Mon, 26 Sep 2011 10:24:05 -0700
7081933: Use zeroing elimination optimization for large array
Summary: Don't zero new typeArray during runtime call if the allocation is followed by arraycopy into it.
Reviewed-by: twisti
1 /*
2 * Copyright (c) 1997, 2011, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "memory/allocation.inline.hpp"
27 #include "opto/addnode.hpp"
28 #include "opto/callnode.hpp"
29 #include "opto/connode.hpp"
30 #include "opto/idealGraphPrinter.hpp"
31 #include "opto/matcher.hpp"
32 #include "opto/memnode.hpp"
33 #include "opto/opcodes.hpp"
34 #include "opto/regmask.hpp"
35 #include "opto/rootnode.hpp"
36 #include "opto/runtime.hpp"
37 #include "opto/type.hpp"
38 #include "runtime/atomic.hpp"
39 #include "runtime/os.hpp"
40 #ifdef TARGET_ARCH_MODEL_x86_32
41 # include "adfiles/ad_x86_32.hpp"
42 #endif
43 #ifdef TARGET_ARCH_MODEL_x86_64
44 # include "adfiles/ad_x86_64.hpp"
45 #endif
46 #ifdef TARGET_ARCH_MODEL_sparc
47 # include "adfiles/ad_sparc.hpp"
48 #endif
49 #ifdef TARGET_ARCH_MODEL_zero
50 # include "adfiles/ad_zero.hpp"
51 #endif
52 #ifdef TARGET_ARCH_MODEL_arm
53 # include "adfiles/ad_arm.hpp"
54 #endif
55 #ifdef TARGET_ARCH_MODEL_ppc
56 # include "adfiles/ad_ppc.hpp"
57 #endif
59 OptoReg::Name OptoReg::c_frame_pointer;
63 const int Matcher::base2reg[Type::lastype] = {
64 Node::NotAMachineReg,0,0, Op_RegI, Op_RegL, 0, Op_RegN,
65 Node::NotAMachineReg, Node::NotAMachineReg, /* tuple, array */
66 Op_RegP, Op_RegP, Op_RegP, Op_RegP, Op_RegP, Op_RegP, /* the pointers */
67 0, 0/*abio*/,
68 Op_RegP /* Return address */, 0, /* the memories */
69 Op_RegF, Op_RegF, Op_RegF, Op_RegD, Op_RegD, Op_RegD,
70 0 /*bottom*/
71 };
73 const RegMask *Matcher::idealreg2regmask[_last_machine_leaf];
74 RegMask Matcher::mreg2regmask[_last_Mach_Reg];
75 RegMask Matcher::STACK_ONLY_mask;
76 RegMask Matcher::c_frame_ptr_mask;
77 const uint Matcher::_begin_rematerialize = _BEGIN_REMATERIALIZE;
78 const uint Matcher::_end_rematerialize = _END_REMATERIALIZE;
80 //---------------------------Matcher-------------------------------------------
81 Matcher::Matcher( Node_List &proj_list ) :
82 PhaseTransform( Phase::Ins_Select ),
83 #ifdef ASSERT
84 _old2new_map(C->comp_arena()),
85 _new2old_map(C->comp_arena()),
86 #endif
87 _shared_nodes(C->comp_arena()),
88 _reduceOp(reduceOp), _leftOp(leftOp), _rightOp(rightOp),
89 _swallowed(swallowed),
90 _begin_inst_chain_rule(_BEGIN_INST_CHAIN_RULE),
91 _end_inst_chain_rule(_END_INST_CHAIN_RULE),
92 _must_clone(must_clone), _proj_list(proj_list),
93 _register_save_policy(register_save_policy),
94 _c_reg_save_policy(c_reg_save_policy),
95 _register_save_type(register_save_type),
96 _ruleName(ruleName),
97 _allocation_started(false),
98 _states_arena(Chunk::medium_size),
99 _visited(&_states_arena),
100 _shared(&_states_arena),
101 _dontcare(&_states_arena) {
102 C->set_matcher(this);
104 idealreg2spillmask [Op_RegI] = NULL;
105 idealreg2spillmask [Op_RegN] = NULL;
106 idealreg2spillmask [Op_RegL] = NULL;
107 idealreg2spillmask [Op_RegF] = NULL;
108 idealreg2spillmask [Op_RegD] = NULL;
109 idealreg2spillmask [Op_RegP] = NULL;
111 idealreg2debugmask [Op_RegI] = NULL;
112 idealreg2debugmask [Op_RegN] = NULL;
113 idealreg2debugmask [Op_RegL] = NULL;
114 idealreg2debugmask [Op_RegF] = NULL;
115 idealreg2debugmask [Op_RegD] = NULL;
116 idealreg2debugmask [Op_RegP] = 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;
125 debug_only(_mem_node = NULL;) // Ideal memory node consumed by mach node
126 }
128 //------------------------------warp_incoming_stk_arg------------------------
129 // This warps a VMReg into an OptoReg::Name
130 OptoReg::Name Matcher::warp_incoming_stk_arg( VMReg reg ) {
131 OptoReg::Name warped;
132 if( reg->is_stack() ) { // Stack slot argument?
133 warped = OptoReg::add(_old_SP, reg->reg2stack() );
134 warped = OptoReg::add(warped, C->out_preserve_stack_slots());
135 if( warped >= _in_arg_limit )
136 _in_arg_limit = OptoReg::add(warped, 1); // Bump max stack slot seen
137 if (!RegMask::can_represent(warped)) {
138 // the compiler cannot represent this method's calling sequence
139 C->record_method_not_compilable_all_tiers("unsupported incoming calling sequence");
140 return OptoReg::Bad;
141 }
142 return warped;
143 }
144 return OptoReg::as_OptoReg(reg);
145 }
147 //---------------------------compute_old_SP------------------------------------
148 OptoReg::Name Compile::compute_old_SP() {
149 int fixed = fixed_slots();
150 int preserve = in_preserve_stack_slots();
151 return OptoReg::stack2reg(round_to(fixed + preserve, Matcher::stack_alignment_in_slots()));
152 }
156 #ifdef ASSERT
157 void Matcher::verify_new_nodes_only(Node* xroot) {
158 // Make sure that the new graph only references new nodes
159 ResourceMark rm;
160 Unique_Node_List worklist;
161 VectorSet visited(Thread::current()->resource_area());
162 worklist.push(xroot);
163 while (worklist.size() > 0) {
164 Node* n = worklist.pop();
165 visited <<= n->_idx;
166 assert(C->node_arena()->contains(n), "dead node");
167 for (uint j = 0; j < n->req(); j++) {
168 Node* in = n->in(j);
169 if (in != NULL) {
170 assert(C->node_arena()->contains(in), "dead node");
171 if (!visited.test(in->_idx)) {
172 worklist.push(in);
173 }
174 }
175 }
176 }
177 }
178 #endif
181 //---------------------------match---------------------------------------------
182 void Matcher::match( ) {
183 if( MaxLabelRootDepth < 100 ) { // Too small?
184 assert(false, "invalid MaxLabelRootDepth, increase it to 100 minimum");
185 MaxLabelRootDepth = 100;
186 }
187 // One-time initialization of some register masks.
188 init_spill_mask( C->root()->in(1) );
189 _return_addr_mask = return_addr();
190 #ifdef _LP64
191 // Pointers take 2 slots in 64-bit land
192 _return_addr_mask.Insert(OptoReg::add(return_addr(),1));
193 #endif
195 // Map a Java-signature return type into return register-value
196 // machine registers for 0, 1 and 2 returned values.
197 const TypeTuple *range = C->tf()->range();
198 if( range->cnt() > TypeFunc::Parms ) { // If not a void function
199 // Get ideal-register return type
200 int ireg = base2reg[range->field_at(TypeFunc::Parms)->base()];
201 // Get machine return register
202 uint sop = C->start()->Opcode();
203 OptoRegPair regs = return_value(ireg, false);
205 // And mask for same
206 _return_value_mask = RegMask(regs.first());
207 if( OptoReg::is_valid(regs.second()) )
208 _return_value_mask.Insert(regs.second());
209 }
211 // ---------------
212 // Frame Layout
214 // Need the method signature to determine the incoming argument types,
215 // because the types determine which registers the incoming arguments are
216 // in, and this affects the matched code.
217 const TypeTuple *domain = C->tf()->domain();
218 uint argcnt = domain->cnt() - TypeFunc::Parms;
219 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt );
220 VMRegPair *vm_parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt );
221 _parm_regs = NEW_RESOURCE_ARRAY( OptoRegPair, argcnt );
222 _calling_convention_mask = NEW_RESOURCE_ARRAY( RegMask, argcnt );
223 uint i;
224 for( i = 0; i<argcnt; i++ ) {
225 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type();
226 }
228 // Pass array of ideal registers and length to USER code (from the AD file)
229 // that will convert this to an array of register numbers.
230 const StartNode *start = C->start();
231 start->calling_convention( sig_bt, vm_parm_regs, argcnt );
232 #ifdef ASSERT
233 // Sanity check users' calling convention. Real handy while trying to
234 // get the initial port correct.
235 { for (uint i = 0; i<argcnt; i++) {
236 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) {
237 assert(domain->field_at(i+TypeFunc::Parms)==Type::HALF, "only allowed on halve" );
238 _parm_regs[i].set_bad();
239 continue;
240 }
241 VMReg parm_reg = vm_parm_regs[i].first();
242 assert(parm_reg->is_valid(), "invalid arg?");
243 if (parm_reg->is_reg()) {
244 OptoReg::Name opto_parm_reg = OptoReg::as_OptoReg(parm_reg);
245 assert(can_be_java_arg(opto_parm_reg) ||
246 C->stub_function() == CAST_FROM_FN_PTR(address, OptoRuntime::rethrow_C) ||
247 opto_parm_reg == inline_cache_reg(),
248 "parameters in register must be preserved by runtime stubs");
249 }
250 for (uint j = 0; j < i; j++) {
251 assert(parm_reg != vm_parm_regs[j].first(),
252 "calling conv. must produce distinct regs");
253 }
254 }
255 }
256 #endif
258 // Do some initial frame layout.
260 // Compute the old incoming SP (may be called FP) as
261 // OptoReg::stack0() + locks + in_preserve_stack_slots + pad2.
262 _old_SP = C->compute_old_SP();
263 assert( is_even(_old_SP), "must be even" );
265 // Compute highest incoming stack argument as
266 // _old_SP + out_preserve_stack_slots + incoming argument size.
267 _in_arg_limit = OptoReg::add(_old_SP, C->out_preserve_stack_slots());
268 assert( is_even(_in_arg_limit), "out_preserve must be even" );
269 for( i = 0; i < argcnt; i++ ) {
270 // Permit args to have no register
271 _calling_convention_mask[i].Clear();
272 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) {
273 continue;
274 }
275 // calling_convention returns stack arguments as a count of
276 // slots beyond OptoReg::stack0()/VMRegImpl::stack0. We need to convert this to
277 // the allocators point of view, taking into account all the
278 // preserve area, locks & pad2.
280 OptoReg::Name reg1 = warp_incoming_stk_arg(vm_parm_regs[i].first());
281 if( OptoReg::is_valid(reg1))
282 _calling_convention_mask[i].Insert(reg1);
284 OptoReg::Name reg2 = warp_incoming_stk_arg(vm_parm_regs[i].second());
285 if( OptoReg::is_valid(reg2))
286 _calling_convention_mask[i].Insert(reg2);
288 // Saved biased stack-slot register number
289 _parm_regs[i].set_pair(reg2, reg1);
290 }
292 // Finally, make sure the incoming arguments take up an even number of
293 // words, in case the arguments or locals need to contain doubleword stack
294 // slots. The rest of the system assumes that stack slot pairs (in
295 // particular, in the spill area) which look aligned will in fact be
296 // aligned relative to the stack pointer in the target machine. Double
297 // stack slots will always be allocated aligned.
298 _new_SP = OptoReg::Name(round_to(_in_arg_limit, RegMask::SlotsPerLong));
300 // Compute highest outgoing stack argument as
301 // _new_SP + out_preserve_stack_slots + max(outgoing argument size).
302 _out_arg_limit = OptoReg::add(_new_SP, C->out_preserve_stack_slots());
303 assert( is_even(_out_arg_limit), "out_preserve must be even" );
305 if (!RegMask::can_represent(OptoReg::add(_out_arg_limit,-1))) {
306 // the compiler cannot represent this method's calling sequence
307 C->record_method_not_compilable("must be able to represent all call arguments in reg mask");
308 }
310 if (C->failing()) return; // bailed out on incoming arg failure
312 // ---------------
313 // Collect roots of matcher trees. Every node for which
314 // _shared[_idx] is cleared is guaranteed to not be shared, and thus
315 // can be a valid interior of some tree.
