Tue, 17 May 2011 19:11:51 -0700
7045513: JSR 292 inlining causes crashes in methodHandleWalk.cpp
Reviewed-by: jrose
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
56 OptoReg::Name OptoReg::c_frame_pointer;
60 const int Matcher::base2reg[Type::lastype] = {
61 Node::NotAMachineReg,0,0, Op_RegI, Op_RegL, 0, Op_RegN,
62 Node::NotAMachineReg, Node::NotAMachineReg, /* tuple, array */
63 Op_RegP, Op_RegP, Op_RegP, Op_RegP, Op_RegP, Op_RegP, /* the pointers */
64 0, 0/*abio*/,
65 Op_RegP /* Return address */, 0, /* the memories */
66 Op_RegF, Op_RegF, Op_RegF, Op_RegD, Op_RegD, Op_RegD,
67 0 /*bottom*/
68 };
70 const RegMask *Matcher::idealreg2regmask[_last_machine_leaf];
71 RegMask Matcher::mreg2regmask[_last_Mach_Reg];
72 RegMask Matcher::STACK_ONLY_mask;
73 RegMask Matcher::c_frame_ptr_mask;
74 const uint Matcher::_begin_rematerialize = _BEGIN_REMATERIALIZE;
75 const uint Matcher::_end_rematerialize = _END_REMATERIALIZE;
77 //---------------------------Matcher-------------------------------------------
78 Matcher::Matcher( Node_List &proj_list ) :
79 PhaseTransform( Phase::Ins_Select ),
80 #ifdef ASSERT
81 _old2new_map(C->comp_arena()),
82 _new2old_map(C->comp_arena()),
83 #endif
84 _shared_nodes(C->comp_arena()),
85 _reduceOp(reduceOp), _leftOp(leftOp), _rightOp(rightOp),
86 _swallowed(swallowed),
87 _begin_inst_chain_rule(_BEGIN_INST_CHAIN_RULE),
88 _end_inst_chain_rule(_END_INST_CHAIN_RULE),
89 _must_clone(must_clone), _proj_list(proj_list),
90 _register_save_policy(register_save_policy),
91 _c_reg_save_policy(c_reg_save_policy),
92 _register_save_type(register_save_type),
93 _ruleName(ruleName),
94 _allocation_started(false),
95 _states_arena(Chunk::medium_size),
96 _visited(&_states_arena),
97 _shared(&_states_arena),
98 _dontcare(&_states_arena) {
99 C->set_matcher(this);
101 idealreg2spillmask [Op_RegI] = NULL;
102 idealreg2spillmask [Op_RegN] = NULL;
103 idealreg2spillmask [Op_RegL] = NULL;
104 idealreg2spillmask [Op_RegF] = NULL;
105 idealreg2spillmask [Op_RegD] = NULL;
106 idealreg2spillmask [Op_RegP] = NULL;
108 idealreg2debugmask [Op_RegI] = NULL;
109 idealreg2debugmask [Op_RegN] = NULL;
110 idealreg2debugmask [Op_RegL] = NULL;
111 idealreg2debugmask [Op_RegF] = NULL;
112 idealreg2debugmask [Op_RegD] = NULL;
113 idealreg2debugmask [Op_RegP] = NULL;
115 idealreg2mhdebugmask[Op_RegI] = NULL;
116 idealreg2mhdebugmask[Op_RegN] = NULL;
117 idealreg2mhdebugmask[Op_RegL] = NULL;
118 idealreg2mhdebugmask[Op_RegF] = NULL;
119 idealreg2mhdebugmask[Op_RegD] = NULL;
120 idealreg2mhdebugmask[Op_RegP] = NULL;
122 debug_only(_mem_node = NULL;) // Ideal memory node consumed by mach node
123 }
125 //------------------------------warp_incoming_stk_arg------------------------
126 // This warps a VMReg into an OptoReg::Name
127 OptoReg::Name Matcher::warp_incoming_stk_arg( VMReg reg ) {
128 OptoReg::Name warped;
129 if( reg->is_stack() ) { // Stack slot argument?
130 warped = OptoReg::add(_old_SP, reg->reg2stack() );
131 warped = OptoReg::add(warped, C->out_preserve_stack_slots());
132 if( warped >= _in_arg_limit )
133 _in_arg_limit = OptoReg::add(warped, 1); // Bump max stack slot seen
134 if (!RegMask::can_represent(warped)) {
135 // the compiler cannot represent this method's calling sequence
136 C->record_method_not_compilable_all_tiers("unsupported incoming calling sequence");
137 return OptoReg::Bad;
138 }
139 return warped;
140 }
141 return OptoReg::as_OptoReg(reg);
142 }
144 //---------------------------compute_old_SP------------------------------------
145 OptoReg::Name Compile::compute_old_SP() {
146 int fixed = fixed_slots();
147 int preserve = in_preserve_stack_slots();
148 return OptoReg::stack2reg(round_to(fixed + preserve, Matcher::stack_alignment_in_slots()));
149 }
153 #ifdef ASSERT
154 void Matcher::verify_new_nodes_only(Node* xroot) {
155 // Make sure that the new graph only references new nodes
156 ResourceMark rm;
157 Unique_Node_List worklist;
158 VectorSet visited(Thread::current()->resource_area());
159 worklist.push(xroot);
160 while (worklist.size() > 0) {
161 Node* n = worklist.pop();
162 visited <<= n->_idx;
163 assert(C->node_arena()->contains(n), "dead node");
164 for (uint j = 0; j < n->req(); j++) {
165 Node* in = n->in(j);
166 if (in != NULL) {
167 assert(C->node_arena()->contains(in), "dead node");
168 if (!visited.test(in->_idx)) {
169 worklist.push(in);
170 }
171 }
172 }
173 }
174 }
175 #endif
178 //---------------------------match---------------------------------------------
179 void Matcher::match( ) {
180 if( MaxLabelRootDepth < 100 ) { // Too small?
181 assert(false, "invalid MaxLabelRootDepth, increase it to 100 minimum");
182 MaxLabelRootDepth = 100;
183 }
184 // One-time initialization of some register masks.
185 init_spill_mask( C->root()->in(1) );
186 _return_addr_mask = return_addr();
187 #ifdef _LP64
188 // Pointers take 2 slots in 64-bit land
189 _return_addr_mask.Insert(OptoReg::add(return_addr(),1));
190 #endif
192 // Map a Java-signature return type into return register-value
193 // machine registers for 0, 1 and 2 returned values.
194 const TypeTuple *range = C->tf()->range();
195 if( range->cnt() > TypeFunc::Parms ) { // If not a void function
196 // Get ideal-register return type
197 int ireg = base2reg[range->field_at(TypeFunc::Parms)->base()];
198 // Get machine return register
199 uint sop = C->start()->Opcode();
200 OptoRegPair regs = return_value(ireg, false);
202 // And mask for same
203 _return_value_mask = RegMask(regs.first());
204 if( OptoReg::is_valid(regs.second()) )
205 _return_value_mask.Insert(regs.second());
206 }
208 // ---------------
209 // Frame Layout
211 // Need the method signature to determine the incoming argument types,
212 // because the types determine which registers the incoming arguments are
213 // in, and this affects the matched code.
214 const TypeTuple *domain = C->tf()->domain();
215 uint argcnt = domain->cnt() - TypeFunc::Parms;
216 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt );
217 VMRegPair *vm_parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt );
218 _parm_regs = NEW_RESOURCE_ARRAY( OptoRegPair, argcnt );
219 _calling_convention_mask = NEW_RESOURCE_ARRAY( RegMask, argcnt );
220 uint i;
221 for( i = 0; i<argcnt; i++ ) {
222 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type();
223 }
225 // Pass array of ideal registers and length to USER code (from the AD file)
226 // that will convert this to an array of register numbers.
227 const StartNode *start = C->start();
228 start->calling_convention( sig_bt, vm_parm_regs, argcnt );
229 #ifdef ASSERT
230 // Sanity check users' calling convention. Real handy while trying to
231 // get the initial port correct.
