&
1353 msfpt->mach_constant_base_node_input() <= msfpt->jvms()->debug_start() + msfpt->_jvmadj) {
1354 // We added an edge before jvms, so we must adapt the position of the ins.
1355 msfpt->jvms()->adapt_position(+1);
1356 }
1357 }
1359 // Registers killed by the call are set in the local scheduling pass
1360 // of Global Code Motion.
1361 return msfpt;
1362 }
1364 //---------------------------match_tree----------------------------------------
1365 // Match a Ideal Node DAG - turn it into a tree; Label & Reduce. Used as part
1366 // of the whole-sale conversion from Ideal to Mach Nodes. Also used for
1367 // making GotoNodes while building the CFG and in init_spill_mask() to identify
1368 // a Load's result RegMask for memoization in idealreg2regmask[]
1369 MachNode *Matcher::match_tree( const Node *n ) {
1370 assert( n->Opcode() != Op_Phi, "cannot match" );
1371 assert( !n->is_block_start(), "cannot match" );
1372 // Set the mark for all locally allocated State objects.
1373 // When this call returns, the _states_arena arena will be reset
1374 // freeing all State objects.
1375 ResourceMark rm( &_states_arena );
1377 LabelRootDepth = 0;
1379 // StoreNodes require their Memory input to match any LoadNodes
1380 Node *mem = n->is_Store() ? n->in(MemNode::Memory) : (Node*)1 ;
1381 #ifdef ASSERT
1382 Node* save_mem_node = _mem_node;
1383 _mem_node = n->is_Store() ? (Node*)n : NULL;
1384 #endif
1385 // State object for root node of match tree
1386 // Allocate it on _states_arena - stack allocation can cause stack overflow.
1387 State *s = new (&_states_arena) State;
1388 s->_kids[0] = NULL;
1389 s->_kids[1] = NULL;
1390 s->_leaf = (Node*)n;
1391 // Label the input tree, allocating labels from top-level arena
1392 Label_Root( n, s, n->in(0), mem );
1393 if (C->failing()) return NULL;
1395 // The minimum cost match for the whole tree is found at the root State
1396 uint mincost = max_juint;
1397 uint cost = max_juint;
1398 uint i;
1399 for( i = 0; i < NUM_OPERANDS; i++ ) {
1400 if( s->valid(i) && // valid entry and
1401 s->_cost[i] < cost && // low cost and
1402 s->_rule[i] >= NUM_OPERANDS ) // not an operand
1403 cost = s->_cost[mincost=i];
1404 }
1405 if (mincost == max_juint) {
1406 #ifndef PRODUCT
1407 tty->print("No matching rule for:");
1408 s->dump();
1409 #endif
1410 Matcher::soft_match_failure();
1411 return NULL;
1412 }
1413 // Reduce input tree based upon the state labels to machine Nodes
1414 MachNode *m = ReduceInst( s, s->_rule[mincost], mem );
1415 #ifdef ASSERT
1416 _old2new_map.map(n->_idx, m);
1417 _new2old_map.map(m->_idx, (Node*)n);
1418 #endif
1420 // Add any Matcher-ignored edges
1421 uint cnt = n->req();
1422 uint start = 1;
1423 if( mem != (Node*)1 ) start = MemNode::Memory+1;
1424 if( n->is_AddP() ) {
1425 assert( mem == (Node*)1, "" );
1426 start = AddPNode::Base+1;
1427 }
1428 for( i = start; i < cnt; i++ ) {
1429 if( !n->match_edge(i) ) {
1430 if( i < m->req() )
1431 m->ins_req( i, n->in(i) );
1432 else
1433 m->add_req( n->in(i) );
1434 }
1435 }
1437 debug_only( _mem_node = save_mem_node; )
1438 return m;
1439 }
1442 //------------------------------match_into_reg---------------------------------
1443 // Choose to either match this Node in a register or part of the current
1444 // match tree. Return true for requiring a register and false for matching
1445 // as part of the current match tree.
1446 static bool match_into_reg( const Node *n, Node *m, Node *control, int i, bool shared ) {
1448 const Type *t = m->bottom_type();
1450 if (t->singleton()) {
1451 // Never force constants into registers. Allow them to match as
1452 // constants or registers. Copies of the same value will share
1453 // the same register. See find_shared_node.
1454 return false;
1455 } else { // Not a constant
1456 // Stop recursion if they have different Controls.
1457 Node* m_control = m->in(0);
1458 // Control of load's memory can post-dominates load's control.
1459 // So use it since load can't float above its memory.
1460 Node* mem_control = (m->is_Load()) ? m->in(MemNode::Memory)->in(0) : NULL;
1461 if (control && m_control && control != m_control && control != mem_control) {
1463 // Actually, we can live with the most conservative control we
1464 // find, if it post-dominates the others. This allows us to
1465 // pick up load/op/store trees where the load can float a little
1466 // above the store.
1467 Node *x = control;
1468 const uint max_scan = 6; // Arbitrary scan cutoff
1469 uint j;
1470 for (j=0; j<max_scan; j++) {
1471 if (x->is_Region()) // Bail out at merge points
1472 return true;
1473 x = x->in(0);
1474 if (x == m_control) // Does 'control' post-dominate
1475 break; // m->in(0)? If so, we can use it
1476 if (x == mem_control) // Does 'control' post-dominate
1477 break; // mem_control? If so, we can use it
1478 }
1479 if (j == max_scan) // No post-domination before scan end?
1480 return true; // Then break the match tree up
1481 }
1482 if ((m->is_DecodeN() && Matcher::narrow_oop_use_complex_address()) ||
1483 (m->is_DecodeNKlass() && Matcher::narrow_klass_use_complex_address())) {
1484 // These are commonly used in address expressions and can
1485 // efficiently fold into them on X64 in some cases.
1486 return false;
1487 }
1488 }
1490 // Not forceable cloning. If shared, put it into a register.
1491 return shared;
1492 }
1495 //------------------------------Instruction Selection--------------------------
1496 // Label method walks a "tree" of nodes, using the ADLC generated DFA to match
1497 // ideal nodes to machine instructions. Trees are delimited by shared Nodes,
1498 // things the Matcher does not match (e.g., Memory), and things with different
1499 // Controls (hence forced into different blocks). We pass in the Control
1500 // selected for this entire State tree.
1502 // The Matcher works on Trees, but an Intel add-to-memory requires a DAG: the
1503 // Store and the Load must have identical Memories (as well as identical
1504 // pointers). Since the Matcher does not have anything for Memory (and
1505 // does not handle DAGs), I have to match the Memory input myself. If the
1506 // Tree root is a Store, I require all Loads to have the identical memory.
1507 Node *Matcher::Label_Root( const Node *n, State *svec, Node *control, const Node *mem){
1508 // Since Label_Root is a recursive function, its possible that we might run
1509 // out of stack space. See bugs 6272980 & 6227033 for more info.