316 find_shared( C->root() );
317 find_shared( C->top() );
319 C->print_method("Before Matching");
321 // Create new ideal node ConP #NULL even if it does exist in old space
322 // to avoid false sharing if the corresponding mach node is not used.
323 // The corresponding mach node is only used in rare cases for derived
324 // pointers.
325 Node* new_ideal_null = ConNode::make(C, TypePtr::NULL_PTR);
327 // Swap out to old-space; emptying new-space
328 Arena *old = C->node_arena()->move_contents(C->old_arena());
330 // Save debug and profile information for nodes in old space:
331 _old_node_note_array = C->node_note_array();
332 if (_old_node_note_array != NULL) {
333 C->set_node_note_array(new(C->comp_arena()) GrowableArray<Node_Notes*>
334 (C->comp_arena(), _old_node_note_array->length(),
335 0, NULL));
336 }
338 // Pre-size the new_node table to avoid the need for range checks.
339 grow_new_node_array(C->unique());
341 // Reset node counter so MachNodes start with _idx at 0
342 int nodes = C->unique(); // save value
343 C->set_unique(0);
345 // Recursively match trees from old space into new space.
346 // Correct leaves of new-space Nodes; they point to old-space.
347 _visited.Clear(); // Clear visit bits for xform call
348 C->set_cached_top_node(xform( C->top(), nodes ));
349 if (!C->failing()) {
350 Node* xroot = xform( C->root(), 1 );
351 if (xroot == NULL) {
352 Matcher::soft_match_failure(); // recursive matching process failed
353 C->record_method_not_compilable("instruction match failed");
354 } else {
355 // During matching shared constants were attached to C->root()
356 // because xroot wasn't available yet, so transfer the uses to
357 // the xroot.
358 for( DUIterator_Fast jmax, j = C->root()->fast_outs(jmax); j < jmax; j++ ) {
359 Node* n = C->root()->fast_out(j);
360 if (C->node_arena()->contains(n)) {
361 assert(n->in(0) == C->root(), "should be control user");
362 n->set_req(0, xroot);
363 --j;
364 --jmax;
365 }
366 }
368 // Generate new mach node for ConP #NULL
369 assert(new_ideal_null != NULL, "sanity");
370 _mach_null = match_tree(new_ideal_null);
371 // Don't set control, it will confuse GCM since there are no uses.
372 // The control will be set when this node is used first time
373 // in find_base_for_derived().
374 assert(_mach_null != NULL, "");
376 C->set_root(xroot->is_Root() ? xroot->as_Root() : NULL);
378 #ifdef ASSERT
379 verify_new_nodes_only(xroot);
380 #endif
381 }
382 }
383 if (C->top() == NULL || C->root() == NULL) {
384 C->record_method_not_compilable("graph lost"); // %%% cannot happen?
385 }
386 if (C->failing()) {
387 // delete old;
388 old->destruct_contents();
389 return;
390 }
391 assert( C->top(), "" );
392 assert( C->root(), "" );
393 validate_null_checks();
395 // Now smoke old-space
396 NOT_DEBUG( old->destruct_contents() );
398 // ------------------------
399 // Set up save-on-entry registers
400 Fixup_Save_On_Entry( );
401 }
404 //------------------------------Fixup_Save_On_Entry----------------------------
405 // The stated purpose of this routine is to take care of save-on-entry
406 // registers. However, the overall goal of the Match phase is to convert into
407 // machine-specific instructions which have RegMasks to guide allocation.
408 // So what this procedure really does is put a valid RegMask on each input
409 // to the machine-specific variations of all Return, TailCall and Halt
410 // instructions. It also adds edgs to define the save-on-entry values (and of
411 // course gives them a mask).
413 static RegMask *init_input_masks( uint size, RegMask &ret_adr, RegMask &fp ) {
414 RegMask *rms = NEW_RESOURCE_ARRAY( RegMask, size );
415 // Do all the pre-defined register masks
416 rms[TypeFunc::Control ] = RegMask::Empty;
417 rms[TypeFunc::I_O ] = RegMask::Empty;
418 rms[TypeFunc::Memory ] = RegMask::Empty;
419 rms[TypeFunc::ReturnAdr] = ret_adr;
420 rms[TypeFunc::FramePtr ] = fp;
421 return rms;
422 }
424 //---------------------------init_first_stack_mask-----------------------------
425 // Create the initial stack mask used by values spilling to the stack.
426 // Disallow any debug info in outgoing argument areas by setting the
427 // initial mask accordingly.
428 void Matcher::init_first_stack_mask() {
430 // Allocate storage for spill masks as masks for the appropriate load type.
431 RegMask *rms = (RegMask*)C->comp_arena()->Amalloc_D(sizeof(RegMask) * 3*6);
433 idealreg2spillmask [Op_RegN] = &rms[0];
434 idealreg2spillmask [Op_RegI] = &rms[1];
435 idealreg2spillmask [Op_RegL] = &rms[2];
436 idealreg2spillmask [Op_RegF] = &rms[3];
437 idealreg2spillmask [Op_RegD] = &rms[4];
438 idealreg2spillmask [Op_RegP] = &rms[5];
440 idealreg2debugmask [Op_RegN] = &rms[6];
441 idealreg2debugmask [Op_RegI] = &rms[7];
442 idealreg2debugmask [Op_RegL] = &rms[8];
443 idealreg2debugmask [Op_RegF] = &rms[9];
444 idealreg2debugmask [Op_RegD] = &rms[10];
445 idealreg2debugmask [Op_RegP] = &rms[11];
447 idealreg2mhdebugmask[Op_RegN] = &rms[12];
448 idealreg2mhdebugmask[Op_RegI] = &rms[13];
449 idealreg2mhdebugmask[Op_RegL] = &rms[14];
450 idealreg2mhdebugmask[Op_RegF] = &rms[15];
451 idealreg2mhdebugmask[Op_RegD] = &rms[16];
452 idealreg2mhdebugmask[Op_RegP] = &rms[17];
454 OptoReg::Name i;
456 // At first, start with the empty mask
457 C->FIRST_STACK_mask().Clear();
459 // Add in the incoming argument area
460 OptoReg::Name init = OptoReg::add(_old_SP, C->out_preserve_stack_slots());
461 for (i = init; i < _in_arg_limit; i = OptoReg::add(i,1))
462 C->FIRST_STACK_mask().Insert(i);
464 // Add in all bits past the outgoing argument area
465 guarantee(RegMask::can_represent(OptoReg::add(_out_arg_limit,-1)),
466 "must be able to represent all call arguments in reg mask");
467 init = _out_arg_limit;
468 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1))
469 C->FIRST_STACK_mask().Insert(i);
471 // Finally, set the "infinite stack" bit.
472 C->FIRST_STACK_mask().set_AllStack();
474 // Make spill masks. Registers for their class, plus FIRST_STACK_mask.
475 #ifdef _LP64
476 *idealreg2spillmask[Op_RegN] = *idealreg2regmask[Op_RegN];
477 idealreg2spillmask[Op_RegN]->OR(C->FIRST_STACK_mask());
478 #endif
479 *idealreg2spillmask[Op_RegI] = *idealreg2regmask[Op_RegI];
480 idealreg2spillmask[Op_RegI]->OR(C->FIRST_STACK_mask());
481 *idealreg2spillmask[Op_RegL] = *idealreg2regmask[Op_RegL];
482 idealreg2spillmask[Op_RegL]->OR(C->FIRST_STACK_mask());
483 *idealreg2spillmask[Op_RegF] = *idealreg2regmask[Op_RegF];
484 idealreg2spillmask[Op_RegF]->OR(C->FIRST_STACK_mask());
485 *idealreg2spillmask[Op_RegD] = *idealreg2regmask[Op_RegD];
486 idealreg2spillmask[Op_RegD]->OR(C->FIRST_STACK_mask());
487 *idealreg2spillmask[Op_RegP] = *idealreg2regmask[Op_RegP];
488 idealreg2spillmask[Op_RegP]->OR(C->FIRST_STACK_mask());
490 if (UseFPUForSpilling) {
491 // This mask logic assumes that the spill operations are
492 // symmetric and that the registers involved are the same size.
493 // On sparc for instance we may have to use 64 bit moves will
494 // kill 2 registers when used with F0-F31.
495 idealreg2spillmask[Op_RegI]->OR(*idealreg2regmask[Op_RegF]);
496 idealreg2spillmask[Op_RegF]->OR(*idealreg2regmask[Op_RegI]);
497 #ifdef _LP64
498 idealreg2spillmask[Op_RegN]->OR(*idealreg2regmask[Op_RegF]);
499 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]);
500 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]);
501 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegD]);
502 #else
503 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegF]);
504 #ifdef ARM
505 // ARM has support for moving 64bit values between a pair of
506 // integer registers and a double register
507 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]);
508 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]);
509 #endif
510 #endif
511 }
513 // Make up debug masks. Any spill slot plus callee-save registers.
514 // Caller-save registers are assumed to be trashable by the various
515 // inline-cache fixup routines.
516 *idealreg2debugmask [Op_RegN]= *idealreg2spillmask[Op_RegN];
517 *idealreg2debugmask [Op_RegI]= *idealreg2spillmask[Op_RegI];
518 *idealreg2debugmask [Op_RegL]= *idealreg2spillmask[Op_RegL];
519 *idealreg2debugmask [Op_RegF]= *idealreg2spillmask[Op_RegF];
520 *idealreg2debugmask [Op_RegD]= *idealreg2spillmask[Op_RegD];
521 *idealreg2debugmask [Op_RegP]= *idealreg2spillmask[Op_RegP];
523 *idealreg2mhdebugmask[Op_RegN]= *idealreg2spillmask[Op_RegN];
524 *idealreg2mhdebugmask[Op_RegI]= *idealreg2spillmask[Op_RegI];
525 *idealreg2mhdebugmask[Op_RegL]= *idealreg2spillmask[Op_RegL];
526 *idealreg2mhdebugmask[Op_RegF]= *idealreg2spillmask[Op_RegF];
527 *idealreg2mhdebugmask[Op_RegD]= *idealreg2spillmask[Op_RegD];
528 *idealreg2mhdebugmask[Op_RegP]= *idealreg2spillmask[Op_RegP];
530 // Prevent stub compilations from attempting to reference
531 // callee-saved registers from debug info
532 bool exclude_soe = !Compile::current()->is_method_compilation();
534 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) {
535 // registers the caller has to save do not work
536 if( _register_save_policy[i] == 'C' ||
537 _register_save_policy[i] == 'A' ||
538 (_register_save_policy[i] == 'E' && exclude_soe) ) {
539 idealreg2debugmask [Op_RegN]->Remove(i);
540 idealreg2debugmask [Op_RegI]->Remove(i); // Exclude save-on-call
541 idealreg2debugmask [Op_RegL]->Remove(i); // registers from debug
542 idealreg2debugmask [Op_RegF]->Remove(i); // masks
543 idealreg2debugmask [Op_RegD]->Remove(i);
544 idealreg2debugmask [Op_RegP]->Remove(i);
546 idealreg2mhdebugmask[Op_RegN]->Remove(i);
547 idealreg2mhdebugmask[Op_RegI]->Remove(i);
548 idealreg2mhdebugmask[Op_RegL]->Remove(i);
549 idealreg2mhdebugmask[Op_RegF]->Remove(i);
550 idealreg2mhdebugmask[Op_RegD]->Remove(i);
551 idealreg2mhdebugmask[Op_RegP]->Remove(i);
552 }
553 }
555 // Subtract the register we use to save the SP for MethodHandle
556 // invokes to from the debug mask.
557 const RegMask save_mask = method_handle_invoke_SP_save_mask();
558 idealreg2mhdebugmask[Op_RegN]->SUBTRACT(save_mask);
559 idealreg2mhdebugmask[Op_RegI]->SUBTRACT(save_mask);
560 idealreg2mhdebugmask[Op_RegL]->SUBTRACT(save_mask);
561 idealreg2mhdebugmask[Op_RegF]->SUBTRACT(save_mask);
562 idealreg2mhdebugmask[Op_RegD]->SUBTRACT(save_mask);
563 idealreg2mhdebugmask[Op_RegP]->SUBTRACT(save_mask);
564 }
566 //---------------------------is_save_on_entry----------------------------------
567 bool Matcher::is_save_on_entry( int reg ) {
568 return
569 _register_save_policy[reg] == 'E' ||
570 _register_save_policy[reg] == 'A' || // Save-on-entry register?
571 // Also save argument registers in the trampolining stubs
572 (C->save_argument_registers() && is_spillable_arg(reg));
573 }
575 //---------------------------Fixup_Save_On_Entry-------------------------------
576 void Matcher::Fixup_Save_On_Entry( ) {
577 init_first_stack_mask();
579 Node *root = C->root(); // Short name for root
580 // Count number of save-on-entry registers.