232 { for (uint i = 0; i<argcnt; i++) {
233 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) {
234 assert(domain->field_at(i+TypeFunc::Parms)==Type::HALF, "only allowed on halve" );
235 _parm_regs[i].set_bad();
236 continue;
237 }
238 VMReg parm_reg = vm_parm_regs[i].first();
239 assert(parm_reg->is_valid(), "invalid arg?");
240 if (parm_reg->is_reg()) {
241 OptoReg::Name opto_parm_reg = OptoReg::as_OptoReg(parm_reg);
242 assert(can_be_java_arg(opto_parm_reg) ||
243 C->stub_function() == CAST_FROM_FN_PTR(address, OptoRuntime::rethrow_C) ||
244 opto_parm_reg == inline_cache_reg(),
245 "parameters in register must be preserved by runtime stubs");
246 }
247 for (uint j = 0; j < i; j++) {
248 assert(parm_reg != vm_parm_regs[j].first(),
249 "calling conv. must produce distinct regs");
250 }
251 }
252 }
253 #endif
255 // Do some initial frame layout.
257 // Compute the old incoming SP (may be called FP) as
258 // OptoReg::stack0() + locks + in_preserve_stack_slots + pad2.
259 _old_SP = C->compute_old_SP();
260 assert( is_even(_old_SP), "must be even" );
262 // Compute highest incoming stack argument as
263 // _old_SP + out_preserve_stack_slots + incoming argument size.
264 _in_arg_limit = OptoReg::add(_old_SP, C->out_preserve_stack_slots());
265 assert( is_even(_in_arg_limit), "out_preserve must be even" );
266 for( i = 0; i < argcnt; i++ ) {
267 // Permit args to have no register
268 _calling_convention_mask[i].Clear();
269 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) {
270 continue;
271 }
272 // calling_convention returns stack arguments as a count of
273 // slots beyond OptoReg::stack0()/VMRegImpl::stack0. We need to convert this to
274 // the allocators point of view, taking into account all the
275 // preserve area, locks & pad2.
277 OptoReg::Name reg1 = warp_incoming_stk_arg(vm_parm_regs[i].first());
278 if( OptoReg::is_valid(reg1))
279 _calling_convention_mask[i].Insert(reg1);
281 OptoReg::Name reg2 = warp_incoming_stk_arg(vm_parm_regs[i].second());
282 if( OptoReg::is_valid(reg2))
283 _calling_convention_mask[i].Insert(reg2);
285 // Saved biased stack-slot register number
286 _parm_regs[i].set_pair(reg2, reg1);
287 }
289 // Finally, make sure the incoming arguments take up an even number of
290 // words, in case the arguments or locals need to contain doubleword stack
291 // slots. The rest of the system assumes that stack slot pairs (in
292 // particular, in the spill area) which look aligned will in fact be
293 // aligned relative to the stack pointer in the target machine. Double
294 // stack slots will always be allocated aligned.
295 _new_SP = OptoReg::Name(round_to(_in_arg_limit, RegMask::SlotsPerLong));
297 // Compute highest outgoing stack argument as
298 // _new_SP + out_preserve_stack_slots + max(outgoing argument size).
299 _out_arg_limit = OptoReg::add(_new_SP, C->out_preserve_stack_slots());
300 assert( is_even(_out_arg_limit), "out_preserve must be even" );
302 if (!RegMask::can_represent(OptoReg::add(_out_arg_limit,-1))) {
303 // the compiler cannot represent this method's calling sequence
304 C->record_method_not_compilable("must be able to represent all call arguments in reg mask");
305 }
307 if (C->failing()) return; // bailed out on incoming arg failure
309 // ---------------
310 // Collect roots of matcher trees. Every node for which
311 // _shared[_idx] is cleared is guaranteed to not be shared, and thus
312 // can be a valid interior of some tree.
313 find_shared( C->root() );
314 find_shared( C->top() );
316 C->print_method("Before Matching");
318 // Create new ideal node ConP #NULL even if it does exist in old space
319 // to avoid false sharing if the corresponding mach node is not used.
320 // The corresponding mach node is only used in rare cases for derived
321 // pointers.
322 Node* new_ideal_null = ConNode::make(C, TypePtr::NULL_PTR);
324 // Swap out to old-space; emptying new-space
325 Arena *old = C->node_arena()->move_contents(C->old_arena());
327 // Save debug and profile information for nodes in old space:
328 _old_node_note_array = C->node_note_array();
329 if (_old_node_note_array != NULL) {
330 C->set_node_note_array(new(C->comp_arena()) GrowableArray<Node_Notes*>
331 (C->comp_arena(), _old_node_note_array->length(),
332 0, NULL));
333 }
335 // Pre-size the new_node table to avoid the need for range checks.
336 grow_new_node_array(C->unique());
338 // Reset node counter so MachNodes start with _idx at 0
339 int nodes = C->unique(); // save value
340 C->set_unique(0);
342 // Recursively match trees from old space into new space.
343 // Correct leaves of new-space Nodes; they point to old-space.
344 _visited.Clear(); // Clear visit bits for xform call
345 C->set_cached_top_node(xform( C->top(), nodes ));
346 if (!C->failing()) {
347 Node* xroot = xform( C->root(), 1 );
348 if (xroot == NULL) {
349 Matcher::soft_match_failure(); // recursive matching process failed
350 C->record_method_not_compilable("instruction match failed");
351 } else {
352 // During matching shared constants were attached to C->root()
353 // because xroot wasn't available yet, so transfer the uses to
354 // the xroot.
355 for( DUIterator_Fast jmax, j = C->root()->fast_outs(jmax); j < jmax; j++ ) {
356 Node* n = C->root()->fast_out(j);
357 if (C->node_arena()->contains(n)) {
358 assert(n->in(0) == C->root(), "should be control user");
359 n->set_req(0, xroot);
360 --j;
361 --jmax;
362 }
363 }
365 // Generate new mach node for ConP #NULL
366 assert(new_ideal_null != NULL, "sanity");
367 _mach_null = match_tree(new_ideal_null);
368 // Don't set control, it will confuse GCM since there are no uses.
369 // The control will be set when this node is used first time
370 // in find_base_for_derived().
371 assert(_mach_null != NULL, "");
373 C->set_root(xroot->is_Root() ? xroot->as_Root() : NULL);
375 #ifdef ASSERT
376 verify_new_nodes_only(xroot);
377 #endif
378 }
379 }
380 if (C->top() == NULL || C->root() == NULL) {
381 C->record_method_not_compilable("graph lost"); // %%% cannot happen?
382 }
383 if (C->failing()) {
384 // delete old;
385 old->destruct_contents();
386 return;
387 }
388 assert( C->top(), "" );
389 assert( C->root(), "" );
390 validate_null_checks();
392 // Now smoke old-space
393 NOT_DEBUG( old->destruct_contents() );
395 // ------------------------
396 // Set up save-on-entry registers
397 Fixup_Save_On_Entry( );
398 }
401 //------------------------------Fixup_Save_On_Entry----------------------------
402 // The stated purpose of this routine is to take care of save-on-entry
403 // registers. However, the overall goal of the Match phase is to convert into
404 // machine-specific instructions which have RegMasks to guide allocation.
405 // So what this procedure really does is put a valid RegMask on each input
406 // to the machine-specific variations of all Return, TailCall and Halt
407 // instructions. It also adds edgs to define the save-on-entry values (and of
408 // course gives them a mask).
410 static RegMask *init_input_masks( uint size, RegMask &ret_adr, RegMask &fp ) {
411 RegMask *rms = NEW_RESOURCE_ARRAY( RegMask, size );
412 // Do all the pre-defined register masks
413 rms[TypeFunc::Control ] = RegMask::Empty;
414 rms[TypeFunc::I_O ] = RegMask::Empty;
415 rms[TypeFunc::Memory ] = RegMask::Empty;
416 rms[TypeFunc::ReturnAdr] = ret_adr;
417 rms[TypeFunc::FramePtr ] = fp;
418 return rms;
419 }
421 //---------------------------init_first_stack_mask-----------------------------
422 // Create the initial stack mask used by values spilling to the stack.