1510 LabelRootDepth++;
1511 if (LabelRootDepth > MaxLabelRootDepth) {
1512 C->record_method_not_compilable_all_tiers("Out of stack space, increase MaxLabelRootDepth");
1513 return NULL;
1514 }
1515 uint care = 0; // Edges matcher cares about
1516 uint cnt = n->req();
1517 uint i = 0;
1519 // Examine children for memory state
1520 // Can only subsume a child into your match-tree if that child's memory state
1521 // is not modified along the path to another input.
1522 // It is unsafe even if the other inputs are separate roots.
1523 Node *input_mem = NULL;
1524 for( i = 1; i < cnt; i++ ) {
1525 if( !n->match_edge(i) ) continue;
1526 Node *m = n->in(i); // Get ith input
1527 assert( m, "expect non-null children" );
1528 if( m->is_Load() ) {
1529 if( input_mem == NULL ) {
1530 input_mem = m->in(MemNode::Memory);
1531 } else if( input_mem != m->in(MemNode::Memory) ) {
1532 input_mem = NodeSentinel;
1533 }
1534 }
1535 }
1537 for( i = 1; i < cnt; i++ ){// For my children
1538 if( !n->match_edge(i) ) continue;
1539 Node *m = n->in(i); // Get ith input
1540 // Allocate states out of a private arena
1541 State *s = new (&_states_arena) State;
1542 svec->_kids[care++] = s;
1543 assert( care <= 2, "binary only for now" );
1545 // Recursively label the State tree.
1546 s->_kids[0] = NULL;
1547 s->_kids[1] = NULL;
1548 s->_leaf = m;
1550 // Check for leaves of the State Tree; things that cannot be a part of
1551 // the current tree. If it finds any, that value is matched as a
1552 // register operand. If not, then the normal matching is used.
1553 if( match_into_reg(n, m, control, i, is_shared(m)) ||
1554 //
1555 // Stop recursion if this is LoadNode and the root of this tree is a
1556 // StoreNode and the load & store have different memories.
1557 ((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) ||
1558 // Can NOT include the match of a subtree when its memory state
1559 // is used by any of the other subtrees
1560 (input_mem == NodeSentinel) ) {
1561 #ifndef PRODUCT
1562 // Print when we exclude matching due to different memory states at input-loads
1563 if( PrintOpto && (Verbose && WizardMode) && (input_mem == NodeSentinel)
1564 && !((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) ) {
1565 tty->print_cr("invalid input_mem");
1566 }
1567 #endif
1568 // Switch to a register-only opcode; this value must be in a register
1569 // and cannot be subsumed as part of a larger instruction.
1570 s->DFA( m->ideal_reg(), m );
1572 } else {
1573 // If match tree has no control and we do, adopt it for entire tree
1574 if( control == NULL && m->in(0) != NULL && m->req() > 1 )
1575 control = m->in(0); // Pick up control
1576 // Else match as a normal part of the match tree.
1577 control = Label_Root(m,s,control,mem);
1578 if (C->failing()) return NULL;
1579 }
1580 }
1583 // Call DFA to match this node, and return
1584 svec->DFA( n->Opcode(), n );
1586 #ifdef ASSERT
1587 uint x;
1588 for( x = 0; x < _LAST_MACH_OPER; x++ )
1589 if( svec->valid(x) )
1590 break;
1592 if (x >= _LAST_MACH_OPER) {
1593 n->dump();
1594 svec->dump();
1595 assert( false, "bad AD file" );
1596 }
1597 #endif
1598 return control;
1599 }
1602 // Con nodes reduced using the same rule can share their MachNode
1603 // which reduces the number of copies of a constant in the final
1604 // program. The register allocator is free to split uses later to
1605 // split live ranges.
1606 MachNode* Matcher::find_shared_node(Node* leaf, uint rule) {
1607 if (!leaf->is_Con() && !leaf->is_DecodeNarrowPtr()) return NULL;
1609 // See if this Con has already been reduced using this rule.
1610 if (_shared_nodes.Size() <= leaf->_idx) return NULL;
1611 MachNode* last = (MachNode*)_shared_nodes.at(leaf->_idx);
1612 if (last != NULL && rule == last->rule()) {
1613 // Don't expect control change for DecodeN
1614 if (leaf->is_DecodeNarrowPtr())
1615 return last;
1616 // Get the new space root.
1617 Node* xroot = new_node(C->root());
1618 if (xroot == NULL) {
1619 // This shouldn't happen give the order of matching.
1620 return NULL;
1621 }
1623 // Shared constants need to have their control be root so they
1624 // can be scheduled properly.
1625 Node* control = last->in(0);
1626 if (control != xroot) {
1627 if (control == NULL || control == C->root()) {
1628 last->set_req(0, xroot);
1629 } else {
1630 assert(false, "unexpected control");
1631 return NULL;
1632 }
1633 }
1634 return last;
1635 }
1636 return NULL;
1637 }
1640 //------------------------------ReduceInst-------------------------------------
1641 // Reduce a State tree (with given Control) into a tree of MachNodes.
1642 // This routine (and it's cohort ReduceOper) convert Ideal Nodes into
1643 // complicated machine Nodes. Each MachNode covers some tree of Ideal Nodes.
1644 // Each MachNode has a number of complicated MachOper operands; each
1645 // MachOper also covers a further tree of Ideal Nodes.
1647 // The root of the Ideal match tree is always an instruction, so we enter
1648 // the recursion here. After building the MachNode, we need to recurse
1649 // the tree checking for these cases:
1650 // (1) Child is an instruction -
1651 // Build the instruction (recursively), add it as an edge.
1652 // Build a simple operand (register) to hold the result of the instruction.
1653 // (2) Child is an interior part of an instruction -
1654 // Skip over it (do nothing)
1655 // (3) Child is the start of a operand -
1656 // Build the operand, place it inside the instruction
1657 // Call ReduceOper.
1658 MachNode *Matcher::ReduceInst( State *s, int rule, Node *&mem ) {
1659 assert( rule >= NUM_OPERANDS, "called with operand rule" );
1661 MachNode* shared_node = find_shared_node(s->_leaf, rule);
1662 if (shared_node != NULL) {
1663 return shared_node;
1664 }
1666 // Build the object to represent this state & prepare for recursive calls
1667 MachNode *mach = s->MachNodeGenerator( rule, C );
1668 guarantee(mach != NULL, "Missing MachNode");
1669 mach->_opnds[0] = s->MachOperGenerator( _reduceOp[rule], C );
1670 assert( mach->_opnds[0] != NULL, "Missing result operand" );
1671 Node *leaf = s->_leaf;
1672 // Check for instruction or instruction chain rule
1673 if( rule >= _END_INST_CHAIN_RULE || rule < _BEGIN_INST_CHAIN_RULE ) {
1674 assert(C->node_arena()->contains(s->_leaf) || !has_new_node(s->_leaf),
1675 "duplicating node that's already been matched");
1676 // Instruction
1677 mach->add_req( leaf->in(0) ); // Set initial control
1678 // Reduce interior of complex instruction
1679 ReduceInst_Interior( s, rule, mem, mach, 1 );
1680 } else {
1681 // Instruction chain rules are data-dependent on their inputs
1682 mach->add_req(0); // Set initial control to none
1683 ReduceInst_Chain_Rule( s, rule, mem, mach );
1684 }
1686 // If a Memory was used, insert a Memory edge
1687 if( mem != (Node*)1 ) {
1688 mach->ins_req(MemNode::Memory,mem);
1689 #ifdef ASSERT
1690 // Verify adr type after matching memory operation
1691 const MachOper* oper = mach->memory_operand();
1692 if (oper != NULL && oper != (MachOper*)-1) {
1693 // It has a unique memory operand. Find corresponding ideal mem node.