581 uint soe_cnt = number_of_saved_registers();
582 uint i;
584 // Find the procedure Start Node
585 StartNode *start = C->start();
586 assert( start, "Expect a start node" );
588 // Save argument registers in the trampolining stubs
589 if( C->save_argument_registers() )
590 for( i = 0; i < _last_Mach_Reg; i++ )
591 if( is_spillable_arg(i) )
592 soe_cnt++;
594 // Input RegMask array shared by all Returns.
595 // The type for doubles and longs has a count of 2, but
596 // there is only 1 returned value
597 uint ret_edge_cnt = TypeFunc::Parms + ((C->tf()->range()->cnt() == TypeFunc::Parms) ? 0 : 1);
598 RegMask *ret_rms = init_input_masks( ret_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
599 // Returns have 0 or 1 returned values depending on call signature.
600 // Return register is specified by return_value in the AD file.
601 if (ret_edge_cnt > TypeFunc::Parms)
602 ret_rms[TypeFunc::Parms+0] = _return_value_mask;
604 // Input RegMask array shared by all Rethrows.
605 uint reth_edge_cnt = TypeFunc::Parms+1;
606 RegMask *reth_rms = init_input_masks( reth_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
607 // Rethrow takes exception oop only, but in the argument 0 slot.
608 reth_rms[TypeFunc::Parms] = mreg2regmask[find_receiver(false)];
609 #ifdef _LP64
610 // Need two slots for ptrs in 64-bit land
611 reth_rms[TypeFunc::Parms].Insert(OptoReg::add(OptoReg::Name(find_receiver(false)),1));
612 #endif
614 // Input RegMask array shared by all TailCalls
615 uint tail_call_edge_cnt = TypeFunc::Parms+2;
616 RegMask *tail_call_rms = init_input_masks( tail_call_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
618 // Input RegMask array shared by all TailJumps
619 uint tail_jump_edge_cnt = TypeFunc::Parms+2;
620 RegMask *tail_jump_rms = init_input_masks( tail_jump_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
622 // TailCalls have 2 returned values (target & moop), whose masks come
623 // from the usual MachNode/MachOper mechanism. Find a sample
624 // TailCall to extract these masks and put the correct masks into
625 // the tail_call_rms array.
626 for( i=1; i < root->req(); i++ ) {
627 MachReturnNode *m = root->in(i)->as_MachReturn();
628 if( m->ideal_Opcode() == Op_TailCall ) {
629 tail_call_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0);
630 tail_call_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1);
631 break;
632 }
633 }
635 // TailJumps have 2 returned values (target & ex_oop), whose masks come
636 // from the usual MachNode/MachOper mechanism. Find a sample
637 // TailJump to extract these masks and put the correct masks into
638 // the tail_jump_rms array.
639 for( i=1; i < root->req(); i++ ) {
640 MachReturnNode *m = root->in(i)->as_MachReturn();
641 if( m->ideal_Opcode() == Op_TailJump ) {
642 tail_jump_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0);
643 tail_jump_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1);
644 break;
645 }
646 }
648 // Input RegMask array shared by all Halts
649 uint halt_edge_cnt = TypeFunc::Parms;
650 RegMask *halt_rms = init_input_masks( halt_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
652 // Capture the return input masks into each exit flavor
653 for( i=1; i < root->req(); i++ ) {
654 MachReturnNode *exit = root->in(i)->as_MachReturn();
655 switch( exit->ideal_Opcode() ) {
656 case Op_Return : exit->_in_rms = ret_rms; break;
657 case Op_Rethrow : exit->_in_rms = reth_rms; break;
658 case Op_TailCall : exit->_in_rms = tail_call_rms; break;
659 case Op_TailJump : exit->_in_rms = tail_jump_rms; break;
660 case Op_Halt : exit->_in_rms = halt_rms; break;
661 default : ShouldNotReachHere();
662 }
663 }
665 // Next unused projection number from Start.
666 int proj_cnt = C->tf()->domain()->cnt();
668 // Do all the save-on-entry registers. Make projections from Start for
669 // them, and give them a use at the exit points. To the allocator, they
670 // look like incoming register arguments.
671 for( i = 0; i < _last_Mach_Reg; i++ ) {
672 if( is_save_on_entry(i) ) {
674 // Add the save-on-entry to the mask array
675 ret_rms [ ret_edge_cnt] = mreg2regmask[i];
676 reth_rms [ reth_edge_cnt] = mreg2regmask[i];
677 tail_call_rms[tail_call_edge_cnt] = mreg2regmask[i];
678 tail_jump_rms[tail_jump_edge_cnt] = mreg2regmask[i];
679 // Halts need the SOE registers, but only in the stack as debug info.
680 // A just-prior uncommon-trap or deoptimization will use the SOE regs.
681 halt_rms [ halt_edge_cnt] = *idealreg2spillmask[_register_save_type[i]];
683 Node *mproj;
685 // Is this a RegF low half of a RegD? Double up 2 adjacent RegF's
686 // into a single RegD.
687 if( (i&1) == 0 &&
688 _register_save_type[i ] == Op_RegF &&
689 _register_save_type[i+1] == Op_RegF &&
690 is_save_on_entry(i+1) ) {
691 // Add other bit for double
692 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1));
693 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1));
694 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1));
695 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1));
696 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1));
697 mproj = new (C, 1) MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegD );
698 proj_cnt += 2; // Skip 2 for doubles
699 }
700 else if( (i&1) == 1 && // Else check for high half of double
701 _register_save_type[i-1] == Op_RegF &&
702 _register_save_type[i ] == Op_RegF &&
703 is_save_on_entry(i-1) ) {
704 ret_rms [ ret_edge_cnt] = RegMask::Empty;
705 reth_rms [ reth_edge_cnt] = RegMask::Empty;
706 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty;
707 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty;
708 halt_rms [ halt_edge_cnt] = RegMask::Empty;
709 mproj = C->top();
710 }
711 // Is this a RegI low half of a RegL? Double up 2 adjacent RegI's
712 // into a single RegL.
713 else if( (i&1) == 0 &&
714 _register_save_type[i ] == Op_RegI &&
715 _register_save_type[i+1] == Op_RegI &&
716 is_save_on_entry(i+1) ) {
717 // Add other bit for long
718 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1));
719 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1));
720 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1));
721 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1));
722 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1));
723 mproj = new (C, 1) MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegL );
724 proj_cnt += 2; // Skip 2 for longs
725 }
726 else if( (i&1) == 1 && // Else check for high half of long
727 _register_save_type[i-1] == Op_RegI &&
728 _register_save_type[i ] == Op_RegI &&
729 is_save_on_entry(i-1) ) {
730 ret_rms [ ret_edge_cnt] = RegMask::Empty;
731 reth_rms [ reth_edge_cnt] = RegMask::Empty;
732 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty;
733 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty;
734 halt_rms [ halt_edge_cnt] = RegMask::Empty;
735 mproj = C->top();
736 } else {
737 // Make a projection for it off the Start
738 mproj = new (C, 1) MachProjNode( start, proj_cnt++, ret_rms[ret_edge_cnt], _register_save_type[i] );
739 }
741 ret_edge_cnt ++;
742 reth_edge_cnt ++;
743 tail_call_edge_cnt ++;
744 tail_jump_edge_cnt ++;
745 halt_edge_cnt ++;
747 // Add a use of the SOE register to all exit paths
748 for( uint j=1; j < root->req(); j++ )
749 root->in(j)->add_req(mproj);
750 } // End of if a save-on-entry register
751 } // End of for all machine registers
752 }
754 //------------------------------init_spill_mask--------------------------------
755 void Matcher::init_spill_mask( Node *ret ) {
756 if( idealreg2regmask[Op_RegI] ) return; // One time only init
758 OptoReg::c_frame_pointer = c_frame_pointer();
759 c_frame_ptr_mask = c_frame_pointer();
760 #ifdef _LP64
761 // pointers are twice as big
762 c_frame_ptr_mask.Insert(OptoReg::add(c_frame_pointer(),1));
763 #endif
765 // Start at OptoReg::stack0()
766 STACK_ONLY_mask.Clear();
767 OptoReg::Name init = OptoReg::stack2reg(0);
768 // STACK_ONLY_mask is all stack bits
769 OptoReg::Name i;
770 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1))
771 STACK_ONLY_mask.Insert(i);
772 // Also set the "infinite stack" bit.
773 STACK_ONLY_mask.set_AllStack();
775 // Copy the register names over into the shared world
776 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) {
777 // SharedInfo::regName[i] = regName[i];
778 // Handy RegMasks per machine register
779 mreg2regmask[i].Insert(i);
780 }
782 // Grab the Frame Pointer
783 Node *fp = ret->in(TypeFunc::FramePtr);
784 Node *mem = ret->in(TypeFunc::Memory);
785 const TypePtr* atp = TypePtr::BOTTOM;
786 // Share frame pointer while making spill ops
787 set_shared(fp);
789 // Compute generic short-offset Loads
790 #ifdef _LP64
791 MachNode *spillCP = match_tree(new (C, 3) LoadNNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM));
792 #endif
793 MachNode *spillI = match_tree(new (C, 3) LoadINode(NULL,mem,fp,atp));
794 MachNode *spillL = match_tree(new (C, 3) LoadLNode(NULL,mem,fp,atp));
795 MachNode *spillF = match_tree(new (C, 3) LoadFNode(NULL,mem,fp,atp));
796 MachNode *spillD = match_tree(new (C, 3) LoadDNode(NULL,mem,fp,atp));
797 MachNode *spillP = match_tree(new (C, 3) LoadPNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM));
798 assert(spillI != NULL && spillL != NULL && spillF != NULL &&
799 spillD != NULL && spillP != NULL, "");
801 // Get the ADLC notion of the right regmask, for each basic type.
802 #ifdef _LP64
803 idealreg2regmask[Op_RegN] = &spillCP->out_RegMask();
804 #endif
805 idealreg2regmask[Op_RegI] = &spillI->out_RegMask();
806 idealreg2regmask[Op_RegL] = &spillL->out_RegMask();
807 idealreg2regmask[Op_RegF] = &spillF->out_RegMask();
808 idealreg2regmask[Op_RegD] = &spillD->out_RegMask();
809 idealreg2regmask[Op_RegP] = &spillP->out_RegMask();
810 }
812 #ifdef ASSERT
813 static void match_alias_type(Compile* C, Node* n, Node* m) {
814 if (!VerifyAliases) return; // do not go looking for trouble by default
815 const TypePtr* nat = n->adr_type();
816 const TypePtr* mat = m->adr_type();
817 int nidx = C->get_alias_index(nat);
818 int midx = C->get_alias_index(mat);
819 // Detune the assert for cases like (AndI 0xFF (LoadB p)).
820 if (nidx == Compile::AliasIdxTop && midx >= Compile::AliasIdxRaw) {
821 for (uint i = 1; i < n->req(); i++) {
822 Node* n1 = n->in(i);
823 const TypePtr* n1at = n1->adr_type();
824 if (n1at != NULL) {
825 nat = n1at;
826 nidx = C->get_alias_index(n1at);
827 }
828 }
829 }
830 // %%% Kludgery. Instead, fix ideal adr_type methods for all these cases:
831 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxRaw) {
832 switch (n->Opcode()) {
833 case Op_PrefetchRead:
834 case Op_PrefetchWrite:
835 case Op_PrefetchAllocation:
836 nidx = Compile::AliasIdxRaw;
837 nat = TypeRawPtr::BOTTOM;
838 break;
839 }
840 }
841 if (nidx == Compile::AliasIdxRaw && midx == Compile::AliasIdxTop) {
842 switch (n->Opcode()) {
843 case Op_ClearArray:
844 midx = Compile::AliasIdxRaw;
845 mat = TypeRawPtr::BOTTOM;
846 break;
847 }
848 }
849 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxBot) {
850 switch (n->Opcode()) {
851 case Op_Return:
852 case Op_Rethrow:
853 case Op_Halt:
854 case Op_TailCall:
855 case Op_TailJump:
856 nidx = Compile::AliasIdxBot;
857 nat = TypePtr::BOTTOM;
858 break;
859 }
860 }
861 if (nidx == Compile::AliasIdxBot && midx == Compile::AliasIdxTop) {
862 switch (n->Opcode()) {
863 case Op_StrComp:
864 case Op_StrEquals:
865 case Op_StrIndexOf:
866 case Op_AryEq:
867 case Op_MemBarVolatile:
868 case Op_MemBarCPUOrder: // %%% these ideals should have narrower adr_type?
869 nidx = Compile::AliasIdxTop;
870 nat = NULL;
871 break;
872 }
873 }
874 if (nidx != midx) {
875 if (PrintOpto || (PrintMiscellaneous && (WizardMode || Verbose))) {
876 tty->print_cr("==== Matcher alias shift %d => %d", nidx, midx);
877 n->dump();
878 m->dump();
879 }
880 assert(C->subsume_loads() && C->must_alias(nat, midx),
881 "must not lose alias info when matching");
882 }
883 }
884 #endif
887 //------------------------------MStack-----------------------------------------
888 // State and MStack class used in xform() and find_shared() iterative methods.