423 // Disallow any debug info in outgoing argument areas by setting the
424 // initial mask accordingly.
425 void Matcher::init_first_stack_mask() {
427 // Allocate storage for spill masks as masks for the appropriate load type.
428 RegMask *rms = (RegMask*)C->comp_arena()->Amalloc_D(sizeof(RegMask) * 3*6);
430 idealreg2spillmask [Op_RegN] = &rms[0];
431 idealreg2spillmask [Op_RegI] = &rms[1];
432 idealreg2spillmask [Op_RegL] = &rms[2];
433 idealreg2spillmask [Op_RegF] = &rms[3];
434 idealreg2spillmask [Op_RegD] = &rms[4];
435 idealreg2spillmask [Op_RegP] = &rms[5];
437 idealreg2debugmask [Op_RegN] = &rms[6];
438 idealreg2debugmask [Op_RegI] = &rms[7];
439 idealreg2debugmask [Op_RegL] = &rms[8];
440 idealreg2debugmask [Op_RegF] = &rms[9];
441 idealreg2debugmask [Op_RegD] = &rms[10];
442 idealreg2debugmask [Op_RegP] = &rms[11];
444 idealreg2mhdebugmask[Op_RegN] = &rms[12];
445 idealreg2mhdebugmask[Op_RegI] = &rms[13];
446 idealreg2mhdebugmask[Op_RegL] = &rms[14];
447 idealreg2mhdebugmask[Op_RegF] = &rms[15];
448 idealreg2mhdebugmask[Op_RegD] = &rms[16];
449 idealreg2mhdebugmask[Op_RegP] = &rms[17];
451 OptoReg::Name i;
453 // At first, start with the empty mask
454 C->FIRST_STACK_mask().Clear();
456 // Add in the incoming argument area
457 OptoReg::Name init = OptoReg::add(_old_SP, C->out_preserve_stack_slots());
458 for (i = init; i < _in_arg_limit; i = OptoReg::add(i,1))
459 C->FIRST_STACK_mask().Insert(i);
461 // Add in all bits past the outgoing argument area
462 guarantee(RegMask::can_represent(OptoReg::add(_out_arg_limit,-1)),
463 "must be able to represent all call arguments in reg mask");
464 init = _out_arg_limit;
465 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1))
466 C->FIRST_STACK_mask().Insert(i);
468 // Finally, set the "infinite stack" bit.
469 C->FIRST_STACK_mask().set_AllStack();
471 // Make spill masks. Registers for their class, plus FIRST_STACK_mask.
472 #ifdef _LP64
473 *idealreg2spillmask[Op_RegN] = *idealreg2regmask[Op_RegN];
474 idealreg2spillmask[Op_RegN]->OR(C->FIRST_STACK_mask());
475 #endif
476 *idealreg2spillmask[Op_RegI] = *idealreg2regmask[Op_RegI];
477 idealreg2spillmask[Op_RegI]->OR(C->FIRST_STACK_mask());
478 *idealreg2spillmask[Op_RegL] = *idealreg2regmask[Op_RegL];
479 idealreg2spillmask[Op_RegL]->OR(C->FIRST_STACK_mask());
480 *idealreg2spillmask[Op_RegF] = *idealreg2regmask[Op_RegF];
481 idealreg2spillmask[Op_RegF]->OR(C->FIRST_STACK_mask());
482 *idealreg2spillmask[Op_RegD] = *idealreg2regmask[Op_RegD];
483 idealreg2spillmask[Op_RegD]->OR(C->FIRST_STACK_mask());
484 *idealreg2spillmask[Op_RegP] = *idealreg2regmask[Op_RegP];
485 idealreg2spillmask[Op_RegP]->OR(C->FIRST_STACK_mask());
487 if (UseFPUForSpilling) {
488 // This mask logic assumes that the spill operations are
489 // symmetric and that the registers involved are the same size.
490 // On sparc for instance we may have to use 64 bit moves will
491 // kill 2 registers when used with F0-F31.
492 idealreg2spillmask[Op_RegI]->OR(*idealreg2regmask[Op_RegF]);
493 idealreg2spillmask[Op_RegF]->OR(*idealreg2regmask[Op_RegI]);
494 #ifdef _LP64
495 idealreg2spillmask[Op_RegN]->OR(*idealreg2regmask[Op_RegF]);
496 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]);
497 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]);
498 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegD]);
499 #else
500 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegF]);
501 #endif
502 }
504 // Make up debug masks. Any spill slot plus callee-save registers.
505 // Caller-save registers are assumed to be trashable by the various
506 // inline-cache fixup routines.
507 *idealreg2debugmask [Op_RegN]= *idealreg2spillmask[Op_RegN];
508 *idealreg2debugmask [Op_RegI]= *idealreg2spillmask[Op_RegI];
509 *idealreg2debugmask [Op_RegL]= *idealreg2spillmask[Op_RegL];
510 *idealreg2debugmask [Op_RegF]= *idealreg2spillmask[Op_RegF];
511 *idealreg2debugmask [Op_RegD]= *idealreg2spillmask[Op_RegD];
512 *idealreg2debugmask [Op_RegP]= *idealreg2spillmask[Op_RegP];
514 *idealreg2mhdebugmask[Op_RegN]= *idealreg2spillmask[Op_RegN];
515 *idealreg2mhdebugmask[Op_RegI]= *idealreg2spillmask[Op_RegI];
516 *idealreg2mhdebugmask[Op_RegL]= *idealreg2spillmask[Op_RegL];
517 *idealreg2mhdebugmask[Op_RegF]= *idealreg2spillmask[Op_RegF];
518 *idealreg2mhdebugmask[Op_RegD]= *idealreg2spillmask[Op_RegD];
519 *idealreg2mhdebugmask[Op_RegP]= *idealreg2spillmask[Op_RegP];
521 // Prevent stub compilations from attempting to reference
522 // callee-saved registers from debug info
523 bool exclude_soe = !Compile::current()->is_method_compilation();
525 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) {
526 // registers the caller has to save do not work
527 if( _register_save_policy[i] == 'C' ||
528 _register_save_policy[i] == 'A' ||
529 (_register_save_policy[i] == 'E' && exclude_soe) ) {
530 idealreg2debugmask [Op_RegN]->Remove(i);
531 idealreg2debugmask [Op_RegI]->Remove(i); // Exclude save-on-call
532 idealreg2debugmask [Op_RegL]->Remove(i); // registers from debug
533 idealreg2debugmask [Op_RegF]->Remove(i); // masks
534 idealreg2debugmask [Op_RegD]->Remove(i);
535 idealreg2debugmask [Op_RegP]->Remove(i);
537 idealreg2mhdebugmask[Op_RegN]->Remove(i);
538 idealreg2mhdebugmask[Op_RegI]->Remove(i);
539 idealreg2mhdebugmask[Op_RegL]->Remove(i);
540 idealreg2mhdebugmask[Op_RegF]->Remove(i);
541 idealreg2mhdebugmask[Op_RegD]->Remove(i);
542 idealreg2mhdebugmask[Op_RegP]->Remove(i);
543 }
544 }
546 // Subtract the register we use to save the SP for MethodHandle
547 // invokes to from the debug mask.
548 const RegMask save_mask = method_handle_invoke_SP_save_mask();
549 idealreg2mhdebugmask[Op_RegN]->SUBTRACT(save_mask);
550 idealreg2mhdebugmask[Op_RegI]->SUBTRACT(save_mask);
551 idealreg2mhdebugmask[Op_RegL]->SUBTRACT(save_mask);
552 idealreg2mhdebugmask[Op_RegF]->SUBTRACT(save_mask);
553 idealreg2mhdebugmask[Op_RegD]->SUBTRACT(save_mask);
554 idealreg2mhdebugmask[Op_RegP]->SUBTRACT(save_mask);
555 }
557 //---------------------------is_save_on_entry----------------------------------
558 bool Matcher::is_save_on_entry( int reg ) {
559 return
560 _register_save_policy[reg] == 'E' ||
561 _register_save_policy[reg] == 'A' || // Save-on-entry register?