1694 Node* m = NULL;
1695 if (leaf->is_Mem()) {
1696 m = leaf;
1697 } else {
1698 m = _mem_node;
1699 assert(m != NULL && m->is_Mem(), "expecting memory node");
1700 }
1701 const Type* mach_at = mach->adr_type();
1702 // DecodeN node consumed by an address may have different type
1703 // then its input. Don't compare types for such case.
1704 if (m->adr_type() != mach_at &&
1705 (m->in(MemNode::Address)->is_DecodeNarrowPtr() ||
1706 m->in(MemNode::Address)->is_AddP() &&
1707 m->in(MemNode::Address)->in(AddPNode::Address)->is_DecodeNarrowPtr() ||
1708 m->in(MemNode::Address)->is_AddP() &&
1709 m->in(MemNode::Address)->in(AddPNode::Address)->is_AddP() &&
1710 m->in(MemNode::Address)->in(AddPNode::Address)->in(AddPNode::Address)->is_DecodeNarrowPtr())) {
1711 mach_at = m->adr_type();
1712 }
1713 if (m->adr_type() != mach_at) {
1714 m->dump();
1715 tty->print_cr("mach:");
1716 mach->dump(1);
1717 }
1718 assert(m->adr_type() == mach_at, "matcher should not change adr type");
1719 }
1720 #endif
1721 }
1723 // If the _leaf is an AddP, insert the base edge
1724 if (leaf->is_AddP()) {
1725 mach->ins_req(AddPNode::Base,leaf->in(AddPNode::Base));
1726 }
1728 uint number_of_projections_prior = number_of_projections();
1730 // Perform any 1-to-many expansions required
1731 MachNode *ex = mach->Expand(s, _projection_list, mem);
1732 if (ex != mach) {
1733 assert(ex->ideal_reg() == mach->ideal_reg(), "ideal types should match");
1734 if( ex->in(1)->is_Con() )
1735 ex->in(1)->set_req(0, C->root());
1736 // Remove old node from the graph
1737 for( uint i=0; i<mach->req(); i++ ) {
1738 mach->set_req(i,NULL);
1739 }
1740 #ifdef ASSERT
1741 _new2old_map.map(ex->_idx, s->_leaf);
1742 #endif
1743 }
1745 // PhaseChaitin::fixup_spills will sometimes generate spill code
1746 // via the matcher. By the time, nodes have been wired into the CFG,
1747 // and any further nodes generated by expand rules will be left hanging
1748 // in space, and will not get emitted as output code. Catch this.
1749 // Also, catch any new register allocation constraints ("projections")
1750 // generated belatedly during spill code generation.
1751 if (_allocation_started) {
1752 guarantee(ex == mach, "no expand rules during spill generation");
1753 guarantee(number_of_projections_prior == number_of_projections(), "no allocation during spill generation");
1754 }
1756 if (leaf->is_Con() || leaf->is_DecodeNarrowPtr()) {
1757 // Record the con for sharing
1758 _shared_nodes.map(leaf->_idx, ex);
1759 }
1761 return ex;
1762 }
1764 void Matcher::ReduceInst_Chain_Rule( State *s, int rule, Node *&mem, MachNode *mach ) {
1765 // 'op' is what I am expecting to receive
1766 int op = _leftOp[rule];
1767 // Operand type to catch childs result
1768 // This is what my child will give me.
1769 int opnd_class_instance = s->_rule[op];
1770 // Choose between operand class or not.
1771 // This is what I will receive.
1772 int catch_op = (FIRST_OPERAND_CLASS <= op && op < NUM_OPERANDS) ? opnd_class_instance : op;
1773 // New rule for child. Chase operand classes to get the actual rule.
1774 int newrule = s->_rule[catch_op];
1776 if( newrule < NUM_OPERANDS ) {
1777 // Chain from operand or operand class, may be output of shared node
1778 assert( 0 <= opnd_class_instance && opnd_class_instance < NUM_OPERANDS,
1779 "Bad AD file: Instruction chain rule must chain from operand");
1780 // Insert operand into array of operands for this instruction
1781 mach->_opnds[1] = s->MachOperGenerator( opnd_class_instance, C );
1783 ReduceOper( s, newrule, mem, mach );
1784 } else {
1785 // Chain from the result of an instruction
1786 assert( newrule >= _LAST_MACH_OPER, "Do NOT chain from internal operand");
1787 mach->_opnds[1] = s->MachOperGenerator( _reduceOp[catch_op], C );
1788 Node *mem1 = (Node*)1;
1789 debug_only(Node *save_mem_node = _mem_node;)
1790 mach->add_req( ReduceInst(s, newrule, mem1) );
1791 debug_only(_mem_node = save_mem_node;)
1792 }
1793 return;
1794 }
1797 uint Matcher::ReduceInst_Interior( State *s, int rule, Node *&mem, MachNode *mach, uint num_opnds ) {
1798 if( s->_leaf->is_Load() ) {
1799 Node *mem2 = s->_leaf->in(MemNode::Memory);
1800 assert( mem == (Node*)1 || mem == mem2, "multiple Memories being matched at once?" );
1801 debug_only( if( mem == (Node*)1 ) _mem_node = s->_leaf;)
1802 mem = mem2;
1803 }
1804 if( s->_leaf->in(0) != NULL && s->_leaf->req() > 1) {
1805 if( mach->in(0) == NULL )
1806 mach->set_req(0, s->_leaf->in(0));
1807 }
1809 // Now recursively walk the state tree & add operand list.
1810 for( uint i=0; i<2; i++ ) { // binary tree
1811 State *newstate = s->_kids[i];
1812 if( newstate == NULL ) break; // Might only have 1 child
1813 // 'op' is what I am expecting to receive
1814 int op;
1815 if( i == 0 ) {
1816 op = _leftOp[rule];
1817 } else {
1818 op = _rightOp[rule];
1819 }
1820 // Operand type to catch childs result
1821 // This is what my child will give me.
1822 int opnd_class_instance = newstate->_rule[op];
1823 // Choose between operand class or not.
1824 // This is what I will receive.
1825 int catch_op = (op >= FIRST_OPERAND_CLASS && op < NUM_OPERANDS) ? opnd_class_instance : op;
1826 // New rule for child. Chase operand classes to get the actual rule.