889 enum Node_State { Pre_Visit, // node has to be pre-visited
890 Visit, // visit node
891 Post_Visit, // post-visit node
892 Alt_Post_Visit // alternative post-visit path
893 };
895 class MStack: public Node_Stack {
896 public:
897 MStack(int size) : Node_Stack(size) { }
899 void push(Node *n, Node_State ns) {
900 Node_Stack::push(n, (uint)ns);
901 }
902 void push(Node *n, Node_State ns, Node *parent, int indx) {
903 ++_inode_top;
904 if ((_inode_top + 1) >= _inode_max) grow();
905 _inode_top->node = parent;
906 _inode_top->indx = (uint)indx;
907 ++_inode_top;
908 _inode_top->node = n;
909 _inode_top->indx = (uint)ns;
910 }
911 Node *parent() {
912 pop();
913 return node();
914 }
915 Node_State state() const {
916 return (Node_State)index();
917 }
918 void set_state(Node_State ns) {
919 set_index((uint)ns);
920 }
921 };
924 //------------------------------xform------------------------------------------
925 // Given a Node in old-space, Match him (Label/Reduce) to produce a machine
926 // Node in new-space. Given a new-space Node, recursively walk his children.
927 Node *Matcher::transform( Node *n ) { ShouldNotCallThis(); return n; }
928 Node *Matcher::xform( Node *n, int max_stack ) {
929 // Use one stack to keep both: child's node/state and parent's node/index
930 MStack mstack(max_stack * 2 * 2); // C->unique() * 2 * 2
931 mstack.push(n, Visit, NULL, -1); // set NULL as parent to indicate root
933 while (mstack.is_nonempty()) {
934 n = mstack.node(); // Leave node on stack
935 Node_State nstate = mstack.state();
936 if (nstate == Visit) {
937 mstack.set_state(Post_Visit);
938 Node *oldn = n;
939 // Old-space or new-space check
940 if (!C->node_arena()->contains(n)) {
941 // Old space!
942 Node* m;
943 if (has_new_node(n)) { // Not yet Label/Reduced
944 m = new_node(n);
945 } else {
946 if (!is_dontcare(n)) { // Matcher can match this guy
947 // Calls match special. They match alone with no children.
948 // Their children, the incoming arguments, match normally.
949 m = n->is_SafePoint() ? match_sfpt(n->as_SafePoint()):match_tree(n);
950 if (C->failing()) return NULL;
951 if (m == NULL) { Matcher::soft_match_failure(); return NULL; }
952 } else { // Nothing the matcher cares about
953 if( n->is_Proj() && n->in(0)->is_Multi()) { // Projections?
954 // Convert to machine-dependent projection
955 m = n->in(0)->as_Multi()->match( n->as_Proj(), this );
956 #ifdef ASSERT
957 _new2old_map.map(m->_idx, n);
958 #endif
959 if (m->in(0) != NULL) // m might be top
960 collect_null_checks(m, n);
961 } else { // Else just a regular 'ol guy
962 m = n->clone(); // So just clone into new-space
963 #ifdef ASSERT
964 _new2old_map.map(m->_idx, n);
965 #endif
966 // Def-Use edges will be added incrementally as Uses
967 // of this node are matched.
968 assert(m->outcnt() == 0, "no Uses of this clone yet");
969 }
970 }
972 set_new_node(n, m); // Map old to new
973 if (_old_node_note_array != NULL) {
974 Node_Notes* nn = C->locate_node_notes(_old_node_note_array,
975 n->_idx);
976 C->set_node_notes_at(m->_idx, nn);
977 }
978 debug_only(match_alias_type(C, n, m));
979 }
980 n = m; // n is now a new-space node
981 mstack.set_node(n);
982 }
984 // New space!
985 if (_visited.test_set(n->_idx)) continue; // while(mstack.is_nonempty())
987 int i;
988 // Put precedence edges on stack first (match them last).
989 for (i = oldn->req(); (uint)i < oldn->len(); i++) {
990 Node *m = oldn->in(i);
991 if (m == NULL) break;
992 // set -1 to call add_prec() instead of set_req() during Step1
993 mstack.push(m, Visit, n, -1);
994 }
996 // For constant debug info, I'd rather have unmatched constants.
997 int cnt = n->req();
998 JVMState* jvms = n->jvms();
999 int debug_cnt = jvms ? jvms->debug_start() : cnt;
1001 // Now do only debug info. Clone constants rather than matching.
1002 // Constants are represented directly in the debug info without
1003 // the need for executable machine instructions.
1004 // Monitor boxes are also represented directly.
1005 for (i = cnt - 1; i >= debug_cnt; --i) { // For all debug inputs do
1006 Node *m = n->in(i); // Get input
1007 int op = m->Opcode();
1008 assert((op == Op_BoxLock) == jvms->is_monitor_use(i), "boxes only at monitor sites");
1009 if( op == Op_ConI || op == Op_ConP || op == Op_ConN ||
1010 op == Op_ConF || op == Op_ConD || op == Op_ConL
1011 // || op == Op_BoxLock // %%%% enable this and remove (+++) in chaitin.cpp
1012 ) {
1013 m = m->clone();
1014 #ifdef ASSERT
1015 _new2old_map.map(m->_idx, n);
1016 #endif
1017 mstack.push(m, Post_Visit, n, i); // Don't need to visit
1018 mstack.push(m->in(0), Visit, m, 0);
1019 } else {
1020 mstack.push(m, Visit, n, i);
1021 }
1022 }
1024 // And now walk his children, and convert his inputs to new-space.
1025 for( ; i >= 0; --i ) { // For all normal inputs do
1026 Node *m = n->in(i); // Get input
1027 if(m != NULL)
1028 mstack.push(m, Visit, n, i);
1029 }
1031 }
1032 else if (nstate == Post_Visit) {
1033 // Set xformed input
1034 Node *p = mstack.parent();
1035 if (p != NULL) { // root doesn't have parent
1036 int i = (int)mstack.index();
1037 if (i >= 0)
1038 p->set_req(i, n); // required input
1039 else if (i == -1)
1040 p->add_prec(n); // precedence input
1041 else
1042 ShouldNotReachHere();
1043 }
1044 mstack.pop(); // remove processed node from stack
1045 }
1046 else {
1047 ShouldNotReachHere();
1048 }
1049 } // while (mstack.is_nonempty())
1050 return n; // Return new-space Node
1051 }
1053 //------------------------------warp_outgoing_stk_arg------------------------
1054 OptoReg::Name Matcher::warp_outgoing_stk_arg( VMReg reg, OptoReg::Name begin_out_arg_area, OptoReg::Name &out_arg_limit_per_call ) {
1055 // Convert outgoing argument location to a pre-biased stack offset
1056 if (reg->is_stack()) {
1057 OptoReg::Name warped = reg->reg2stack();
1058 // Adjust the stack slot offset to be the register number used
1059 // by the allocator.
1060 warped = OptoReg::add(begin_out_arg_area, warped);
1061 // Keep track of the largest numbered stack slot used for an arg.
1062 // Largest used slot per call-site indicates the amount of stack
1063 // that is killed by the call.
1064 if( warped >= out_arg_limit_per_call )
1065 out_arg_limit_per_call = OptoReg::add(warped,1);
1066 if (!RegMask::can_represent(warped)) {
1067 C->record_method_not_compilable_all_tiers("unsupported calling sequence");
1068 return OptoReg::Bad;
1069 }
1070 return warped;
1071 }
1072 return OptoReg::as_OptoReg(reg);
1073 }
1076 //------------------------------match_sfpt-------------------------------------
1077 // Helper function to match call instructions. Calls match special.
1078 // They match alone with no children. Their children, the incoming
1079 // arguments, match normally.
1080 MachNode *Matcher::match_sfpt( SafePointNode *sfpt ) {
1081 MachSafePointNode *msfpt = NULL;
1082 MachCallNode *mcall = NULL;
1083 uint cnt;
1084 // Split out case for SafePoint vs Call
1085 CallNode *call;
1086 const TypeTuple *domain;
1087 ciMethod* method = NULL;
1088 bool is_method_handle_invoke = false; // for special kill effects
1089 if( sfpt->is_Call() ) {
1090 call = sfpt->as_Call();
1091 domain = call->tf()->domain();
1092 cnt = domain->cnt();
1094 // Match just the call, nothing else
1095 MachNode *m = match_tree(call);
1096 if (C->failing()) return NULL;
1097 if( m == NULL ) { Matcher::soft_match_failure(); return NULL; }
1099 // Copy data from the Ideal SafePoint to the machine version
1100 mcall = m->as_MachCall();
1102 mcall->set_tf( call->tf());
1103 mcall->set_entry_point(call->entry_point());
1104 mcall->set_cnt( call->cnt());
1106 if( mcall->is_MachCallJava() ) {
1107 MachCallJavaNode *mcall_java = mcall->as_MachCallJava();
1108 const CallJavaNode *call_java = call->as_CallJava();
1109 method = call_java->method();
1110 mcall_java->_method = method;
1111 mcall_java->_bci = call_java->_bci;
1112 mcall_java->_optimized_virtual = call_java->is_optimized_virtual();
1113 is_method_handle_invoke = call_java->is_method_handle_invoke();
1114 mcall_java->_method_handle_invoke = is_method_handle_invoke;
1115 if (is_method_handle_invoke) {
1116 C->set_has_method_handle_invokes(true);
1117 }
1118 if( mcall_java->is_MachCallStaticJava() )
1119 mcall_java->as_MachCallStaticJava()->_name =
1120 call_java->as_CallStaticJava()->_name;
1121 if( mcall_java->is_MachCallDynamicJava() )
1122 mcall_java->as_MachCallDynamicJava()->_vtable_index =
1123 call_java->as_CallDynamicJava()->_vtable_index;
1124 }
1125 else if( mcall->is_MachCallRuntime() ) {
1126 mcall->as_MachCallRuntime()->_name = call->as_CallRuntime()->_name;
1127 }
1128 msfpt = mcall;
1129 }
1130 // This is a non-call safepoint
1131 else {
1132 call = NULL;
1133 domain = NULL;
1134 MachNode *mn = match_tree(sfpt);
1135 if (C->failing()) return NULL;
1136 msfpt = mn->as_MachSafePoint();
1137 cnt = TypeFunc::Parms;
1138 }
1140 // Advertise the correct memory effects (for anti-dependence computation).
1141 msfpt->set_adr_type(sfpt->adr_type());
1143 // Allocate a private array of RegMasks. These RegMasks are not shared.
1144 msfpt->_in_rms = NEW_RESOURCE_ARRAY( RegMask, cnt );
1145 // Empty them all.
1146 memset( msfpt->_in_rms, 0, sizeof(RegMask)*cnt );
1148 // Do all the pre-defined non-Empty register masks
1149 msfpt->_in_rms[TypeFunc::ReturnAdr] = _return_addr_mask;
1150 msfpt->_in_rms[TypeFunc::FramePtr ] = c_frame_ptr_mask;
1152 // Place first outgoing argument can possibly be put.
1153 OptoReg::Name begin_out_arg_area = OptoReg::add(_new_SP, C->out_preserve_stack_slots());
1154 assert( is_even(begin_out_arg_area), "" );
1155 // Compute max outgoing register number per call site.
1156 OptoReg::Name out_arg_limit_per_call = begin_out_arg_area;
1157 // Calls to C may hammer extra stack slots above and beyond any arguments.
1158 // These are usually backing store for register arguments for varargs.
1159 if( call != NULL && call->is_CallRuntime() )
1160 out_arg_limit_per_call = OptoReg::add(out_arg_limit_per_call,C->varargs_C_out_slots_killed());
1163 // Do the normal argument list (parameters) register masks
1164 int argcnt = cnt - TypeFunc::Parms;
1165 if( argcnt > 0 ) { // Skip it all if we have no args
1166 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt );
1167 VMRegPair *parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt );
1168 int i;
1169 for( i = 0; i < argcnt; i++ ) {
1170 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type();
1171 }
1172 // V-call to pick proper calling convention
1173 call->calling_convention( sig_bt, parm_regs, argcnt );
1175 #ifdef ASSERT
1176 // Sanity check users' calling convention. Really handy during
1177 // the initial porting effort. Fairly expensive otherwise.