562 // Also save argument registers in the trampolining stubs
563 (C->save_argument_registers() && is_spillable_arg(reg));
564 }
566 //---------------------------Fixup_Save_On_Entry-------------------------------
567 void Matcher::Fixup_Save_On_Entry( ) {
568 init_first_stack_mask();
570 Node *root = C->root(); // Short name for root
571 // Count number of save-on-entry registers.
572 uint soe_cnt = number_of_saved_registers();
573 uint i;
575 // Find the procedure Start Node
576 StartNode *start = C->start();
577 assert( start, "Expect a start node" );
579 // Save argument registers in the trampolining stubs
580 if( C->save_argument_registers() )
581 for( i = 0; i < _last_Mach_Reg; i++ )
582 if( is_spillable_arg(i) )
583 soe_cnt++;
585 // Input RegMask array shared by all Returns.
586 // The type for doubles and longs has a count of 2, but
587 // there is only 1 returned value
588 uint ret_edge_cnt = TypeFunc::Parms + ((C->tf()->range()->cnt() == TypeFunc::Parms) ? 0 : 1);
589 RegMask *ret_rms = init_input_masks( ret_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
590 // Returns have 0 or 1 returned values depending on call signature.
591 // Return register is specified by return_value in the AD file.
592 if (ret_edge_cnt > TypeFunc::Parms)
593 ret_rms[TypeFunc::Parms+0] = _return_value_mask;
595 // Input RegMask array shared by all Rethrows.
596 uint reth_edge_cnt = TypeFunc::Parms+1;
597 RegMask *reth_rms = init_input_masks( reth_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
598 // Rethrow takes exception oop only, but in the argument 0 slot.
599 reth_rms[TypeFunc::Parms] = mreg2regmask[find_receiver(false)];
600 #ifdef _LP64
601 // Need two slots for ptrs in 64-bit land
602 reth_rms[TypeFunc::Parms].Insert(OptoReg::add(OptoReg::Name(find_receiver(false)),1));
603 #endif
605 // Input RegMask array shared by all TailCalls
606 uint tail_call_edge_cnt = TypeFunc::Parms+2;
607 RegMask *tail_call_rms = init_input_masks( tail_call_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
609 // Input RegMask array shared by all TailJumps
610 uint tail_jump_edge_cnt = TypeFunc::Parms+2;
611 RegMask *tail_jump_rms = init_input_masks( tail_jump_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
613 // TailCalls have 2 returned values (target & moop), whose masks come
614 // from the usual MachNode/MachOper mechanism. Find a sample
615 // TailCall to extract these masks and put the correct masks into
616 // the tail_call_rms array.
617 for( i=1; i < root->req(); i++ ) {
618 MachReturnNode *m = root->in(i)->as_MachReturn();
619 if( m->ideal_Opcode() == Op_TailCall ) {
620 tail_call_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0);
621 tail_call_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1);
622 break;
623 }
624 }
626 // TailJumps have 2 returned values (target & ex_oop), whose masks come
627 // from the usual MachNode/MachOper mechanism. Find a sample
628 // TailJump to extract these masks and put the correct masks into
629 // the tail_jump_rms array.
630 for( i=1; i < root->req(); i++ ) {
631 MachReturnNode *m = root->in(i)->as_MachReturn();
632 if( m->ideal_Opcode() == Op_TailJump ) {
633 tail_jump_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0);
634 tail_jump_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1);
635 break;
636 }
637 }
639 // Input RegMask array shared by all Halts
640 uint halt_edge_cnt = TypeFunc::Parms;
641 RegMask *halt_rms = init_input_masks( halt_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
643 // Capture the return input masks into each exit flavor
644 for( i=1; i < root->req(); i++ ) {
645 MachReturnNode *exit = root->in(i)->as_MachReturn();
646 switch( exit->ideal_Opcode() ) {
647 case Op_Return : exit->_in_rms = ret_rms; break;
648 case Op_Rethrow : exit->_in_rms = reth_rms; break;
649 case Op_TailCall : exit->_in_rms = tail_call_rms; break;
650 case Op_TailJump : exit->_in_rms = tail_jump_rms; break;
651 case Op_Halt : exit->_in_rms = halt_rms; break;
652 default : ShouldNotReachHere();
653 }
654 }
656 // Next unused projection number from Start.
657 int proj_cnt = C->tf()->domain()->cnt();
659 // Do all the save-on-entry registers. Make projections from Start for
660 // them, and give them a use at the exit points. To the allocator, they
661 // look like incoming register arguments.
662 for( i = 0; i < _last_Mach_Reg; i++ ) {
663 if( is_save_on_entry(i) ) {
665 // Add the save-on-entry to the mask array
666 ret_rms [ ret_edge_cnt] = mreg2regmask[i];
667 reth_rms [ reth_edge_cnt] = mreg2regmask[i];
668 tail_call_rms[tail_call_edge_cnt] = mreg2regmask[i];
669 tail_jump_rms[tail_jump_edge_cnt] = mreg2regmask[i];
670 // Halts need the SOE registers, but only in the stack as debug info.
671 // A just-prior uncommon-trap or deoptimization will use the SOE regs.
672 halt_rms [ halt_edge_cnt] = *idealreg2spillmask[_register_save_type[i]];
674 Node *mproj;
676 // Is this a RegF low half of a RegD? Double up 2 adjacent RegF's
677 // into a single RegD.
678 if( (i&1) == 0 &&
679 _register_save_type[i ] == Op_RegF &&
680 _register_save_type[i+1] == Op_RegF &&
681 is_save_on_entry(i+1) ) {
682 // Add other bit for double
683 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1));
684 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1));
685 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1));
686 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1));
687 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1));
688 mproj = new (C, 1) MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegD );
689 proj_cnt += 2; // Skip 2 for doubles
690 }
691 else if( (i&1) == 1 && // Else check for high half of double
692 _register_save_type[i-1] == Op_RegF &&
693 _register_save_type[i ] == Op_RegF &&
694 is_save_on_entry(i-1) ) {
695 ret_rms [ ret_edge_cnt] = RegMask::Empty;
696 reth_rms [ reth_edge_cnt] = RegMask::Empty;
697 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty;
698 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty;
699 halt_rms [ halt_edge_cnt] = RegMask::Empty;
700 mproj = C->top();
701 }
702 // Is this a RegI low half of a RegL? Double up 2 adjacent RegI's
703 // into a single RegL.
704 else if( (i&1) == 0 &&
705 _register_save_type[i ] == Op_RegI &&
706 _register_save_type[i+1] == Op_RegI &&
707 is_save_on_entry(i+1) ) {
708 // Add other bit for long
709 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1));
710 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1));
711 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1));
712 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1));
713 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1));
714 mproj = new (C, 1) MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegL );
715 proj_cnt += 2; // Skip 2 for longs
716 }
717 else if( (i&1) == 1 && // Else check for high half of long
718 _register_save_type[i-1] == Op_RegI &&
719 _register_save_type[i ] == Op_RegI &&
720 is_save_on_entry(i-1) ) {
721 ret_rms [ ret_edge_cnt] = RegMask::Empty;
722 reth_rms [ reth_edge_cnt] = RegMask::Empty;
723 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty;
724 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty;
725 halt_rms [ halt_edge_cnt] = RegMask::Empty;
726 mproj = C->top();
727 } else {
728 // Make a projection for it off the Start
729 mproj = new (C, 1) MachProjNode( start, proj_cnt++, ret_rms[ret_edge_cnt], _register_save_type[i] );
730 }
732 ret_edge_cnt ++;
733 reth_edge_cnt ++;
734 tail_call_edge_cnt ++;
735 tail_jump_edge_cnt ++;
736 halt_edge_cnt ++;
738 // Add a use of the SOE register to all exit paths
739 for( uint j=1; j < root->req(); j++ )
740 root->in(j)->add_req(mproj);
741 } // End of if a save-on-entry register
742 } // End of for all machine registers
743 }
745 //------------------------------init_spill_mask--------------------------------
746 void Matcher::init_spill_mask( Node *ret ) {
747 if( idealreg2regmask[Op_RegI] ) return; // One time only init
749 OptoReg::c_frame_pointer = c_frame_pointer();
750 c_frame_ptr_mask = c_frame_pointer();
751 #ifdef _LP64
752 // pointers are twice as big
753 c_frame_ptr_mask.Insert(OptoReg::add(c_frame_pointer(),1));
754 #endif
756 // Start at OptoReg::stack0()
757 STACK_ONLY_mask.Clear();
758 OptoReg::Name init = OptoReg::stack2reg(0);
759 // STACK_ONLY_mask is all stack bits
760 OptoReg::Name i;
761 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1))
762 STACK_ONLY_mask.Insert(i);
763 // Also set the "infinite stack" bit.