1827 int newrule = newstate->_rule[catch_op];
1829 if( newrule < NUM_OPERANDS ) { // Operand/operandClass or internalOp/instruction?
1830 // Operand/operandClass
1831 // Insert operand into array of operands for this instruction
1832 mach->_opnds[num_opnds++] = newstate->MachOperGenerator( opnd_class_instance, C );
1833 ReduceOper( newstate, newrule, mem, mach );
1835 } else { // Child is internal operand or new instruction
1836 if( newrule < _LAST_MACH_OPER ) { // internal operand or instruction?
1837 // internal operand --> call ReduceInst_Interior
1838 // Interior of complex instruction. Do nothing but recurse.
1839 num_opnds = ReduceInst_Interior( newstate, newrule, mem, mach, num_opnds );
1840 } else {
1841 // instruction --> call build operand( ) to catch result
1842 // --> ReduceInst( newrule )
1843 mach->_opnds[num_opnds++] = s->MachOperGenerator( _reduceOp[catch_op], C );
1844 Node *mem1 = (Node*)1;
1845 debug_only(Node *save_mem_node = _mem_node;)
1846 mach->add_req( ReduceInst( newstate, newrule, mem1 ) );
1847 debug_only(_mem_node = save_mem_node;)
1848 }
1849 }
1850 assert( mach->_opnds[num_opnds-1], "" );
1851 }
1852 return num_opnds;
1853 }
1855 // This routine walks the interior of possible complex operands.
1856 // At each point we check our children in the match tree:
1857 // (1) No children -
1858 // We are a leaf; add _leaf field as an input to the MachNode
1859 // (2) Child is an internal operand -
1860 // Skip over it ( do nothing )
1861 // (3) Child is an instruction -
1862 // Call ReduceInst recursively and
1863 // and instruction as an input to the MachNode
1864 void Matcher::ReduceOper( State *s, int rule, Node *&mem, MachNode *mach ) {
1865 assert( rule < _LAST_MACH_OPER, "called with operand rule" );
1866 State *kid = s->_kids[0];
1867 assert( kid == NULL || s->_leaf->in(0) == NULL, "internal operands have no control" );
1869 // Leaf? And not subsumed?
1870 if( kid == NULL && !_swallowed[rule] ) {
1871 mach->add_req( s->_leaf ); // Add leaf pointer
1872 return; // Bail out
1873 }
1875 if( s->_leaf->is_Load() ) {
1876 assert( mem == (Node*)1, "multiple Memories being matched at once?" );
1877 mem = s->_leaf->in(MemNode::Memory);
1878 debug_only(_mem_node = s->_leaf;)
1879 }
1880 if( s->_leaf->in(0) && s->_leaf->req() > 1) {
1881 if( !mach->in(0) )
1882 mach->set_req(0,s->_leaf->in(0));
1883 else {
1884 assert( s->_leaf->in(0) == mach->in(0), "same instruction, differing controls?" );
1885 }
1886 }
1888 for( uint i=0; kid != NULL && i<2; kid = s->_kids[1], i++ ) { // binary tree
1889 int newrule;
1890 if( i == 0)
1891 newrule = kid->_rule[_leftOp[rule]];
1892 else
1893 newrule = kid->_rule[_rightOp[rule]];
1895 if( newrule < _LAST_MACH_OPER ) { // Operand or instruction?
1896 // Internal operand; recurse but do nothing else
1897 ReduceOper( kid, newrule, mem, mach );
1899 } else { // Child is a new instruction
1900 // Reduce the instruction, and add a direct pointer from this
1901 // machine instruction to the newly reduced one.
1902 Node *mem1 = (Node*)1;
1903 debug_only(Node *save_mem_node = _mem_node;)
1904 mach->add_req( ReduceInst( kid, newrule, mem1 ) );
1905 debug_only(_mem_node = save_mem_node;)
1906 }
1907 }
1908 }
1911 // -------------------------------------------------------------------------
1912 // Java-Java calling convention
1913 // (what you use when Java calls Java)
1915 //------------------------------find_receiver----------------------------------
1916 // For a given signature, return the OptoReg for parameter 0.
1917 OptoReg::Name Matcher::find_receiver( bool is_outgoing ) {
1918 VMRegPair regs;
1919 BasicType sig_bt = T_OBJECT;
1920 calling_convention(&sig_bt, ®s, 1, is_outgoing);
1921 // Return argument 0 register. In the LP64 build pointers
1922 // take 2 registers, but the VM wants only the 'main' name.
1923 return OptoReg::as_OptoReg(regs.first());
1924 }
1926 // This function identifies sub-graphs in which a 'load' node is
1927 // input to two different nodes, and such that it can be matched
1928 // with BMI instructions like blsi, blsr, etc.
1929 // Example : for b = -a[i] & a[i] can be matched to blsi r32, m32.
1930 // The graph is (AndL (SubL Con0 LoadL*) LoadL*), where LoadL*
1931 // refers to the same node.
1932 #ifdef X86
1933 // Match the generic fused operations pattern (op1 (op2 Con{ConType} mop) mop)
1934 // This is a temporary solution until we make DAGs expressible in ADL.