1178 { for (int i = 0; i<argcnt; i++) {
1179 if( !parm_regs[i].first()->is_valid() &&
1180 !parm_regs[i].second()->is_valid() ) continue;
1181 VMReg reg1 = parm_regs[i].first();
1182 VMReg reg2 = parm_regs[i].second();
1183 for (int j = 0; j < i; j++) {
1184 if( !parm_regs[j].first()->is_valid() &&
1185 !parm_regs[j].second()->is_valid() ) continue;
1186 VMReg reg3 = parm_regs[j].first();
1187 VMReg reg4 = parm_regs[j].second();
1188 if( !reg1->is_valid() ) {
1189 assert( !reg2->is_valid(), "valid halvsies" );
1190 } else if( !reg3->is_valid() ) {
1191 assert( !reg4->is_valid(), "valid halvsies" );
1192 } else {
1193 assert( reg1 != reg2, "calling conv. must produce distinct regs");
1194 assert( reg1 != reg3, "calling conv. must produce distinct regs");
1195 assert( reg1 != reg4, "calling conv. must produce distinct regs");
1196 assert( reg2 != reg3, "calling conv. must produce distinct regs");
1197 assert( reg2 != reg4 || !reg2->is_valid(), "calling conv. must produce distinct regs");
1198 assert( reg3 != reg4, "calling conv. must produce distinct regs");
1199 }
1200 }
1201 }
1202 }
1203 #endif
1205 // Visit each argument. Compute its outgoing register mask.
1206 // Return results now can have 2 bits returned.
1207 // Compute max over all outgoing arguments both per call-site
1208 // and over the entire method.
1209 for( i = 0; i < argcnt; i++ ) {
1210 // Address of incoming argument mask to fill in
1211 RegMask *rm = &mcall->_in_rms[i+TypeFunc::Parms];
1212 if( !parm_regs[i].first()->is_valid() &&
1213 !parm_regs[i].second()->is_valid() ) {
1214 continue; // Avoid Halves
1215 }
1216 // Grab first register, adjust stack slots and insert in mask.
1217 OptoReg::Name reg1 = warp_outgoing_stk_arg(parm_regs[i].first(), begin_out_arg_area, out_arg_limit_per_call );
1218 if (OptoReg::is_valid(reg1))
1219 rm->Insert( reg1 );
1220 // Grab second register (if any), adjust stack slots and insert in mask.
1221 OptoReg::Name reg2 = warp_outgoing_stk_arg(parm_regs[i].second(), begin_out_arg_area, out_arg_limit_per_call );
1222 if (OptoReg::is_valid(reg2))
1223 rm->Insert( reg2 );
1224 } // End of for all arguments
1226 // Compute number of stack slots needed to restore stack in case of
1227 // Pascal-style argument popping.
1228 mcall->_argsize = out_arg_limit_per_call - begin_out_arg_area;
1229 }
1231 if (is_method_handle_invoke) {
1232 // Kill some extra stack space in case method handles want to do
1233 // a little in-place argument insertion.
1234 int regs_per_word = NOT_LP64(1) LP64_ONLY(2); // %%% make a global const!
1235 out_arg_limit_per_call += MethodHandlePushLimit * regs_per_word;
1236 // Do not update mcall->_argsize because (a) the extra space is not
1237 // pushed as arguments and (b) _argsize is dead (not used anywhere).
1238 }
1240 // Compute the max stack slot killed by any call. These will not be
1241 // available for debug info, and will be used to adjust FIRST_STACK_mask
1242 // after all call sites have been visited.
1243 if( _out_arg_limit < out_arg_limit_per_call)
1244 _out_arg_limit = out_arg_limit_per_call;
1246 if (mcall) {
1247 // Kill the outgoing argument area, including any non-argument holes and
1248 // any legacy C-killed slots. Use Fat-Projections to do the killing.
1249 // Since the max-per-method covers the max-per-call-site and debug info
1250 // is excluded on the max-per-method basis, debug info cannot land in
1251 // this killed area.
1252 uint r_cnt = mcall->tf()->range()->cnt();
1253 MachProjNode *proj = new (C, 1) MachProjNode( mcall, r_cnt+10000, RegMask::Empty, MachProjNode::fat_proj );
1254 if (!RegMask::can_represent(OptoReg::Name(out_arg_limit_per_call-1))) {
1255 C->record_method_not_compilable_all_tiers("unsupported outgoing calling sequence");
1256 } else {
1257 for (int i = begin_out_arg_area; i < out_arg_limit_per_call; i++)
1258 proj->_rout.Insert(OptoReg::Name(i));
1259 }
1260 if( proj->_rout.is_NotEmpty() )
1261 _proj_list.push(proj);
1262 }
1263 // Transfer the safepoint information from the call to the mcall
1264 // Move the JVMState list
1265 msfpt->set_jvms(sfpt->jvms());
1266 for (JVMState* jvms = msfpt->jvms(); jvms; jvms = jvms->caller()) {
1267 jvms->set_map(sfpt);
1268 }
1270 // Debug inputs begin just after the last incoming parameter
1271 assert( (mcall == NULL) || (mcall->jvms() == NULL) ||
1272 (mcall->jvms()->debug_start() + mcall->_jvmadj == mcall->tf()->domain()->cnt()), "" );
1274 // Move the OopMap
1275 msfpt->_oop_map = sfpt->_oop_map;
1277 // Registers killed by the call are set in the local scheduling pass
1278 // of Global Code Motion.
1279 return msfpt;
1280 }
1282 //---------------------------match_tree----------------------------------------
1283 // Match a Ideal Node DAG - turn it into a tree; Label & Reduce. Used as part
1284 // of the whole-sale conversion from Ideal to Mach Nodes. Also used for
1285 // making GotoNodes while building the CFG and in init_spill_mask() to identify
1286 // a Load's result RegMask for memoization in idealreg2regmask[]
1287 MachNode *Matcher::match_tree( const Node *n ) {
1288 assert( n->Opcode() != Op_Phi, "cannot match" );
1289 assert( !n->is_block_start(), "cannot match" );
1290 // Set the mark for all locally allocated State objects.
1291 // When this call returns, the _states_arena arena will be reset
1292 // freeing all State objects.
1293 ResourceMark rm( &_states_arena );
1295 LabelRootDepth = 0;
1297 // StoreNodes require their Memory input to match any LoadNodes
1298 Node *mem = n->is_Store() ? n->in(MemNode::Memory) : (Node*)1 ;
1299 #ifdef ASSERT
1300 Node* save_mem_node = _mem_node;
1301 _mem_node = n->is_Store() ? (Node*)n : NULL;
1302 #endif
1303 // State object for root node of match tree
1304 // Allocate it on _states_arena - stack allocation can cause stack overflow.
1305 State *s = new (&_states_arena) State;
1306 s->_kids[0] = NULL;
1307 s->_kids[1] = NULL;
1308 s->_leaf = (Node*)n;
1309 // Label the input tree, allocating labels from top-level arena
1310 Label_Root( n, s, n->in(0), mem );
1311 if (C->failing()) return NULL;
1313 // The minimum cost match for the whole tree is found at the root State
1314 uint mincost = max_juint;
1315 uint cost = max_juint;
1316 uint i;
1317 for( i = 0; i < NUM_OPERANDS; i++ ) {
1318 if( s->valid(i) && // valid entry and
1319 s->_cost[i] < cost && // low cost and
1320 s->_rule[i] >= NUM_OPERANDS ) // not an operand
1321 cost = s->_cost[mincost=i];
1322 }
1323 if (mincost == max_juint) {
1324 #ifndef PRODUCT
1325 tty->print("No matching rule for:");
1326 s->dump();
1327 #endif
1328 Matcher::soft_match_failure();
1329 return NULL;
1330 }
1331 // Reduce input tree based upon the state labels to machine Nodes
1332 MachNode *m = ReduceInst( s, s->_rule[mincost], mem );
1333 #ifdef ASSERT
1334 _old2new_map.map(n->_idx, m);
1335 _new2old_map.map(m->_idx, (Node*)n);
1336 #endif
1338 // Add any Matcher-ignored edges
1339 uint cnt = n->req();
1340 uint start = 1;
1341 if( mem != (Node*)1 ) start = MemNode::Memory+1;
1342 if( n->is_AddP() ) {
1343 assert( mem == (Node*)1, "" );
1344 start = AddPNode::Base+1;
1345 }
1346 for( i = start; i < cnt; i++ ) {
1347 if( !n->match_edge(i) ) {
1348 if( i < m->req() )
1349 m->ins_req( i, n->in(i) );
1350 else
1351 m->add_req( n->in(i) );
1352 }
1353 }
1355 debug_only( _mem_node = save_mem_node; )
1356 return m;
1357 }
1360 //------------------------------match_into_reg---------------------------------
1361 // Choose to either match this Node in a register or part of the current
1362 // match tree. Return true for requiring a register and false for matching
1363 // as part of the current match tree.
1364 static bool match_into_reg( const Node *n, Node *m, Node *control, int i, bool shared ) {
1366 const Type *t = m->bottom_type();
1368 if( t->singleton() ) {
1369 // Never force constants into registers. Allow them to match as
1370 // constants or registers. Copies of the same value will share
1371 // the same register. See find_shared_node.
1372 return false;
1373 } else { // Not a constant
1374 // Stop recursion if they have different Controls.
1375 // Slot 0 of constants is not really a Control.
1376 if( control && m->in(0) && control != m->in(0) ) {
1378 // Actually, we can live with the most conservative control we
1379 // find, if it post-dominates the others. This allows us to
1380 // pick up load/op/store trees where the load can float a little
1381 // above the store.
1382 Node *x = control;
1383 const uint max_scan = 6; // Arbitrary scan cutoff
1384 uint j;
1385 for( j=0; j<max_scan; j++ ) {
1386 if( x->is_Region() ) // Bail out at merge points
1387 return true;
1388 x = x->in(0);
1389 if( x == m->in(0) ) // Does 'control' post-dominate
1390 break; // m->in(0)? If so, we can use it
1391 }
1392 if( j == max_scan ) // No post-domination before scan end?
1393 return true; // Then break the match tree up
1394 }
1395 if (m->is_DecodeN() && Matcher::narrow_oop_use_complex_address()) {
1396 // These are commonly used in address expressions and can
1397 // efficiently fold into them on X64 in some cases.
1398 return false;
1399 }
1400 }
1402 // Not forceable cloning. If shared, put it into a register.
1403 return shared;
1404 }
1407 //------------------------------Instruction Selection--------------------------
1408 // Label method walks a "tree" of nodes, using the ADLC generated DFA to match
1409 // ideal nodes to machine instructions. Trees are delimited by shared Nodes,
1410 // things the Matcher does not match (e.g., Memory), and things with different
1411 // Controls (hence forced into different blocks). We pass in the Control
1412 // selected for this entire State tree.
1414 // The Matcher works on Trees, but an Intel add-to-memory requires a DAG: the
1415 // Store and the Load must have identical Memories (as well as identical
1416 // pointers). Since the Matcher does not have anything for Memory (and
1417 // does not handle DAGs), I have to match the Memory input myself. If the
1418 // Tree root is a Store, I require all Loads to have the identical memory.
1419 Node *Matcher::Label_Root( const Node *n, State *svec, Node *control, const Node *mem){
1420 // Since Label_Root is a recursive function, its possible that we might run
1421 // out of stack space. See bugs 6272980 & 6227033 for more info.
1422 LabelRootDepth++;
1423 if (LabelRootDepth > MaxLabelRootDepth) {
1424 C->record_method_not_compilable_all_tiers("Out of stack space, increase MaxLabelRootDepth");
1425 return NULL;
1426 }
1427 uint care = 0; // Edges matcher cares about
1428 uint cnt = n->req();
1429 uint i = 0;
1431 // Examine children for memory state
1432 // Can only subsume a child into your match-tree if that child's memory state
1433 // is not modified along the path to another input.
1434 // It is unsafe even if the other inputs are separate roots.
1435 Node *input_mem = NULL;
1436 for( i = 1; i < cnt; i++ ) {
1437 if( !n->match_edge(i) ) continue;
1438 Node *m = n->in(i); // Get ith input
1439 assert( m, "expect non-null children" );
1440 if( m->is_Load() ) {
1441 if( input_mem == NULL ) {
1442 input_mem = m->in(MemNode::Memory);
1443 } else if( input_mem != m->in(MemNode::Memory) ) {
1444 input_mem = NodeSentinel;
1445 }
1446 }
1447 }
1449 for( i = 1; i < cnt; i++ ){// For my children
1450 if( !n->match_edge(i) ) continue;
1451 Node *m = n->in(i); // Get ith input
1452 // Allocate states out of a private arena
1453 State *s = new (&_states_arena) State;
1454 svec->_kids[care++] = s;
1455 assert( care <= 2, "binary only for now" );
1457 // Recursively label the State tree.
1458 s->_kids[0] = NULL;
1459 s->_kids[1] = NULL;
1460 s->_leaf = m;
1462 // Check for leaves of the State Tree; things that cannot be a part of
1463 // the current tree. If it finds any, that value is matched as a
1464 // register operand. If not, then the normal matching is used.
1465 if( match_into_reg(n, m, control, i, is_shared(m)) ||
1466 //
1467 // Stop recursion if this is LoadNode and the root of this tree is a
1468 // StoreNode and the load & store have different memories.