764 STACK_ONLY_mask.set_AllStack();
766 // Copy the register names over into the shared world
767 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) {
768 // SharedInfo::regName[i] = regName[i];
769 // Handy RegMasks per machine register
770 mreg2regmask[i].Insert(i);
771 }
773 // Grab the Frame Pointer
774 Node *fp = ret->in(TypeFunc::FramePtr);
775 Node *mem = ret->in(TypeFunc::Memory);
776 const TypePtr* atp = TypePtr::BOTTOM;
777 // Share frame pointer while making spill ops
778 set_shared(fp);
780 // Compute generic short-offset Loads
781 #ifdef _LP64
782 MachNode *spillCP = match_tree(new (C, 3) LoadNNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM));
783 #endif
784 MachNode *spillI = match_tree(new (C, 3) LoadINode(NULL,mem,fp,atp));
785 MachNode *spillL = match_tree(new (C, 3) LoadLNode(NULL,mem,fp,atp));
786 MachNode *spillF = match_tree(new (C, 3) LoadFNode(NULL,mem,fp,atp));
787 MachNode *spillD = match_tree(new (C, 3) LoadDNode(NULL,mem,fp,atp));
788 MachNode *spillP = match_tree(new (C, 3) LoadPNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM));
789 assert(spillI != NULL && spillL != NULL && spillF != NULL &&
790 spillD != NULL && spillP != NULL, "");
792 // Get the ADLC notion of the right regmask, for each basic type.
793 #ifdef _LP64
794 idealreg2regmask[Op_RegN] = &spillCP->out_RegMask();
795 #endif
796 idealreg2regmask[Op_RegI] = &spillI->out_RegMask();
797 idealreg2regmask[Op_RegL] = &spillL->out_RegMask();
798 idealreg2regmask[Op_RegF] = &spillF->out_RegMask();
799 idealreg2regmask[Op_RegD] = &spillD->out_RegMask();
800 idealreg2regmask[Op_RegP] = &spillP->out_RegMask();
801 }
803 #ifdef ASSERT
804 static void match_alias_type(Compile* C, Node* n, Node* m) {
805 if (!VerifyAliases) return; // do not go looking for trouble by default
806 const TypePtr* nat = n->adr_type();
807 const TypePtr* mat = m->adr_type();
808 int nidx = C->get_alias_index(nat);
809 int midx = C->get_alias_index(mat);
810 // Detune the assert for cases like (AndI 0xFF (LoadB p)).
811 if (nidx == Compile::AliasIdxTop && midx >= Compile::AliasIdxRaw) {
812 for (uint i = 1; i < n->req(); i++) {
813 Node* n1 = n->in(i);
814 const TypePtr* n1at = n1->adr_type();
815 if (n1at != NULL) {
816 nat = n1at;
817 nidx = C->get_alias_index(n1at);
818 }
819 }
820 }
821 // %%% Kludgery. Instead, fix ideal adr_type methods for all these cases:
822 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxRaw) {
823 switch (n->Opcode()) {
824 case Op_PrefetchRead:
825 case Op_PrefetchWrite:
826 nidx = Compile::AliasIdxRaw;
827 nat = TypeRawPtr::BOTTOM;
828 break;
829 }
830 }
831 if (nidx == Compile::AliasIdxRaw && midx == Compile::AliasIdxTop) {
832 switch (n->Opcode()) {
833 case Op_ClearArray:
834 midx = Compile::AliasIdxRaw;
835 mat = TypeRawPtr::BOTTOM;
836 break;
837 }
838 }
839 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxBot) {
840 switch (n->Opcode()) {
841 case Op_Return:
842 case Op_Rethrow:
843 case Op_Halt:
844 case Op_TailCall:
845 case Op_TailJump:
846 nidx = Compile::AliasIdxBot;
847 nat = TypePtr::BOTTOM;
848 break;
849 }
850 }
851 if (nidx == Compile::AliasIdxBot && midx == Compile::AliasIdxTop) {
852 switch (n->Opcode()) {
853 case Op_StrComp:
854 case Op_StrEquals:
855 case Op_StrIndexOf:
856 case Op_AryEq:
857 case Op_MemBarVolatile:
858 case Op_MemBarCPUOrder: // %%% these ideals should have narrower adr_type?
859 nidx = Compile::AliasIdxTop;
860 nat = NULL;
861 break;
862 }
863 }
864 if (nidx != midx) {
865 if (PrintOpto || (PrintMiscellaneous && (WizardMode || Verbose))) {
866 tty->print_cr("==== Matcher alias shift %d => %d", nidx, midx);
867 n->dump();
868 m->dump();
869 }
870 assert(C->subsume_loads() && C->must_alias(nat, midx),
871 "must not lose alias info when matching");
872 }
873 }
874 #endif
877 //------------------------------MStack-----------------------------------------
878 // State and MStack class used in xform() and find_shared() iterative methods.
879 enum Node_State { Pre_Visit, // node has to be pre-visited
880 Visit, // visit node
881 Post_Visit, // post-visit node
882 Alt_Post_Visit // alternative post-visit path
883 };
885 class MStack: public Node_Stack {
886 public:
887 MStack(int size) : Node_Stack(size) { }
889 void push(Node *n, Node_State ns) {
890 Node_Stack::push(n, (uint)ns);
891 }
892 void push(Node *n, Node_State ns, Node *parent, int indx) {
893 ++_inode_top;
894 if ((_inode_top + 1) >= _inode_max) grow();
895 _inode_top->node = parent;
896 _inode_top->indx = (uint)indx;
897 ++_inode_top;
898 _inode_top->node = n;
899 _inode_top->indx = (uint)ns;
900 }
901 Node *parent() {
902 pop();
903 return node();
904 }
905 Node_State state() const {
906 return (Node_State)index();
907 }
908 void set_state(Node_State ns) {
909 set_index((uint)ns);
910 }
911 };
914 //------------------------------xform------------------------------------------
915 // Given a Node in old-space, Match him (Label/Reduce) to produce a machine
916 // Node in new-space. Given a new-space Node, recursively walk his children.
917 Node *Matcher::transform( Node *n ) { ShouldNotCallThis(); return n; }
918 Node *Matcher::xform( Node *n, int max_stack ) {
919 // Use one stack to keep both: child's node/state and parent's node/index
920 MStack mstack(max_stack * 2 * 2); // C->unique() * 2 * 2
921 mstack.push(n, Visit, NULL, -1); // set NULL as parent to indicate root
923 while (mstack.is_nonempty()) {
924 n = mstack.node(); // Leave node on stack
925 Node_State nstate = mstack.state();
926 if (nstate == Visit) {
927 mstack.set_state(Post_Visit);
928 Node *oldn = n;
929 // Old-space or new-space check
930 if (!C->node_arena()->contains(n)) {
931 // Old space!
932 Node* m;
933 if (has_new_node(n)) { // Not yet Label/Reduced
934 m = new_node(n);
935 } else {
936 if (!is_dontcare(n)) { // Matcher can match this guy
937 // Calls match special. They match alone with no children.