1935 template<typename ConType>
1936 class FusedPatternMatcher {
1937 Node* _op1_node;
1938 Node* _mop_node;
1939 int _con_op;
1941 static int match_next(Node* n, int next_op, int next_op_idx) {
1942 if (n->in(1) == NULL || n->in(2) == NULL) {
1943 return -1;
1944 }
1946 if (next_op_idx == -1) { // n is commutative, try rotations
1947 if (n->in(1)->Opcode() == next_op) {
1948 return 1;
1949 } else if (n->in(2)->Opcode() == next_op) {
1950 return 2;
1951 }
1952 } else {
1953 assert(next_op_idx > 0 && next_op_idx <= 2, "Bad argument index");
1954 if (n->in(next_op_idx)->Opcode() == next_op) {
1955 return next_op_idx;
1956 }
1957 }
1958 return -1;
1959 }
1960 public:
1961 FusedPatternMatcher(Node* op1_node, Node *mop_node, int con_op) :
1962 _op1_node(op1_node), _mop_node(mop_node), _con_op(con_op) { }
1964 bool match(int op1, int op1_op2_idx, // op1 and the index of the op1->op2 edge, -1 if op1 is commutative
1965 int op2, int op2_con_idx, // op2 and the index of the op2->con edge, -1 if op2 is commutative
1966 typename ConType::NativeType con_value) {
1967 if (_op1_node->Opcode() != op1) {
1968 return false;
1969 }
1970 if (_mop_node->outcnt() > 2) {
1971 return false;
1972 }
1973 op1_op2_idx = match_next(_op1_node, op2, op1_op2_idx);
1974 if (op1_op2_idx == -1) {
1975 return false;
1976 }
1977 // Memory operation must be the other edge
1978 int op1_mop_idx = (op1_op2_idx & 1) + 1;
1980 // Check that the mop node is really what we want
1981 if (_op1_node->in(op1_mop_idx) == _mop_node) {
1982 Node *op2_node = _op1_node->in(op1_op2_idx);
1983 if (op2_node->outcnt() > 1) {
1984 return false;
1985 }
1986 assert(op2_node->Opcode() == op2, "Should be");
1987 op2_con_idx = match_next(op2_node, _con_op, op2_con_idx);
1988 if (op2_con_idx == -1) {
1989 return false;
1990 }
1991 // Memory operation must be the other edge
1992 int op2_mop_idx = (op2_con_idx & 1) + 1;
1993 // Check that the memory operation is the same node
1994 if (op2_node->in(op2_mop_idx) == _mop_node) {
1995 // Now check the constant
1996 const Type* con_type = op2_node->in(op2_con_idx)->bottom_type();
1997 if (con_type != Type::TOP && ConType::as_self(con_type)->get_con() == con_value) {
1998 return true;
1999 }
2000 }
2001 }
2002 return false;
2003 }
2004 };
2007 bool Matcher::is_bmi_pattern(Node *n, Node *m) {
2008 if (n != NULL && m != NULL) {
2009 if (m->Opcode() == Op_LoadI) {
2010 FusedPatternMatcher<TypeInt> bmii(n, m, Op_ConI);
2011 return bmii.match(Op_AndI, -1, Op_SubI, 1, 0) ||
2012 bmii.match(Op_AndI, -1, Op_AddI, -1, -1) ||
2013 bmii.match(Op_XorI, -1, Op_AddI, -1, -1);
2014 } else if (m->Opcode() == Op_LoadL) {
2015 FusedPatternMatcher<TypeLong> bmil(n, m, Op_ConL);
2016 return bmil.match(Op_AndL, -1, Op_SubL, 1, 0) ||
2017 bmil.match(Op_AndL, -1, Op_AddL, -1, -1) ||
2018 bmil.match(Op_XorL, -1, Op_AddL, -1, -1);
2019 }
2020 }
2021 return false;
2022 }
2023 #endif // X86
2025 // A method-klass-holder may be passed in the inline_cache_reg
2026 // and then expanded into the inline_cache_reg and a method_oop register
2027 // defined in ad_<arch>.cpp
2030 //------------------------------find_shared------------------------------------
2031 // Set bits if Node is shared or otherwise a root
2032 void Matcher::find_shared( Node *n ) {
2033 // Allocate stack of size C->live_nodes() * 2 to avoid frequent realloc
2034 MStack mstack(C->live_nodes() * 2);
2035 // Mark nodes as address_visited if they are inputs to an address expression
2036 VectorSet address_visited(Thread::current()->resource_area());
2037 mstack.push(n, Visit); // Don't need to pre-visit root node
2038 while (mstack.is_nonempty()) {
2039 n = mstack.node(); // Leave node on stack
2040 Node_State nstate = mstack.state();
2041 uint nop = n->Opcode();
2042 if (nstate == Pre_Visit) {
2043 if (address_visited.test(n->_idx)) { // Visited in address already?
2044 // Flag as visited and shared now.
2045 set_visited(n);
2046 }
2047 if (is_visited(n)) { // Visited already?
2048 // Node is shared and has no reason to clone. Flag it as shared.
2049 // This causes it to match into a register for the sharing.
2050 set_shared(n); // Flag as shared and
2051 mstack.pop(); // remove node from stack
2052 continue;
2053 }
2054 nstate = Visit; // Not already visited; so visit now
2055 }
2056 if (nstate == Visit) {
2057 mstack.set_state(Post_Visit);
2058 set_visited(n); // Flag as visited now
2059 bool mem_op = false;
2061 switch( nop ) { // Handle some opcodes special
2062 case Op_Phi: // Treat Phis as shared roots
2063 case Op_Parm:
2064 case Op_Proj: // All handled specially during matching
2065 case Op_SafePointScalarObject:
2066 set_shared(n);
2067 set_dontcare(n);
2068 break;
2069 case Op_If:
2070 case Op_CountedLoopEnd:
2071 mstack.set_state(Alt_Post_Visit); // Alternative way
2072 // Convert (If (Bool (CmpX A B))) into (If (Bool) (CmpX A B)). Helps
2073 // with matching cmp/branch in 1 instruction. The Matcher needs the
2074 // Bool and CmpX side-by-side, because it can only get at constants
2075 // that are at the leaves of Match trees, and the Bool's condition acts
2076 // as a constant here.
2077 mstack.push(n->in(1), Visit); // Clone the Bool
2078 mstack.push(n->in(0), Pre_Visit); // Visit control input
2079 continue; // while (mstack.is_nonempty())
2080 case Op_ConvI2D: // These forms efficiently match with a prior
2081 case Op_ConvI2F: // Load but not a following Store
2082 if( n->in(1)->is_Load() && // Prior load
2083 n->outcnt() == 1 && // Not already shared
2084 n->unique_out()->is_Store() ) // Following store
2085 set_shared(n); // Force it to be a root
2086 break;
2087 case Op_ReverseBytesI:
2088 case Op_ReverseBytesL:
2089 if( n->in(1)->is_Load() && // Prior load
2090 n->outcnt() == 1 ) // Not already shared
2091 set_shared(n); // Force it to be a root
2092 break;
2093 case Op_BoxLock: // Cant match until we get stack-regs in ADLC
2094 case Op_IfFalse:
2095 case Op_IfTrue:
2096 case Op_MachProj:
2097 case Op_MergeMem:
2098 case Op_Catch:
2099 case Op_CatchProj:
2100 case Op_CProj:
2101 case Op_JumpProj:
2102 case Op_JProj:
2103 case Op_NeverBranch:
2104 set_dontcare(n);
2105 break;
2106 case Op_Jump:
2107 mstack.push(n->in(1), Pre_Visit); // Switch Value (could be shared)
2108 mstack.push(n->in(0), Pre_Visit); // Visit Control input
2109 continue; // while (mstack.is_nonempty())
2110 case Op_StrComp:
2111 case Op_StrEquals:
2112 case Op_StrIndexOf:
2113 case Op_AryEq:
2114 case Op_EncodeISOArray:
2115 set_shared(n); // Force result into register (it will be anyways)
2116 break;
2117 case Op_ConP: { // Convert pointers above the centerline to NUL
2118 TypeNode *tn = n->as_Type(); // Constants derive from type nodes
2119 const TypePtr* tp = tn->type()->is_ptr();
2120 if (tp->_ptr == TypePtr::AnyNull) {
2121 tn->set_type(TypePtr::NULL_PTR);
2122 }
2123 break;
2124 }
2125 case Op_ConN: { // Convert narrow pointers above the centerline to NUL
2126 TypeNode *tn = n->as_Type(); // Constants derive from type nodes
2127 const TypePtr* tp = tn->type()->make_ptr();
2128 if (tp && tp->_ptr == TypePtr::AnyNull) {
2129 tn->set_type(TypeNarrowOop::NULL_PTR);
2130 }
2131 break;
2132 }
2133 case Op_Binary: // These are introduced in the Post_Visit state.