1469 ((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) ||
1470 // Can NOT include the match of a subtree when its memory state
1471 // is used by any of the other subtrees
1472 (input_mem == NodeSentinel) ) {
1473 #ifndef PRODUCT
1474 // Print when we exclude matching due to different memory states at input-loads
1475 if( PrintOpto && (Verbose && WizardMode) && (input_mem == NodeSentinel)
1476 && !((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) ) {
1477 tty->print_cr("invalid input_mem");
1478 }
1479 #endif
1480 // Switch to a register-only opcode; this value must be in a register
1481 // and cannot be subsumed as part of a larger instruction.
1482 s->DFA( m->ideal_reg(), m );
1484 } else {
1485 // If match tree has no control and we do, adopt it for entire tree
1486 if( control == NULL && m->in(0) != NULL && m->req() > 1 )
1487 control = m->in(0); // Pick up control
1488 // Else match as a normal part of the match tree.
1489 control = Label_Root(m,s,control,mem);
1490 if (C->failing()) return NULL;
1491 }
1492 }
1495 // Call DFA to match this node, and return
1496 svec->DFA( n->Opcode(), n );
1498 #ifdef ASSERT
1499 uint x;
1500 for( x = 0; x < _LAST_MACH_OPER; x++ )
1501 if( svec->valid(x) )
1502 break;
1504 if (x >= _LAST_MACH_OPER) {
1505 n->dump();
1506 svec->dump();
1507 assert( false, "bad AD file" );
1508 }
1509 #endif
1510 return control;
1511 }
1514 // Con nodes reduced using the same rule can share their MachNode
1515 // which reduces the number of copies of a constant in the final
1516 // program. The register allocator is free to split uses later to
1517 // split live ranges.
1518 MachNode* Matcher::find_shared_node(Node* leaf, uint rule) {
1519 if (!leaf->is_Con() && !leaf->is_DecodeN()) return NULL;
1521 // See if this Con has already been reduced using this rule.
1522 if (_shared_nodes.Size() <= leaf->_idx) return NULL;
1523 MachNode* last = (MachNode*)_shared_nodes.at(leaf->_idx);
1524 if (last != NULL && rule == last->rule()) {
1525 // Don't expect control change for DecodeN
1526 if (leaf->is_DecodeN())
1527 return last;
1528 // Get the new space root.
1529 Node* xroot = new_node(C->root());
1530 if (xroot == NULL) {
1531 // This shouldn't happen give the order of matching.
1532 return NULL;
1533 }
1535 // Shared constants need to have their control be root so they
1536 // can be scheduled properly.
1537 Node* control = last->in(0);
1538 if (control != xroot) {
1539 if (control == NULL || control == C->root()) {
1540 last->set_req(0, xroot);
1541 } else {
1542 assert(false, "unexpected control");
1543 return NULL;
1544 }
1545 }
1546 return last;
1547 }
1548 return NULL;
1549 }
1552 //------------------------------ReduceInst-------------------------------------
1553 // Reduce a State tree (with given Control) into a tree of MachNodes.
1554 // This routine (and it's cohort ReduceOper) convert Ideal Nodes into
1555 // complicated machine Nodes. Each MachNode covers some tree of Ideal Nodes.
1556 // Each MachNode has a number of complicated MachOper operands; each
1557 // MachOper also covers a further tree of Ideal Nodes.
1559 // The root of the Ideal match tree is always an instruction, so we enter
1560 // the recursion here. After building the MachNode, we need to recurse
1561 // the tree checking for these cases:
1562 // (1) Child is an instruction -
1563 // Build the instruction (recursively), add it as an edge.
1564 // Build a simple operand (register) to hold the result of the instruction.
1565 // (2) Child is an interior part of an instruction -
1566 // Skip over it (do nothing)
1567 // (3) Child is the start of a operand -
1568 // Build the operand, place it inside the instruction
1569 // Call ReduceOper.
1570 MachNode *Matcher::ReduceInst( State *s, int rule, Node *&mem ) {
1571 assert( rule >= NUM_OPERANDS, "called with operand rule" );
1573 MachNode* shared_node = find_shared_node(s->_leaf, rule);
1574 if (shared_node != NULL) {
1575 return shared_node;
1576 }
1578 // Build the object to represent this state & prepare for recursive calls
1579 MachNode *mach = s->MachNodeGenerator( rule, C );
1580 mach->_opnds[0] = s->MachOperGenerator( _reduceOp[rule], C );
1581 assert( mach->_opnds[0] != NULL, "Missing result operand" );
1582 Node *leaf = s->_leaf;
1583 // Check for instruction or instruction chain rule
1584 if( rule >= _END_INST_CHAIN_RULE || rule < _BEGIN_INST_CHAIN_RULE ) {
1585 assert(C->node_arena()->contains(s->_leaf) || !has_new_node(s->_leaf),
1586 "duplicating node that's already been matched");
1587 // Instruction
1588 mach->add_req( leaf->in(0) ); // Set initial control
1589 // Reduce interior of complex instruction
1590 ReduceInst_Interior( s, rule, mem, mach, 1 );
1591 } else {
1592 // Instruction chain rules are data-dependent on their inputs
1593 mach->add_req(0); // Set initial control to none
1594 ReduceInst_Chain_Rule( s, rule, mem, mach );
1595 }
1597 // If a Memory was used, insert a Memory edge
1598 if( mem != (Node*)1 ) {
1599 mach->ins_req(MemNode::Memory,mem);
1600 #ifdef ASSERT
1601 // Verify adr type after matching memory operation
1602 const MachOper* oper = mach->memory_operand();
1603 if (oper != NULL && oper != (MachOper*)-1) {
1604 // It has a unique memory operand. Find corresponding ideal mem node.
1605 Node* m = NULL;
1606 if (leaf->is_Mem()) {
1607 m = leaf;
1608 } else {
1609 m = _mem_node;
1610 assert(m != NULL && m->is_Mem(), "expecting memory node");
1611 }
1612 const Type* mach_at = mach->adr_type();
1613 // DecodeN node consumed by an address may have different type
1614 // then its input. Don't compare types for such case.
1615 if (m->adr_type() != mach_at &&
1616 (m->in(MemNode::Address)->is_DecodeN() ||
1617 m->in(MemNode::Address)->is_AddP() &&
1618 m->in(MemNode::Address)->in(AddPNode::Address)->is_DecodeN() ||
1619 m->in(MemNode::Address)->is_AddP() &&
1620 m->in(MemNode::Address)->in(AddPNode::Address)->is_AddP() &&
1621 m->in(MemNode::Address)->in(AddPNode::Address)->in(AddPNode::Address)->is_DecodeN())) {
1622 mach_at = m->adr_type();
1623 }
1624 if (m->adr_type() != mach_at) {
1625 m->dump();
1626 tty->print_cr("mach:");
1627 mach->dump(1);
1628 }
1629 assert(m->adr_type() == mach_at, "matcher should not change adr type");
1630 }
1631 #endif
1632 }
1634 // If the _leaf is an AddP, insert the base edge
1635 if( leaf->is_AddP() )
1636 mach->ins_req(AddPNode::Base,leaf->in(AddPNode::Base));
1638 uint num_proj = _proj_list.size();
1640 // Perform any 1-to-many expansions required
1641 MachNode *ex = mach->Expand(s,_proj_list, mem);
1642 if( ex != mach ) {
1643 assert(ex->ideal_reg() == mach->ideal_reg(), "ideal types should match");
1644 if( ex->in(1)->is_Con() )
1645 ex->in(1)->set_req(0, C->root());
1646 // Remove old node from the graph
1647 for( uint i=0; i<mach->req(); i++ ) {
1648 mach->set_req(i,NULL);
1649 }
1650 #ifdef ASSERT
1651 _new2old_map.map(ex->_idx, s->_leaf);
1652 #endif
1653 }
1655 // PhaseChaitin::fixup_spills will sometimes generate spill code
1656 // via the matcher. By the time, nodes have been wired into the CFG,
1657 // and any further nodes generated by expand rules will be left hanging
1658 // in space, and will not get emitted as output code. Catch this.
1659 // Also, catch any new register allocation constraints ("projections")
1660 // generated belatedly during spill code generation.
1661 if (_allocation_started) {
1662 guarantee(ex == mach, "no expand rules during spill generation");
1663 guarantee(_proj_list.size() == num_proj, "no allocation during spill generation");
1664 }
1666 if (leaf->is_Con() || leaf->is_DecodeN()) {
1667 // Record the con for sharing
1668 _shared_nodes.map(leaf->_idx, ex);
1669 }
1671 return ex;
1672 }
1674 void Matcher::ReduceInst_Chain_Rule( State *s, int rule, Node *&mem, MachNode *mach ) {
1675 // 'op' is what I am expecting to receive
1676 int op = _leftOp[rule];
1677 // Operand type to catch childs result
1678 // This is what my child will give me.
1679 int opnd_class_instance = s->_rule[op];
1680 // Choose between operand class or not.
1681 // This is what I will receive.
1682 int catch_op = (FIRST_OPERAND_CLASS <= op && op < NUM_OPERANDS) ? opnd_class_instance : op;
1683 // New rule for child. Chase operand classes to get the actual rule.
1684 int newrule = s->_rule[catch_op];
1686 if( newrule < NUM_OPERANDS ) {
1687 // Chain from operand or operand class, may be output of shared node
1688 assert( 0 <= opnd_class_instance && opnd_class_instance < NUM_OPERANDS,
1689 "Bad AD file: Instruction chain rule must chain from operand");
1690 // Insert operand into array of operands for this instruction
1691 mach->_opnds[1] = s->MachOperGenerator( opnd_class_instance, C );
1693 ReduceOper( s, newrule, mem, mach );
1694 } else {
1695 // Chain from the result of an instruction
1696 assert( newrule >= _LAST_MACH_OPER, "Do NOT chain from internal operand");
1697 mach->_opnds[1] = s->MachOperGenerator( _reduceOp[catch_op], C );
1698 Node *mem1 = (Node*)1;
1699 debug_only(Node *save_mem_node = _mem_node;)
1700 mach->add_req( ReduceInst(s, newrule, mem1) );
1701 debug_only(_mem_node = save_mem_node;)
1702 }
1703 return;
1704 }
1707 uint Matcher::ReduceInst_Interior( State *s, int rule, Node *&mem, MachNode *mach, uint num_opnds ) {
1708 if( s->_leaf->is_Load() ) {
1709 Node *mem2 = s->_leaf->in(MemNode::Memory);
1710 assert( mem == (Node*)1 || mem == mem2, "multiple Memories being matched at once?" );
1711 debug_only( if( mem == (Node*)1 ) _mem_node = s->_leaf;)
1712 mem = mem2;
1713 }
1714 if( s->_leaf->in(0) != NULL && s->_leaf->req() > 1) {
1715 if( mach->in(0) == NULL )
1716 mach->set_req(0, s->_leaf->in(0));
1717 }
1719 // Now recursively walk the state tree & add operand list.
1720 for( uint i=0; i<2; i++ ) { // binary tree
1721 State *newstate = s->_kids[i];
1722 if( newstate == NULL ) break; // Might only have 1 child
1723 // 'op' is what I am expecting to receive
1724 int op;
1725 if( i == 0 ) {
1726 op = _leftOp[rule];
1727 } else {
1728 op = _rightOp[rule];
1729 }
1730 // Operand type to catch childs result
1731 // This is what my child will give me.
1732 int opnd_class_instance = newstate->_rule[op];
1733 // Choose between operand class or not.
1734 // This is what I will receive.
1735 int catch_op = (op >= FIRST_OPERAND_CLASS && op < NUM_OPERANDS) ? opnd_class_instance : op;
1736 // New rule for child. Chase operand classes to get the actual rule.
1737 int newrule = newstate->_rule[catch_op];
1739 if( newrule < NUM_OPERANDS ) { // Operand/operandClass or internalOp/instruction?
1740 // Operand/operandClass
1741 // Insert operand into array of operands for this instruction
1742 mach->_opnds[num_opnds++] = newstate->MachOperGenerator( opnd_class_instance, C );
1743 ReduceOper( newstate, newrule, mem, mach );
1745 } else { // Child is internal operand or new instruction
1746 if( newrule < _LAST_MACH_OPER ) { // internal operand or instruction?
1747 // internal operand --> call ReduceInst_Interior
1748 // Interior of complex instruction. Do nothing but recurse.
1749 num_opnds = ReduceInst_Interior( newstate, newrule, mem, mach, num_opnds );
1750 } else {
1751 // instruction --> call build operand( ) to catch result
1752 // --> ReduceInst( newrule )
1753 mach->_opnds[num_opnds++] = s->MachOperGenerator( _reduceOp[catch_op], C );
1754 Node *mem1 = (Node*)1;
1755 debug_only(Node *save_mem_node = _mem_node;)
1756 mach->add_req( ReduceInst( newstate, newrule, mem1 ) );
1757 debug_only(_mem_node = save_mem_node;)
1758 }
1759 }
1760 assert( mach->_opnds[num_opnds-1], "" );
1761 }
1762 return num_opnds;
1763 }
1765 // This routine walks the interior of possible complex operands.