938 // Their children, the incoming arguments, match normally.
939 m = n->is_SafePoint() ? match_sfpt(n->as_SafePoint()):match_tree(n);
940 if (C->failing()) return NULL;
941 if (m == NULL) { Matcher::soft_match_failure(); return NULL; }
942 } else { // Nothing the matcher cares about
943 if( n->is_Proj() && n->in(0)->is_Multi()) { // Projections?
944 // Convert to machine-dependent projection
945 m = n->in(0)->as_Multi()->match( n->as_Proj(), this );
946 #ifdef ASSERT
947 _new2old_map.map(m->_idx, n);
948 #endif
949 if (m->in(0) != NULL) // m might be top
950 collect_null_checks(m, n);
951 } else { // Else just a regular 'ol guy
952 m = n->clone(); // So just clone into new-space
953 #ifdef ASSERT
954 _new2old_map.map(m->_idx, n);
955 #endif
956 // Def-Use edges will be added incrementally as Uses
957 // of this node are matched.
958 assert(m->outcnt() == 0, "no Uses of this clone yet");
959 }
960 }
962 set_new_node(n, m); // Map old to new
963 if (_old_node_note_array != NULL) {
964 Node_Notes* nn = C->locate_node_notes(_old_node_note_array,
965 n->_idx);
966 C->set_node_notes_at(m->_idx, nn);
967 }
968 debug_only(match_alias_type(C, n, m));
969 }
970 n = m; // n is now a new-space node
971 mstack.set_node(n);
972 }
974 // New space!
975 if (_visited.test_set(n->_idx)) continue; // while(mstack.is_nonempty())
977 int i;
978 // Put precedence edges on stack first (match them last).
979 for (i = oldn->req(); (uint)i < oldn->len(); i++) {
980 Node *m = oldn->in(i);
981 if (m == NULL) break;
982 // set -1 to call add_prec() instead of set_req() during Step1
983 mstack.push(m, Visit, n, -1);
984 }
986 // For constant debug info, I'd rather have unmatched constants.
987 int cnt = n->req();
988 JVMState* jvms = n->jvms();
989 int debug_cnt = jvms ? jvms->debug_start() : cnt;
991 // Now do only debug info. Clone constants rather than matching.
992 // Constants are represented directly in the debug info without
993 // the need for executable machine instructions.
994 // Monitor boxes are also represented directly.
995 for (i = cnt - 1; i >= debug_cnt; --i) { // For all debug inputs do
996 Node *m = n->in(i); // Get input
997 int op = m->Opcode();
998 assert((op == Op_BoxLock) == jvms->is_monitor_use(i), "boxes only at monitor sites");
999 if( op == Op_ConI || op == Op_ConP || op == Op_ConN ||
1000 op == Op_ConF || op == Op_ConD || op == Op_ConL
1001 // || op == Op_BoxLock // %%%% enable this and remove (+++) in chaitin.cpp
1002 ) {
1003 m = m->clone();
1004 #ifdef ASSERT
1005 _new2old_map.map(m->_idx, n);
1006 #endif
1007 mstack.push(m, Post_Visit, n, i); // Don't need to visit
1008 mstack.push(m->in(0), Visit, m, 0);
1009 } else {
1010 mstack.push(m, Visit, n, i);
1011 }
1012 }
1014 // And now walk his children, and convert his inputs to new-space.
1015 for( ; i >= 0; --i ) { // For all normal inputs do
1016 Node *m = n->in(i); // Get input
1017 if(m != NULL)
1018 mstack.push(m, Visit, n, i);
1019 }
1021 }
1022 else if (nstate == Post_Visit) {
1023 // Set xformed input
1024 Node *p = mstack.parent();
1025 if (p != NULL) { // root doesn't have parent
1026 int i = (int)mstack.index();
1027 if (i >= 0)
1028 p->set_req(i, n); // required input
1029 else if (i == -1)
1030 p->add_prec(n); // precedence input
1031 else
1032 ShouldNotReachHere();
1033 }
1034 mstack.pop(); // remove processed node from stack
1035 }
1036 else {
1037 ShouldNotReachHere();
1038 }
1039 } // while (mstack.is_nonempty())
1040 return n; // Return new-space Node
1041 }
1043 //------------------------------warp_outgoing_stk_arg------------------------
1044 OptoReg::Name Matcher::warp_outgoing_stk_arg( VMReg reg, OptoReg::Name begin_out_arg_area, OptoReg::Name &out_arg_limit_per_call ) {
1045 // Convert outgoing argument location to a pre-biased stack offset
1046 if (reg->is_stack()) {
1047 OptoReg::Name warped = reg->reg2stack();
1048 // Adjust the stack slot offset to be the register number used
1049 // by the allocator.
1050 warped = OptoReg::add(begin_out_arg_area, warped);
1051 // Keep track of the largest numbered stack slot used for an arg.
1052 // Largest used slot per call-site indicates the amount of stack
1053 // that is killed by the call.
1054 if( warped >= out_arg_limit_per_call )
1055 out_arg_limit_per_call = OptoReg::add(warped,1);
1056 if (!RegMask::can_represent(warped)) {
1057 C->record_method_not_compilable_all_tiers("unsupported calling sequence");
1058 return OptoReg::Bad;
1059 }
1060 return warped;
1061 }
1062 return OptoReg::as_OptoReg(reg);
1063 }
1066 //------------------------------match_sfpt-------------------------------------
1067 // Helper function to match call instructions. Calls match special.
1068 // They match alone with no children. Their children, the incoming
1069 // arguments, match normally.
1070 MachNode *Matcher::match_sfpt( SafePointNode *sfpt ) {
1071 MachSafePointNode *msfpt = NULL;
1072 MachCallNode *mcall = NULL;
1073 uint cnt;
1074 // Split out case for SafePoint vs Call
1075 CallNode *call;
1076 const TypeTuple *domain;
1077 ciMethod* method = NULL;
1078 bool is_method_handle_invoke = false; // for special kill effects
1079 if( sfpt->is_Call() ) {
1080 call = sfpt->as_Call();
1081 domain = call->tf()->domain();
1082 cnt = domain->cnt();
1084 // Match just the call, nothing else
1085 MachNode *m = match_tree(call);
1086 if (C->failing()) return NULL;
1087 if( m == NULL ) { Matcher::soft_match_failure(); return NULL; }
1089 // Copy data from the Ideal SafePoint to the machine version
1090 mcall = m->as_MachCall();
1092 mcall->set_tf( call->tf());
1093 mcall->set_entry_point(call->entry_point());
1094 mcall->set_cnt( call->cnt());
1096 if( mcall->is_MachCallJava() ) {
1097 MachCallJavaNode *mcall_java = mcall->as_MachCallJava();
1098 const CallJavaNode *call_java = call->as_CallJava();
1099 method = call_java->method();
1100 mcall_java->_method = method;
1101 mcall_java->_bci = call_java->_bci;
1102 mcall_java->_optimized_virtual = call_java->is_optimized_virtual();
1103 is_method_handle_invoke = call_java->is_method_handle_invoke();
1104 mcall_java->_method_handle_invoke = is_method_handle_invoke;
1105 if( mcall_java->is_MachCallStaticJava() )
1106 mcall_java->as_MachCallStaticJava()->_name =
1107 call_java->as_CallStaticJava()->_name;
1108 if( mcall_java->is_MachCallDynamicJava() )
1109 mcall_java->as_MachCallDynamicJava()->_vtable_index =
1110 call_java->as_CallDynamicJava()->_vtable_index;
1111 }
1112 else if( mcall->is_MachCallRuntime() ) {
1113 mcall->as_MachCallRuntime()->_name = call->as_CallRuntime()->_name;
1114 }
1115 msfpt = mcall;
1116 }
1117 // This is a non-call safepoint
1118 else {
1119 call = NULL;
1120 domain = NULL;
1121 MachNode *mn = match_tree(sfpt);
1122 if (C->failing()) return NULL;
1123 msfpt = mn->as_MachSafePoint();
1124 cnt = TypeFunc::Parms;
1125 }
1127 // Advertise the correct memory effects (for anti-dependence computation).