2134 ShouldNotReachHere();
2135 break;
2136 case Op_ClearArray:
2137 case Op_SafePoint:
2138 mem_op = true;
2139 break;
2140 default:
2141 if( n->is_Store() ) {
2142 // Do match stores, despite no ideal reg
2143 mem_op = true;
2144 break;
2145 }
2146 if( n->is_Mem() ) { // Loads and LoadStores
2147 mem_op = true;
2148 // Loads must be root of match tree due to prior load conflict
2149 if( C->subsume_loads() == false )
2150 set_shared(n);
2151 }
2152 // Fall into default case
2153 if( !n->ideal_reg() )
2154 set_dontcare(n); // Unmatchable Nodes
2155 } // end_switch
2157 for(int i = n->req() - 1; i >= 0; --i) { // For my children
2158 Node *m = n->in(i); // Get ith input
2159 if (m == NULL) continue; // Ignore NULLs
2160 uint mop = m->Opcode();
2162 // Must clone all producers of flags, or we will not match correctly.
2163 // Suppose a compare setting int-flags is shared (e.g., a switch-tree)
2164 // then it will match into an ideal Op_RegFlags. Alas, the fp-flags
2165 // are also there, so we may match a float-branch to int-flags and
2166 // expect the allocator to haul the flags from the int-side to the
2167 // fp-side. No can do.
2168 if( _must_clone[mop] ) {
2169 mstack.push(m, Visit);
2170 continue; // for(int i = ...)
2171 }
2173 if( mop == Op_AddP && m->in(AddPNode::Base)->is_DecodeNarrowPtr()) {
2174 // Bases used in addresses must be shared but since
2175 // they are shared through a DecodeN they may appear
2176 // to have a single use so force sharing here.
2177 set_shared(m->in(AddPNode::Base)->in(1));
2178 }
2180 // if 'n' and 'm' are part of a graph for BMI instruction, clone this node.
2181 #ifdef X86
2182 if (UseBMI1Instructions && is_bmi_pattern(n, m)) {
2183 mstack.push(m, Visit);
2184 continue;
2185 }
2186 #endif
2188 // Clone addressing expressions as they are "free" in memory access instructions
2189 if( mem_op && i == MemNode::Address && mop == Op_AddP ) {
2190 // Some inputs for address expression are not put on stack
2191 // to avoid marking them as shared and forcing them into register
2192 // if they are used only in address expressions.
2193 // But they should be marked as shared if there are other uses
2194 // besides address expressions.
2196 Node *off = m->in(AddPNode::Offset);
2197 if( off->is_Con() &&
2198 // When there are other uses besides address expressions
2199 // put it on stack and mark as shared.
2200 !is_visited(m) ) {
2201 address_visited.test_set(m->_idx); // Flag as address_visited
2202 Node *adr = m->in(AddPNode::Address);
2204 // Intel, ARM and friends can handle 2 adds in addressing mode
2205 if( clone_shift_expressions && adr->is_AddP() &&
2206 // AtomicAdd is not an addressing expression.
2207 // Cheap to find it by looking for screwy base.
2208 !adr->in(AddPNode::Base)->is_top() &&
2209 // Are there other uses besides address expressions?
2210 !is_visited(adr) ) {
2211 address_visited.set(adr->_idx); // Flag as address_visited
2212 Node *shift = adr->in(AddPNode::Offset);
2213 // Check for shift by small constant as well
2214 if( shift->Opcode() == Op_LShiftX && shift->in(2)->is_Con() &&
2215 shift->in(2)->get_int() <= 3 &&
2216 // Are there other uses besides address expressions?
2217 !is_visited(shift) ) {
2218 address_visited.set(shift->_idx); // Flag as address_visited
2219 mstack.push(shift->in(2), Visit);
2220 Node *conv = shift->in(1);
2221 #ifdef _LP64
2222 // Allow Matcher to match the rule which bypass
2223 // ConvI2L operation for an array index on LP64
2224 // if the index value is positive.
2225 if( conv->Opcode() == Op_ConvI2L &&
2226 conv->as_Type()->type()->is_long()->_lo >= 0 &&
2227 // Are there other uses besides address expressions?
2228 !is_visited(conv) ) {
2229 address_visited.set(conv->_idx); // Flag as address_visited
2230 mstack.push(conv->in(1), Pre_Visit);
2231 } else
2232 #endif
2233 mstack.push(conv, Pre_Visit);
2234 } else {
2235 mstack.push(shift, Pre_Visit);
2236 }
2237 mstack.push(adr->in(AddPNode::Address), Pre_Visit);
2238 mstack.push(adr->in(AddPNode::Base), Pre_Visit);
2239 } else { // Sparc, Alpha, PPC and friends
2240 mstack.push(adr, Pre_Visit);
2241 }
2243 // Clone X+offset as it also folds into most addressing expressions
2244 mstack.push(off, Visit);
2245 mstack.push(m->in(AddPNode::Base), Pre_Visit);
2246 continue; // for(int i = ...)
2247 } // if( off->is_Con() )
2248 } // if( mem_op &&
2249 mstack.push(m, Pre_Visit);
2250 } // for(int i = ...)
2251 }
2252 else if (nstate == Alt_Post_Visit) {
2253 mstack.pop(); // Remove node from stack
2254 // We cannot remove the Cmp input from the Bool here, as the Bool may be
2255 // shared and all users of the Bool need to move the Cmp in parallel.
2256 // This leaves both the Bool and the If pointing at the Cmp. To
2257 // prevent the Matcher from trying to Match the Cmp along both paths
2258 // BoolNode::match_edge always returns a zero.
2260 // We reorder the Op_If in a pre-order manner, so we can visit without
2261 // accidentally sharing the Cmp (the Bool and the If make 2 users).
2262 n->add_req( n->in(1)->in(1) ); // Add the Cmp next to the Bool
2263 }
2264 else if (nstate == Post_Visit) {
2265 mstack.pop(); // Remove node from stack
2267 // Now hack a few special opcodes
2268 switch( n->Opcode() ) { // Handle some opcodes special
2269 case Op_StorePConditional:
2270 case Op_StoreIConditional:
2271 case Op_StoreLConditional:
2272 case Op_CompareAndSwapI:
2273 case Op_CompareAndSwapL:
2274 case Op_CompareAndSwapP:
2275 case Op_CompareAndSwapN: { // Convert trinary to binary-tree
2276 Node *newval = n->in(MemNode::ValueIn );
2277 Node *oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
2278 Node *pair = new (C) BinaryNode( oldval, newval );
2279 n->set_req(MemNode::ValueIn,pair);
2280 n->del_req(LoadStoreConditionalNode::ExpectedIn);
2281 break;
2282 }
2283 case Op_CMoveD: // Convert trinary to binary-tree
2284 case Op_CMoveF:
2285 case Op_CMoveI:
2286 case Op_CMoveL:
2287 case Op_CMoveN:
2288 case Op_CMoveP: {
2289 // Restructure into a binary tree for Matching. It's possible that
2290 // we could move this code up next to the graph reshaping for IfNodes
2291 // or vice-versa, but I do not want to debug this for Ladybird.