1766 // At each point we check our children in the match tree:
1767 // (1) No children -
1768 // We are a leaf; add _leaf field as an input to the MachNode
1769 // (2) Child is an internal operand -
1770 // Skip over it ( do nothing )
1771 // (3) Child is an instruction -
1772 // Call ReduceInst recursively and
1773 // and instruction as an input to the MachNode
1774 void Matcher::ReduceOper( State *s, int rule, Node *&mem, MachNode *mach ) {
1775 assert( rule < _LAST_MACH_OPER, "called with operand rule" );
1776 State *kid = s->_kids[0];
1777 assert( kid == NULL || s->_leaf->in(0) == NULL, "internal operands have no control" );
1779 // Leaf? And not subsumed?
1780 if( kid == NULL && !_swallowed[rule] ) {
1781 mach->add_req( s->_leaf ); // Add leaf pointer
1782 return; // Bail out
1783 }
1785 if( s->_leaf->is_Load() ) {
1786 assert( mem == (Node*)1, "multiple Memories being matched at once?" );
1787 mem = s->_leaf->in(MemNode::Memory);
1788 debug_only(_mem_node = s->_leaf;)
1789 }
1790 if( s->_leaf->in(0) && s->_leaf->req() > 1) {
1791 if( !mach->in(0) )
1792 mach->set_req(0,s->_leaf->in(0));
1793 else {
1794 assert( s->_leaf->in(0) == mach->in(0), "same instruction, differing controls?" );
1795 }
1796 }
1798 for( uint i=0; kid != NULL && i<2; kid = s->_kids[1], i++ ) { // binary tree
1799 int newrule;
1800 if( i == 0 )
1801 newrule = kid->_rule[_leftOp[rule]];
1802 else
1803 newrule = kid->_rule[_rightOp[rule]];
1805 if( newrule < _LAST_MACH_OPER ) { // Operand or instruction?
1806 // Internal operand; recurse but do nothing else
1807 ReduceOper( kid, newrule, mem, mach );
1809 } else { // Child is a new instruction
1810 // Reduce the instruction, and add a direct pointer from this
1811 // machine instruction to the newly reduced one.
1812 Node *mem1 = (Node*)1;
1813 debug_only(Node *save_mem_node = _mem_node;)
1814 mach->add_req( ReduceInst( kid, newrule, mem1 ) );
1815 debug_only(_mem_node = save_mem_node;)
1816 }
1817 }
1818 }
1821 // -------------------------------------------------------------------------
1822 // Java-Java calling convention
1823 // (what you use when Java calls Java)
1825 //------------------------------find_receiver----------------------------------
1826 // For a given signature, return the OptoReg for parameter 0.
1827 OptoReg::Name Matcher::find_receiver( bool is_outgoing ) {
1828 VMRegPair regs;
1829 BasicType sig_bt = T_OBJECT;
1830 calling_convention(&sig_bt, ®s, 1, is_outgoing);
1831 // Return argument 0 register. In the LP64 build pointers
1832 // take 2 registers, but the VM wants only the 'main' name.
1833 return OptoReg::as_OptoReg(regs.first());
1834 }
1836 // A method-klass-holder may be passed in the inline_cache_reg
1837 // and then expanded into the inline_cache_reg and a method_oop register
1838 // defined in ad_<arch>.cpp
1841 //------------------------------find_shared------------------------------------
1842 // Set bits if Node is shared or otherwise a root
1843 void Matcher::find_shared( Node *n ) {
1844 // Allocate stack of size C->unique() * 2 to avoid frequent realloc
1845 MStack mstack(C->unique() * 2);
1846 // Mark nodes as address_visited if they are inputs to an address expression
1847 VectorSet address_visited(Thread::current()->resource_area());
1848 mstack.push(n, Visit); // Don't need to pre-visit root node
1849 while (mstack.is_nonempty()) {
1850 n = mstack.node(); // Leave node on stack
1851 Node_State nstate = mstack.state();
1852 uint nop = n->Opcode();
1853 if (nstate == Pre_Visit) {
1854 if (address_visited.test(n->_idx)) { // Visited in address already?
1855 // Flag as visited and shared now.
1856 set_visited(n);
1857 }
1858 if (is_visited(n)) { // Visited already?
1859 // Node is shared and has no reason to clone. Flag it as shared.
1860 // This causes it to match into a register for the sharing.
1861 set_shared(n); // Flag as shared and
1862 mstack.pop(); // remove node from stack
1863 continue;
1864 }
1865 nstate = Visit; // Not already visited; so visit now
1866 }
1867 if (nstate == Visit) {
1868 mstack.set_state(Post_Visit);
1869 set_visited(n); // Flag as visited now
1870 bool mem_op = false;
1872 switch( nop ) { // Handle some opcodes special
1873 case Op_Phi: // Treat Phis as shared roots
1874 case Op_Parm:
1875 case Op_Proj: // All handled specially during matching
1876 case Op_SafePointScalarObject:
1877 set_shared(n);
1878 set_dontcare(n);
1879 break;
1880 case Op_If:
1881 case Op_CountedLoopEnd:
1882 mstack.set_state(Alt_Post_Visit); // Alternative way
1883 // Convert (If (Bool (CmpX A B))) into (If (Bool) (CmpX A B)). Helps
1884 // with matching cmp/branch in 1 instruction. The Matcher needs the
1885 // Bool and CmpX side-by-side, because it can only get at constants
1886 // that are at the leaves of Match trees, and the Bool's condition acts
1887 // as a constant here.
1888 mstack.push(n->in(1), Visit); // Clone the Bool
1889 mstack.push(n->in(0), Pre_Visit); // Visit control input
1890 continue; // while (mstack.is_nonempty())
1891 case Op_ConvI2D: // These forms efficiently match with a prior
1892 case Op_ConvI2F: // Load but not a following Store
1893 if( n->in(1)->is_Load() && // Prior load
1894 n->outcnt() == 1 && // Not already shared
1895 n->unique_out()->is_Store() ) // Following store
1896 set_shared(n); // Force it to be a root
1897 break;
1898 case Op_ReverseBytesI:
1899 case Op_ReverseBytesL:
1900 if( n->in(1)->is_Load() && // Prior load
1901 n->outcnt() == 1 ) // Not already shared
1902 set_shared(n); // Force it to be a root
1903 break;
1904 case Op_BoxLock: // Cant match until we get stack-regs in ADLC
1905 case Op_IfFalse:
1906 case Op_IfTrue:
1907 case Op_MachProj:
1908 case Op_MergeMem:
1909 case Op_Catch:
1910 case Op_CatchProj:
1911 case Op_CProj:
1912 case Op_JumpProj:
1913 case Op_JProj:
1914 case Op_NeverBranch:
1915 set_dontcare(n);
1916 break;
1917 case Op_Jump:
1918 mstack.push(n->in(1), Visit); // Switch Value
1919 mstack.push(n->in(0), Pre_Visit); // Visit Control input
1920 continue; // while (mstack.is_nonempty())
1921 case Op_StrComp:
1922 case Op_StrEquals:
1923 case Op_StrIndexOf:
1924 case Op_AryEq:
1925 set_shared(n); // Force result into register (it will be anyways)
1926 break;
1927 case Op_ConP: { // Convert pointers above the centerline to NUL
1928 TypeNode *tn = n->as_Type(); // Constants derive from type nodes
1929 const TypePtr* tp = tn->type()->is_ptr();
1930 if (tp->_ptr == TypePtr::AnyNull) {
1931 tn->set_type(TypePtr::NULL_PTR);
1932 }
1933 break;
1934 }
1935 case Op_ConN: { // Convert narrow pointers above the centerline to NUL
1936 TypeNode *tn = n->as_Type(); // Constants derive from type nodes
1937 const TypePtr* tp = tn->type()->make_ptr();
1938 if (tp && tp->_ptr == TypePtr::AnyNull) {
1939 tn->set_type(TypeNarrowOop::NULL_PTR);
1940 }
1941 break;
1942 }
1943 case Op_Binary: // These are introduced in the Post_Visit state.
1944 ShouldNotReachHere();
1945 break;
1946 case Op_ClearArray:
1947 case Op_SafePoint:
1948 mem_op = true;
1949 break;
1950 default:
1951 if( n->is_Store() ) {
1952 // Do match stores, despite no ideal reg
1953 mem_op = true;
1954 break;
1955 }
1956 if( n->is_Mem() ) { // Loads and LoadStores
1957 mem_op = true;
1958 // Loads must be root of match tree due to prior load conflict
1959 if( C->subsume_loads() == false )
1960 set_shared(n);
1961 }
1962 // Fall into default case
1963 if( !n->ideal_reg() )
1964 set_dontcare(n); // Unmatchable Nodes
1965 } // end_switch
1967 for(int i = n->req() - 1; i >= 0; --i) { // For my children
1968 Node *m = n->in(i); // Get ith input
1969 if (m == NULL) continue; // Ignore NULLs
1970 uint mop = m->Opcode();
1972 // Must clone all producers of flags, or we will not match correctly.
1973 // Suppose a compare setting int-flags is shared (e.g., a switch-tree)
1974 // then it will match into an ideal Op_RegFlags. Alas, the fp-flags
1975 // are also there, so we may match a float-branch to int-flags and
1976 // expect the allocator to haul the flags from the int-side to the
1977 // fp-side. No can do.
1978 if( _must_clone[mop] ) {
1979 mstack.push(m, Visit);
1980 continue; // for(int i = ...)
1981 }
1983 if( mop == Op_AddP && m->in(AddPNode::Base)->Opcode() == Op_DecodeN ) {
1984 // Bases used in addresses must be shared but since
1985 // they are shared through a DecodeN they may appear
1986 // to have a single use so force sharing here.
1987 set_shared(m->in(AddPNode::Base)->in(1));
1988 }
1990 // Clone addressing expressions as they are "free" in memory access instructions
1991 if( mem_op && i == MemNode::Address && mop == Op_AddP ) {
1992 // Some inputs for address expression are not put on stack
1993 // to avoid marking them as shared and forcing them into register
1994 // if they are used only in address expressions.
1995 // But they should be marked as shared if there are other uses
1996 // besides address expressions.
1998 Node *off = m->in(AddPNode::Offset);
1999 if( off->is_Con() &&
2000 // When there are other uses besides address expressions
2001 // put it on stack and mark as shared.
2002 !is_visited(m) ) {
2003 address_visited.test_set(m->_idx); // Flag as address_visited
2004 Node *adr = m->in(AddPNode::Address);
2006 // Intel, ARM and friends can handle 2 adds in addressing mode
2007 if( clone_shift_expressions && adr->is_AddP() &&
2008 // AtomicAdd is not an addressing expression.
2009 // Cheap to find it by looking for screwy base.
2010 !adr->in(AddPNode::Base)->is_top() &&
2011 // Are there other uses besides address expressions?
2012 !is_visited(adr) ) {
2013 address_visited.set(adr->_idx); // Flag as address_visited
2014 Node *shift = adr->in(AddPNode::Offset);
2015 // Check for shift by small constant as well
2016 if( shift->Opcode() == Op_LShiftX && shift->in(2)->is_Con() &&
2017 shift->in(2)->get_int() <= 3 &&
2018 // Are there other uses besides address expressions?
2019 !is_visited(shift) ) {
2020 address_visited.set(shift->_idx); // Flag as address_visited
2021 mstack.push(shift->in(2), Visit);
2022 Node *conv = shift->in(1);
2023 #ifdef _LP64
2024 // Allow Matcher to match the rule which bypass
2025 // ConvI2L operation for an array index on LP64
2026 // if the index value is positive.
2027 if( conv->Opcode() == Op_ConvI2L &&
2028 conv->as_Type()->type()->is_long()->_lo >= 0 &&
2029 // Are there other uses besides address expressions?
2030 !is_visited(conv) ) {
2031 address_visited.set(conv->_idx); // Flag as address_visited
2032 mstack.push(conv->in(1), Pre_Visit);
2033 } else
2034 #endif
2035 mstack.push(conv, Pre_Visit);
2036 } else {
2037 mstack.push(shift, Pre_Visit);
2038 }
2039 mstack.push(adr->in(AddPNode::Address), Pre_Visit);
2040 mstack.push(adr->in(AddPNode::Base), Pre_Visit);
2041 } else { // Sparc, Alpha, PPC and friends
2042 mstack.push(adr, Pre_Visit);
2043 }
2045 // Clone X+offset as it also folds into most addressing expressions
2046 mstack.push(off, Visit);
2047 mstack.push(m->in(AddPNode::Base), Pre_Visit);
2048 continue; // for(int i = ...)
2049 } // if( off->is_Con() )
2050 } // if( mem_op &&
2051 mstack.push(m, Pre_Visit);
2052 } // for(int i = ...)