1128 msfpt->set_adr_type(sfpt->adr_type());
1130 // Allocate a private array of RegMasks. These RegMasks are not shared.
1131 msfpt->_in_rms = NEW_RESOURCE_ARRAY( RegMask, cnt );
1132 // Empty them all.
1133 memset( msfpt->_in_rms, 0, sizeof(RegMask)*cnt );
1135 // Do all the pre-defined non-Empty register masks
1136 msfpt->_in_rms[TypeFunc::ReturnAdr] = _return_addr_mask;
1137 msfpt->_in_rms[TypeFunc::FramePtr ] = c_frame_ptr_mask;
1139 // Place first outgoing argument can possibly be put.
1140 OptoReg::Name begin_out_arg_area = OptoReg::add(_new_SP, C->out_preserve_stack_slots());
1141 assert( is_even(begin_out_arg_area), "" );
1142 // Compute max outgoing register number per call site.
1143 OptoReg::Name out_arg_limit_per_call = begin_out_arg_area;
1144 // Calls to C may hammer extra stack slots above and beyond any arguments.
1145 // These are usually backing store for register arguments for varargs.
1146 if( call != NULL && call->is_CallRuntime() )
1147 out_arg_limit_per_call = OptoReg::add(out_arg_limit_per_call,C->varargs_C_out_slots_killed());
1150 // Do the normal argument list (parameters) register masks
1151 int argcnt = cnt - TypeFunc::Parms;
1152 if( argcnt > 0 ) { // Skip it all if we have no args
1153 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt );
1154 VMRegPair *parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt );
1155 int i;
1156 for( i = 0; i < argcnt; i++ ) {
1157 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type();
1158 }
1159 // V-call to pick proper calling convention
1160 call->calling_convention( sig_bt, parm_regs, argcnt );
1162 #ifdef ASSERT
1163 // Sanity check users' calling convention. Really handy during
1164 // the initial porting effort. Fairly expensive otherwise.
1165 { for (int i = 0; i<argcnt; i++) {
1166 if( !parm_regs[i].first()->is_valid() &&
1167 !parm_regs[i].second()->is_valid() ) continue;
1168 VMReg reg1 = parm_regs[i].first();
1169 VMReg reg2 = parm_regs[i].second();
1170 for (int j = 0; j < i; j++) {
1171 if( !parm_regs[j].first()->is_valid() &&
1172 !parm_regs[j].second()->is_valid() ) continue;
1173 VMReg reg3 = parm_regs[j].first();
1174 VMReg reg4 = parm_regs[j].second();
1175 if( !reg1->is_valid() ) {
1176 assert( !reg2->is_valid(), "valid halvsies" );
1177 } else if( !reg3->is_valid() ) {
1178 assert( !reg4->is_valid(), "valid halvsies" );
1179 } else {
1180 assert( reg1 != reg2, "calling conv. must produce distinct regs");
1181 assert( reg1 != reg3, "calling conv. must produce distinct regs");
1182 assert( reg1 != reg4, "calling conv. must produce distinct regs");
1183 assert( reg2 != reg3, "calling conv. must produce distinct regs");
1184 assert( reg2 != reg4 || !reg2->is_valid(), "calling conv. must produce distinct regs");
1185 assert( reg3 != reg4, "calling conv. must produce distinct regs");
1186 }
1187 }
1188 }
1189 }
1190 #endif
1192 // Visit each argument. Compute its outgoing register mask.
1193 // Return results now can have 2 bits returned.
1194 // Compute max over all outgoing arguments both per call-site
1195 // and over the entire method.
1196 for( i = 0; i < argcnt; i++ ) {
1197 // Address of incoming argument mask to fill in
1198 RegMask *rm = &mcall->_in_rms[i+TypeFunc::Parms];
1199 if( !parm_regs[i].first()->is_valid() &&
1200 !parm_regs[i].second()->is_valid() ) {
1201 continue; // Avoid Halves
1202 }
1203 // Grab first register, adjust stack slots and insert in mask.
1204 OptoReg::Name reg1 = warp_outgoing_stk_arg(parm_regs[i].first(), begin_out_arg_area, out_arg_limit_per_call );
1205 if (OptoReg::is_valid(reg1))
1206 rm->Insert( reg1 );
1207 // Grab second register (if any), adjust stack slots and insert in mask.
1208 OptoReg::Name reg2 = warp_outgoing_stk_arg(parm_regs[i].second(), begin_out_arg_area, out_arg_limit_per_call );
1209 if (OptoReg::is_valid(reg2))
1210 rm->Insert( reg2 );
1211 } // End of for all arguments
1213 // Compute number of stack slots needed to restore stack in case of
1214 // Pascal-style argument popping.
1215 mcall->_argsize = out_arg_limit_per_call - begin_out_arg_area;
1216 }
1218 if (is_method_handle_invoke) {
1219 // Kill some extra stack space in case method handles want to do
1220 // a little in-place argument insertion.
1221 int regs_per_word = NOT_LP64(1) LP64_ONLY(2); // %%% make a global const!
1222 out_arg_limit_per_call += MethodHandlePushLimit * regs_per_word;
1223 // Do not update mcall->_argsize because (a) the extra space is not
1224 // pushed as arguments and (b) _argsize is dead (not used anywhere).
1225 }
1227 // Compute the max stack slot killed by any call. These will not be
1228 // available for debug info, and will be used to adjust FIRST_STACK_mask
1229 // after all call sites have been visited.
1230 if( _out_arg_limit < out_arg_limit_per_call)
1231 _out_arg_limit = out_arg_limit_per_call;
1233 if (mcall) {
1234 // Kill the outgoing argument area, including any non-argument holes and
1235 // any legacy C-killed slots. Use Fat-Projections to do the killing.
1236 // Since the max-per-method covers the max-per-call-site and debug info
1237 // is excluded on the max-per-method basis, debug info cannot land in
1238 // this killed area.
1239 uint r_cnt = mcall->tf()->range()->cnt();
1240 MachProjNode *proj = new (C, 1) MachProjNode( mcall, r_cnt+10000, RegMask::Empty, MachProjNode::fat_proj );
1241 if (!RegMask::can_represent(OptoReg::Name(out_arg_limit_per_call-1))) {
1242 C->record_method_not_compilable_all_tiers("unsupported outgoing calling sequence");
1243 } else {
1244 for (int i = begin_out_arg_area; i < out_arg_limit_per_call; i++)
1245 proj->_rout.Insert(OptoReg::Name(i));
1246 }
1247 if( proj->_rout.is_NotEmpty() )
1248 _proj_list.push(proj);
1249 }
1250 // Transfer the safepoint information from the call to the mcall
1251 // Move the JVMState list
1252 msfpt->set_jvms(sfpt->jvms());
1253 for (JVMState* jvms = msfpt->jvms(); jvms; jvms = jvms->caller()) {
1254 jvms->set_map(sfpt);
1255 }
1257 // Debug inputs begin just after the last incoming parameter
1258 assert( (mcall == NULL) || (mcall->jvms() == NULL) ||
1259 (mcall->jvms()->debug_start() + mcall->_jvmadj == mcall->tf()->domain()->cnt()), "" );
1261 // Move the OopMap
1262 msfpt->_oop_map = sfpt->_oop_map;
1264 // Registers killed by the call are set in the local scheduling pass
1265 // of Global Code Motion.
1266 return msfpt;
1267 }
1269 //---------------------------match_tree----------------------------------------
1270 // Match a Ideal Node DAG - turn it into a tree; Label & Reduce. Used as part
1271 // of the whole-sale conversion from Ideal to Mach Nodes. Also used for
1272 // making GotoNodes while building the CFG and in init_spill_mask() to identify
1273 // a Load's result RegMask for memoization in idealreg2regmask[]
1274 MachNode *Matcher::match_tree( const Node *n ) {
1275 assert( n->Opcode() != Op_Phi, "cannot match" );
1276 assert( !n->is_block_start(), "cannot match" );
1277 // Set the mark for all locally allocated State objects.