2292 // 10/2/2000 CNC.
2293 Node *pair1 = new (C) BinaryNode(n->in(1),n->in(1)->in(1));
2294 n->set_req(1,pair1);
2295 Node *pair2 = new (C) BinaryNode(n->in(2),n->in(3));
2296 n->set_req(2,pair2);
2297 n->del_req(3);
2298 break;
2299 }
2300 case Op_LoopLimit: {
2301 Node *pair1 = new (C) BinaryNode(n->in(1),n->in(2));
2302 n->set_req(1,pair1);
2303 n->set_req(2,n->in(3));
2304 n->del_req(3);
2305 break;
2306 }
2307 case Op_StrEquals: {
2308 Node *pair1 = new (C) BinaryNode(n->in(2),n->in(3));
2309 n->set_req(2,pair1);
2310 n->set_req(3,n->in(4));
2311 n->del_req(4);
2312 break;
2313 }
2314 case Op_StrComp:
2315 case Op_StrIndexOf: {
2316 Node *pair1 = new (C) BinaryNode(n->in(2),n->in(3));
2317 n->set_req(2,pair1);
2318 Node *pair2 = new (C) BinaryNode(n->in(4),n->in(5));
2319 n->set_req(3,pair2);
2320 n->del_req(5);
2321 n->del_req(4);
2322 break;
2323 }
2324 case Op_EncodeISOArray: {
2325 // Restructure into a binary tree for Matching.
2326 Node* pair = new (C) BinaryNode(n->in(3), n->in(4));
2327 n->set_req(3, pair);
2328 n->del_req(4);
2329 break;
2330 }
2331 default:
2332 break;
2333 }
2334 }
2335 else {
2336 ShouldNotReachHere();
2337 }
2338 } // end of while (mstack.is_nonempty())
2339 }
2341 #ifdef ASSERT
2342 // machine-independent root to machine-dependent root
2343 void Matcher::dump_old2new_map() {
2344 _old2new_map.dump();
2345 }
2346 #endif
2348 //---------------------------collect_null_checks-------------------------------
2349 // Find null checks in the ideal graph; write a machine-specific node for
2350 // it. Used by later implicit-null-check handling. Actually collects
2351 // either an IfTrue or IfFalse for the common NOT-null path, AND the ideal
2352 // value being tested.
2353 void Matcher::collect_null_checks( Node *proj, Node *orig_proj ) {
2354 Node *iff = proj->in(0);
2355 if( iff->Opcode() == Op_If ) {
2356 // During matching If's have Bool & Cmp side-by-side
2357 BoolNode *b = iff->in(1)->as_Bool();
2358 Node *cmp = iff->in(2);
2359 int opc = cmp->Opcode();
2360 if (opc != Op_CmpP && opc != Op_CmpN) return;
2362 const Type* ct = cmp->in(2)->bottom_type();
2363 if (ct == TypePtr::NULL_PTR ||
2364 (opc == Op_CmpN && ct == TypeNarrowOop::NULL_PTR)) {
2366 bool push_it = false;
2367 if( proj->Opcode() == Op_IfTrue ) {
2368 extern int all_null_checks_found;
2369 all_null_checks_found++;
2370 if( b->_test._test == BoolTest::ne ) {
2371 push_it = true;
2372 }
2373 } else {
2374 assert( proj->Opcode() == Op_IfFalse, "" );
2375 if( b->_test._test == BoolTest::eq ) {
2376 push_it = true;
2377 }
2378 }
2379 if( push_it ) {
2380 _null_check_tests.push(proj);
2381 Node* val = cmp->in(1);
2382 #ifdef _LP64
2383 if (val->bottom_type()->isa_narrowoop() &&
2384 !Matcher::narrow_oop_use_complex_address()) {
2385 //
2386 // Look for DecodeN node which should be pinned to orig_proj.
2387 // On platforms (Sparc) which can not handle 2 adds
2388 // in addressing mode we have to keep a DecodeN node and
2389 // use it to do implicit NULL check in address.
2390 //
2391 // DecodeN node was pinned to non-null path (orig_proj) during
2392 // CastPP transformation in final_graph_reshaping_impl().
2393 //
2394 uint cnt = orig_proj->outcnt();
2395 for (uint i = 0; i < orig_proj->outcnt(); i++) {
2396 Node* d = orig_proj->raw_out(i);
2397 if (d->is_DecodeN() && d->in(1) == val) {
2398 val = d;
2399 val->set_req(0, NULL); // Unpin now.
2400 // Mark this as special case to distinguish from
2401 // a regular case: CmpP(DecodeN, NULL).
2402 val = (Node*)(((intptr_t)val) | 1);
2403 break;
2404 }
2405 }
2406 }
2407 #endif
2408 _null_check_tests.push(val);
2409 }
2410 }
2411 }
2412 }
2414 //---------------------------validate_null_checks------------------------------
2415 // Its possible that the value being NULL checked is not the root of a match
2416 // tree. If so, I cannot use the value in an implicit null check.
2417 void Matcher::validate_null_checks( ) {
2418 uint cnt = _null_check_tests.size();
2419 for( uint i=0; i < cnt; i+=2 ) {
2420 Node *test = _null_check_tests[i];
2421 Node *val = _null_check_tests[i+1];
2422 bool is_decoden = ((intptr_t)val) & 1;
2423 val = (Node*)(((intptr_t)val) & ~1);
2424 if (has_new_node(val)) {
2425 Node* new_val = new_node(val);
2426 if (is_decoden) {
2427 assert(val->is_DecodeNarrowPtr() && val->in(0) == NULL, "sanity");
2428 // Note: new_val may have a control edge if
2429 // the original ideal node DecodeN was matched before
2430 // it was unpinned in Matcher::collect_null_checks().
2431 // Unpin the mach node and mark it.
2432 new_val->set_req(0, NULL);
2433 new_val = (Node*)(((intptr_t)new_val) | 1);
2434 }
2435 // Is a match-tree root, so replace with the matched value
2436 _null_check_tests.map(i+1, new_val);
2437 } else {
2438 // Yank from candidate list
2439 _null_check_tests.map(i+1,_null_check_tests[--cnt]);
2440 _null_check_tests.map(i,_null_check_tests[--cnt]);
2441 _null_check_tests.pop();
2442 _null_check_tests.pop();
2443 i-=2;
2444 }
2445 }
2446 }
2448 // Used by the DFA in dfa_xxx.cpp. Check for a following barrier or
2449 // atomic instruction acting as a store_load barrier without any
2450 // intervening volatile load, and thus we don't need a barrier here.