2053 }
2054 else if (nstate == Alt_Post_Visit) {
2055 mstack.pop(); // Remove node from stack
2056 // We cannot remove the Cmp input from the Bool here, as the Bool may be
2057 // shared and all users of the Bool need to move the Cmp in parallel.
2058 // This leaves both the Bool and the If pointing at the Cmp. To
2059 // prevent the Matcher from trying to Match the Cmp along both paths
2060 // BoolNode::match_edge always returns a zero.
2062 // We reorder the Op_If in a pre-order manner, so we can visit without
2063 // accidentally sharing the Cmp (the Bool and the If make 2 users).
2064 n->add_req( n->in(1)->in(1) ); // Add the Cmp next to the Bool
2065 }
2066 else if (nstate == Post_Visit) {
2067 mstack.pop(); // Remove node from stack
2069 // Now hack a few special opcodes
2070 switch( n->Opcode() ) { // Handle some opcodes special
2071 case Op_StorePConditional:
2072 case Op_StoreIConditional:
2073 case Op_StoreLConditional:
2074 case Op_CompareAndSwapI:
2075 case Op_CompareAndSwapL:
2076 case Op_CompareAndSwapP:
2077 case Op_CompareAndSwapN: { // Convert trinary to binary-tree
2078 Node *newval = n->in(MemNode::ValueIn );
2079 Node *oldval = n->in(LoadStoreNode::ExpectedIn);
2080 Node *pair = new (C, 3) BinaryNode( oldval, newval );
2081 n->set_req(MemNode::ValueIn,pair);
2082 n->del_req(LoadStoreNode::ExpectedIn);
2083 break;
2084 }
2085 case Op_CMoveD: // Convert trinary to binary-tree
2086 case Op_CMoveF:
2087 case Op_CMoveI:
2088 case Op_CMoveL:
2089 case Op_CMoveN:
2090 case Op_CMoveP: {
2091 // Restructure into a binary tree for Matching. It's possible that
2092 // we could move this code up next to the graph reshaping for IfNodes
2093 // or vice-versa, but I do not want to debug this for Ladybird.
2094 // 10/2/2000 CNC.
2095 Node *pair1 = new (C, 3) BinaryNode(n->in(1),n->in(1)->in(1));
2096 n->set_req(1,pair1);
2097 Node *pair2 = new (C, 3) BinaryNode(n->in(2),n->in(3));
2098 n->set_req(2,pair2);
2099 n->del_req(3);
2100 break;
2101 }
2102 case Op_LoopLimit: {
2103 Node *pair1 = new (C, 3) BinaryNode(n->in(1),n->in(2));
2104 n->set_req(1,pair1);
2105 n->set_req(2,n->in(3));
2106 n->del_req(3);
2107 break;
2108 }
2109 case Op_StrEquals: {
2110 Node *pair1 = new (C, 3) BinaryNode(n->in(2),n->in(3));
2111 n->set_req(2,pair1);
2112 n->set_req(3,n->in(4));
2113 n->del_req(4);
2114 break;
2115 }
2116 case Op_StrComp:
2117 case Op_StrIndexOf: {
2118 Node *pair1 = new (C, 3) BinaryNode(n->in(2),n->in(3));
2119 n->set_req(2,pair1);
2120 Node *pair2 = new (C, 3) BinaryNode(n->in(4),n->in(5));
2121 n->set_req(3,pair2);
2122 n->del_req(5);
2123 n->del_req(4);
2124 break;
2125 }
2126 default:
2127 break;
2128 }
2129 }
2130 else {
2131 ShouldNotReachHere();
2132 }
2133 } // end of while (mstack.is_nonempty())
2134 }
2136 #ifdef ASSERT
2137 // machine-independent root to machine-dependent root
2138 void Matcher::dump_old2new_map() {
2139 _old2new_map.dump();
2140 }
2141 #endif
2143 //---------------------------collect_null_checks-------------------------------
2144 // Find null checks in the ideal graph; write a machine-specific node for
2145 // it. Used by later implicit-null-check handling. Actually collects
2146 // either an IfTrue or IfFalse for the common NOT-null path, AND the ideal
2147 // value being tested.
2148 void Matcher::collect_null_checks( Node *proj, Node *orig_proj ) {
2149 Node *iff = proj->in(0);
2150 if( iff->Opcode() == Op_If ) {
2151 // During matching If's have Bool & Cmp side-by-side
2152 BoolNode *b = iff->in(1)->as_Bool();
2153 Node *cmp = iff->in(2);
2154 int opc = cmp->Opcode();
2155 if (opc != Op_CmpP && opc != Op_CmpN) return;
2157 const Type* ct = cmp->in(2)->bottom_type();
2158 if (ct == TypePtr::NULL_PTR ||
2159 (opc == Op_CmpN && ct == TypeNarrowOop::NULL_PTR)) {
2161 bool push_it = false;
2162 if( proj->Opcode() == Op_IfTrue ) {
2163 extern int all_null_checks_found;
2164 all_null_checks_found++;
2165 if( b->_test._test == BoolTest::ne ) {
2166 push_it = true;
2167 }
2168 } else {
2169 assert( proj->Opcode() == Op_IfFalse, "" );
2170 if( b->_test._test == BoolTest::eq ) {
2171 push_it = true;
2172 }
2173 }
2174 if( push_it ) {
2175 _null_check_tests.push(proj);
2176 Node* val = cmp->in(1);
2177 #ifdef _LP64
2178 if (val->bottom_type()->isa_narrowoop() &&
2179 !Matcher::narrow_oop_use_complex_address()) {
2180 //
2181 // Look for DecodeN node which should be pinned to orig_proj.
2182 // On platforms (Sparc) which can not handle 2 adds
2183 // in addressing mode we have to keep a DecodeN node and
2184 // use it to do implicit NULL check in address.
2185 //
2186 // DecodeN node was pinned to non-null path (orig_proj) during
2187 // CastPP transformation in final_graph_reshaping_impl().
2188 //
2189 uint cnt = orig_proj->outcnt();
2190 for (uint i = 0; i < orig_proj->outcnt(); i++) {
2191 Node* d = orig_proj->raw_out(i);
2192 if (d->is_DecodeN() && d->in(1) == val) {
2193 val = d;
2194 val->set_req(0, NULL); // Unpin now.
2195 // Mark this as special case to distinguish from
2196 // a regular case: CmpP(DecodeN, NULL).
2197 val = (Node*)(((intptr_t)val) | 1);
2198 break;
2199 }
2200 }
2201 }
2202 #endif
2203 _null_check_tests.push(val);
2204 }
2205 }
2206 }
2207 }
2209 //---------------------------validate_null_checks------------------------------
2210 // Its possible that the value being NULL checked is not the root of a match
2211 // tree. If so, I cannot use the value in an implicit null check.
2212 void Matcher::validate_null_checks( ) {
2213 uint cnt = _null_check_tests.size();
2214 for( uint i=0; i < cnt; i+=2 ) {
2215 Node *test = _null_check_tests[i];
2216 Node *val = _null_check_tests[i+1];
2217 bool is_decoden = ((intptr_t)val) & 1;
2218 val = (Node*)(((intptr_t)val) & ~1);
2219 if (has_new_node(val)) {
2220 Node* new_val = new_node(val);
2221 if (is_decoden) {
2222 assert(val->is_DecodeN() && val->in(0) == NULL, "sanity");
2223 // Note: new_val may have a control edge if
2224 // the original ideal node DecodeN was matched before
2225 // it was unpinned in Matcher::collect_null_checks().
2226 // Unpin the mach node and mark it.
2227 new_val->set_req(0, NULL);
2228 new_val = (Node*)(((intptr_t)new_val) | 1);
2229 }
2230 // Is a match-tree root, so replace with the matched value
2231 _null_check_tests.map(i+1, new_val);
2232 } else {
2233 // Yank from candidate list
2234 _null_check_tests.map(i+1,_null_check_tests[--cnt]);
2235 _null_check_tests.map(i,_null_check_tests[--cnt]);
2236 _null_check_tests.pop();
2237 _null_check_tests.pop();
2238 i-=2;
2239 }
2240 }
2241 }
2243 // Used by the DFA in dfa_xxx.cpp. Check for a following barrier or
2244 // atomic instruction acting as a store_load barrier without any
2245 // intervening volatile load, and thus we don't need a barrier here.
2246 // We retain the Node to act as a compiler ordering barrier.
2247 bool Matcher::post_store_load_barrier(const Node *vmb) {
2248 Compile *C = Compile::current();
2249 assert( vmb->is_MemBar(), "" );
2250 assert( vmb->Opcode() != Op_MemBarAcquire, "" );
2251 const MemBarNode *mem = (const MemBarNode*)vmb;
2253 // Get the Proj node, ctrl, that can be used to iterate forward
2254 Node *ctrl = NULL;
2255 DUIterator_Fast imax, i = mem->fast_outs(imax);
2256 while( true ) {
2257 ctrl = mem->fast_out(i); // Throw out-of-bounds if proj not found
2258 assert( ctrl->is_Proj(), "only projections here" );
2259 ProjNode *proj = (ProjNode*)ctrl;
2260 if( proj->_con == TypeFunc::Control &&
2261 !C->node_arena()->contains(ctrl) ) // Unmatched old-space only
2262 break;
2263 i++;
2264 }
2266 for( DUIterator_Fast jmax, j = ctrl->fast_outs(jmax); j < jmax; j++ ) {
2267 Node *x = ctrl->fast_out(j);
2268 int xop = x->Opcode();
2270 // We don't need current barrier if we see another or a lock
2271 // before seeing volatile load.
2272 //
2273 // Op_Fastunlock previously appeared in the Op_* list below.
2274 // With the advent of 1-0 lock operations we're no longer guaranteed
2275 // that a monitor exit operation contains a serializing instruction.
2277 if (xop == Op_MemBarVolatile ||
2278 xop == Op_FastLock ||
2279 xop == Op_CompareAndSwapL ||
2280 xop == Op_CompareAndSwapP ||
2281 xop == Op_CompareAndSwapN ||
2282 xop == Op_CompareAndSwapI)
2283 return true;
2285 if (x->is_MemBar()) {
2286 // We must retain this membar if there is an upcoming volatile
2287 // load, which will be preceded by acquire membar.
2288 if (xop == Op_MemBarAcquire)
2289 return false;
2290 // For other kinds of barriers, check by pretending we
2291 // are them, and seeing if we can be removed.
2292 else
2293 return post_store_load_barrier((const MemBarNode*)x);
2294 }
2296 // Delicate code to detect case of an upcoming fastlock block
2297 if( x->is_If() && x->req() > 1 &&
2298 !C->node_arena()->contains(x) ) { // Unmatched old-space only
2299 Node *iff = x;
2300 Node *bol = iff->in(1);
2301 // The iff might be some random subclass of If or bol might be Con-Top
2302 if (!bol->is_Bool()) return false;
2303 assert( bol->req() > 1, "" );
2304 return (bol->in(1)->Opcode() == Op_FastUnlock);
2305 }
2306 // probably not necessary to check for these
2307 if (x->is_Call() || x->is_SafePoint() || x->is_block_proj())
2308 return false;
2309 }
2310 return false;
2311 }
2313 //=============================================================================
2314 //---------------------------State---------------------------------------------
2315 State::State(void) {
2316 #ifdef ASSERT
2317 _id = 0;
2318 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe);
2319 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d);
2320 //memset(_cost, -1, sizeof(_cost));
2321 //memset(_rule, -1, sizeof(_rule));
2322 #endif
2323 memset(_valid, 0, sizeof(_valid));
2324 }
2326 #ifdef ASSERT
2327 State::~State() {
2328 _id = 99;
2329 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe);
2330 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d);
2331 memset(_cost, -3, sizeof(_cost));
2332 memset(_rule, -3, sizeof(_rule));
2333 }
2334 #endif
2336 #ifndef PRODUCT
2337 //---------------------------dump----------------------------------------------
2338 void State::dump() {
2339 tty->print("\n");
2340 dump(0);
2341 }
2343 void State::dump(int depth) {
2344 for( int j = 0; j < depth; j++ )
2345 tty->print(" ");
2346 tty->print("--N: ");
2347 _leaf->dump();
2348 uint i;
2349 for( i = 0; i < _LAST_MACH_OPER; i++ )
2350 // Check for valid entry
2351 if( valid(i) ) {
2352 for( int j = 0; j < depth; j++ )
2353 tty->print(" ");
2354 assert(_cost[i] != max_juint, "cost must be a valid value");
2355 assert(_rule[i] < _last_Mach_Node, "rule[i] must be valid rule");
2356 tty->print_cr("%s %d %s",
2357 ruleName[i], _cost[i], ruleName[_rule[i]] );
2358 }
2359 tty->print_cr("");
2361 for( i=0; i<2; i++ )
2362 if( _kids[i] )
2363 _kids[i]->dump(depth+1);
2364 }
2365 #endif