2451 // We retain the Node to act as a compiler ordering barrier.
2452 bool Matcher::post_store_load_barrier(const Node* vmb) {
2453 Compile* C = Compile::current();
2454 assert(vmb->is_MemBar(), "");
2455 assert(vmb->Opcode() != Op_MemBarAcquire && vmb->Opcode() != Op_LoadFence, "");
2456 const MemBarNode* membar = vmb->as_MemBar();
2458 // Get the Ideal Proj node, ctrl, that can be used to iterate forward
2459 Node* ctrl = NULL;
2460 for (DUIterator_Fast imax, i = membar->fast_outs(imax); i < imax; i++) {
2461 Node* p = membar->fast_out(i);
2462 assert(p->is_Proj(), "only projections here");
2463 if ((p->as_Proj()->_con == TypeFunc::Control) &&
2464 !C->node_arena()->contains(p)) { // Unmatched old-space only
2465 ctrl = p;
2466 break;
2467 }
2468 }
2469 assert((ctrl != NULL), "missing control projection");
2471 for (DUIterator_Fast jmax, j = ctrl->fast_outs(jmax); j < jmax; j++) {
2472 Node *x = ctrl->fast_out(j);
2473 int xop = x->Opcode();
2475 // We don't need current barrier if we see another or a lock
2476 // before seeing volatile load.
2477 //
2478 // Op_Fastunlock previously appeared in the Op_* list below.
2479 // With the advent of 1-0 lock operations we're no longer guaranteed
2480 // that a monitor exit operation contains a serializing instruction.
2482 if (xop == Op_MemBarVolatile ||
2483 xop == Op_CompareAndSwapL ||
2484 xop == Op_CompareAndSwapP ||
2485 xop == Op_CompareAndSwapN ||
2486 xop == Op_CompareAndSwapI) {
2487 return true;
2488 }
2490 // Op_FastLock previously appeared in the Op_* list above.
2491 // With biased locking we're no longer guaranteed that a monitor
2492 // enter operation contains a serializing instruction.
2493 if ((xop == Op_FastLock) && !UseBiasedLocking) {
2494 return true;
2495 }
2497 if (x->is_MemBar()) {
2498 // We must retain this membar if there is an upcoming volatile
2499 // load, which will be followed by acquire membar.
2500 if (xop == Op_MemBarAcquire || xop == Op_LoadFence) {
2501 return false;
2502 } else {
2503 // For other kinds of barriers, check by pretending we
2504 // are them, and seeing if we can be removed.
2505 return post_store_load_barrier(x->as_MemBar());
2506 }
2507 }
2509 // probably not necessary to check for these
2510 if (x->is_Call() || x->is_SafePoint() || x->is_block_proj()) {
2511 return false;
2512 }
2513 }
2514 return false;
2515 }
2517 // Check whether node n is a branch to an uncommon trap that we could
2518 // optimize as test with very high branch costs in case of going to
2519 // the uncommon trap. The code must be able to be recompiled to use
2520 // a cheaper test.
2521 bool Matcher::branches_to_uncommon_trap(const Node *n) {
2522 // Don't do it for natives, adapters, or runtime stubs
2523 Compile *C = Compile::current();
2524 if (!C->is_method_compilation()) return false;
2526 assert(n->is_If(), "You should only call this on if nodes.");
2527 IfNode *ifn = n->as_If();
2529 Node *ifFalse = NULL;
2530 for (DUIterator_Fast imax, i = ifn->fast_outs(imax); i < imax; i++) {
2531 if (ifn->fast_out(i)->is_IfFalse()) {
2532 ifFalse = ifn->fast_out(i);
2533 break;
2534 }
2535 }
2536 assert(ifFalse, "An If should have an ifFalse. Graph is broken.");
2538 Node *reg = ifFalse;
2539 int cnt = 4; // We must protect against cycles. Limit to 4 iterations.
2540 // Alternatively use visited set? Seems too expensive.
2541 while (reg != NULL && cnt > 0) {
2542 CallNode *call = NULL;
2543 RegionNode *nxt_reg = NULL;
2544 for (DUIterator_Fast imax, i = reg->fast_outs(imax); i < imax; i++) {
2545 Node *o = reg->fast_out(i);
2546 if (o->is_Call()) {
2547 call = o->as_Call();
2548 }
2549 if (o->is_Region()) {
2550 nxt_reg = o->as_Region();
2551 }
2552 }
2554 if (call &&
2555 call->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) {
2556 const Type* trtype = call->in(TypeFunc::Parms)->bottom_type();
2557 if (trtype->isa_int() && trtype->is_int()->is_con()) {
2558 jint tr_con = trtype->is_int()->get_con();
2559 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
2560 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
2561 assert((int)reason < (int)BitsPerInt, "recode bit map");
2563 if (is_set_nth_bit(C->allowed_deopt_reasons(), (int)reason)
2564 && action != Deoptimization::Action_none) {
2565 // This uncommon trap is sure to recompile, eventually.
2566 // When that happens, C->too_many_traps will prevent
2567 // this transformation from happening again.
2568 return true;
2569 }
2570 }
2571 }
2573 reg = nxt_reg;
2574 cnt--;
2575 }
2577 return false;
2578 }
2580 //=============================================================================
2581 //---------------------------State---------------------------------------------
2582 State::State(void) {
2583 #ifdef ASSERT
2584 _id = 0;
2585 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe);
2586 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d);
2587 //memset(_cost, -1, sizeof(_cost));
2588 //memset(_rule, -1, sizeof(_rule));
2589 #endif
2590 memset(_valid, 0, sizeof(_valid));
2591 }
2593 #ifdef ASSERT
2594 State::~State() {
2595 _id = 99;
2596 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe);
2597 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d);
2598 memset(_cost, -3, sizeof(_cost));
2599 memset(_rule, -3, sizeof(_rule));
2600 }
2601 #endif
2603 #ifndef PRODUCT
2604 //---------------------------dump----------------------------------------------
2605 void State::dump() {
2606 tty->print("\n");
2607 dump(0);
2608 }
2610 void State::dump(int depth) {
2611 for( int j = 0; j < depth; j++ )
2612 tty->print(" ");
2613 tty->print("--N: ");
2614 _leaf->dump();
2615 uint i;
2616 for( i = 0; i < _LAST_MACH_OPER; i++ )
2617 // Check for valid entry
2618 if( valid(i) ) {
2619 for( int j = 0; j < depth; j++ )
2620 tty->print(" ");
2621 assert(_cost[i] != max_juint, "cost must be a valid value");
2622 assert(_rule[i] < _last_Mach_Node, "rule[i] must be valid rule");
2623 tty->print_cr("%s %d %s",
2624 ruleName[i], _cost[i], ruleName[_rule[i]] );
2625 }
2626 tty->cr();
2628 for( i=0; i<2; i++ )
2629 if( _kids[i] )
2630 _kids[i]->dump(depth+1);
2631 }
2632 #endif