1278 // When this call returns, the _states_arena arena will be reset
1279 // freeing all State objects.
1280 ResourceMark rm( &_states_arena );
1282 LabelRootDepth = 0;
1284 // StoreNodes require their Memory input to match any LoadNodes
1285 Node *mem = n->is_Store() ? n->in(MemNode::Memory) : (Node*)1 ;
1286 #ifdef ASSERT
1287 Node* save_mem_node = _mem_node;
1288 _mem_node = n->is_Store() ? (Node*)n : NULL;
1289 #endif
1290 // State object for root node of match tree
1291 // Allocate it on _states_arena - stack allocation can cause stack overflow.
1292 State *s = new (&_states_arena) State;
1293 s->_kids[0] = NULL;
1294 s->_kids[1] = NULL;
1295 s->_leaf = (Node*)n;
1296 // Label the input tree, allocating labels from top-level arena
1297 Label_Root( n, s, n->in(0), mem );
1298 if (C->failing()) return NULL;
1300 // The minimum cost match for the whole tree is found at the root State
1301 uint mincost = max_juint;
1302 uint cost = max_juint;
1303 uint i;
1304 for( i = 0; i < NUM_OPERANDS; i++ ) {
1305 if( s->valid(i) && // valid entry and
1306 s->_cost[i] < cost && // low cost and
1307 s->_rule[i] >= NUM_OPERANDS ) // not an operand
1308 cost = s->_cost[mincost=i];
1309 }
1310 if (mincost == max_juint) {
1311 #ifndef PRODUCT
1312 tty->print("No matching rule for:");
1313 s->dump();
1314 #endif
1315 Matcher::soft_match_failure();
1316 return NULL;
1317 }
1318 // Reduce input tree based upon the state labels to machine Nodes
1319 MachNode *m = ReduceInst( s, s->_rule[mincost], mem );
1320 #ifdef ASSERT
1321 _old2new_map.map(n->_idx, m);
1322 _new2old_map.map(m->_idx, (Node*)n);
1323 #endif
1325 // Add any Matcher-ignored edges
1326 uint cnt = n->req();
1327 uint start = 1;
1328 if( mem != (Node*)1 ) start = MemNode::Memory+1;
1329 if( n->is_AddP() ) {
1330 assert( mem == (Node*)1, "" );
1331 start = AddPNode::Base+1;
1332 }
1333 for( i = start; i < cnt; i++ ) {
1334 if( !n->match_edge(i) ) {
1335 if( i < m->req() )
1336 m->ins_req( i, n->in(i) );
1337 else
1338 m->add_req( n->in(i) );
1339 }
1340 }
1342 debug_only( _mem_node = save_mem_node; )
1343 return m;
1344 }
1347 //------------------------------match_into_reg---------------------------------
1348 // Choose to either match this Node in a register or part of the current
1349 // match tree. Return true for requiring a register and false for matching
1350 // as part of the current match tree.
1351 static bool match_into_reg( const Node *n, Node *m, Node *control, int i, bool shared ) {
1353 const Type *t = m->bottom_type();
1355 if( t->singleton() ) {
1356 // Never force constants into registers. Allow them to match as
1357 // constants or registers. Copies of the same value will share
1358 // the same register. See find_shared_node.
1359 return false;
1360 } else { // Not a constant
1361 // Stop recursion if they have different Controls.
1362 // Slot 0 of constants is not really a Control.
1363 if( control && m->in(0) && control != m->in(0) ) {
1365 // Actually, we can live with the most conservative control we
1366 // find, if it post-dominates the others. This allows us to
1367 // pick up load/op/store trees where the load can float a little
1368 // above the store.
1369 Node *x = control;
1370 const uint max_scan = 6; // Arbitrary scan cutoff
1371 uint j;
1372 for( j=0; j<max_scan; j++ ) {
1373 if( x->is_Region() ) // Bail out at merge points
1374 return true;
1375 x = x->in(0);
1376 if( x == m->in(0) ) // Does 'control' post-dominate
1377 break; // m->in(0)? If so, we can use it
1378 }
1379 if( j == max_scan ) // No post-domination before scan end?
1380 return true; // Then break the match tree up
1381 }
1382 if (m->is_DecodeN() && Matcher::narrow_oop_use_complex_address()) {
1383 // These are commonly used in address expressions and can
1384 // efficiently fold into them on X64 in some cases.
1385 return false;
1386 }
1387 }
1389 // Not forceable cloning. If shared, put it into a register.
1390 return shared;
1391 }
1394 //------------------------------Instruction Selection--------------------------
1395 // Label method walks a "tree" of nodes, using the ADLC generated DFA to match
1396 // ideal nodes to machine instructions. Trees are delimited by shared Nodes,
1397 // things the Matcher does not match (e.g., Memory), and things with different
1398 // Controls (hence forced into different blocks). We pass in the Control
1399 // selected for this entire State tree.
1401 // The Matcher works on Trees, but an Intel add-to-memory requires a DAG: the
1402 // Store and the Load must have identical Memories (as well as identical
1403 // pointers). Since the Matcher does not have anything for Memory (and
1404 // does not handle DAGs), I have to match the Memory input myself. If the
1405 // Tree root is a Store, I require all Loads to have the identical memory.
1406 Node *Matcher::Label_Root( const Node *n, State *svec, Node *control, const Node *mem){
1407 // Since Label_Root is a recursive function, its possible that we might run
1408 // out of stack space. See bugs 6272980 & 6227033 for more info.
1409 LabelRootDepth++;
1410 if (LabelRootDepth > MaxLabelRootDepth) {
1411 C->record_method_not_compilable_all_tiers("Out of stack space, increase MaxLabelRootDepth");
1412 return NULL;
1413 }
1414 uint care = 0; // Edges matcher cares about
1415 uint cnt = n->req();
1416 uint i = 0;
1418 // Examine children for memory state
1419 // Can only subsume a child into your match-tree if that child's memory state
1420 // is not modified along the path to another input.
1421 // It is unsafe even if the other inputs are separate roots.
1422 Node *input_mem = NULL;
1423 for( i = 1; i < cnt; i++ ) {
1424 if( !n->match_edge(i) ) continue;
1425 Node *m = n->in(i); // Get ith input
1426 assert( m, "expect non-null children" );
1427 if( m->is_Load() ) {
1428 if( input_mem == NULL ) {
1429 input_mem = m->in(MemNode::Memory);
1430 } else if( input_mem != m->in(MemNode::Memory) ) {
1431 input_mem = NodeSentinel;
1432 }
1433 }
1434 }
1436 for( i = 1; i < cnt; i++ ){// For my children
1437 if( !n->match_edge(i) ) continue;
1438 Node *m = n->in(i); // Get ith input
1439 // Allocate states out of a private arena
1440 State *s = new (&_states_arena) State;
1441 svec->_kids[care++] = s;
1442 assert( care <= 2, "binary only for now" );
1444 // Recursively label the State tree.
1445 s->_kids[0] = NULL;
1446 s->_kids[1] = NULL;
1447 s->_leaf = m;
1449 // Check for leaves of the State Tree; things that cannot be a part of
1450 // the current tree. If it finds any, that value is matched as a
1451 // register operand. If not, then the normal matching is used.
1452 if( match_into_reg(n, m, control, i, is_shared(m)) ||
1453 //
1454 // Stop recursion if this is LoadNode and the root of this tree is a
1455 // StoreNode and the load & store have different memories.
1456 ((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) ||
1457 // Can NOT include the match of a subtree when its memory state
1458 // is used by any of the other subtrees
1459 (input_mem == NodeSentinel) ) {
1460 #ifndef PRODUCT
1461 // Print when we exclude matching due to different memory states at input-loads
1462 if( PrintOpto && (Verbose && WizardMode) && (input_mem == NodeSentinel)
1463 && !((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) ) {
1464 tty->print_cr("invalid input_mem");
1465 }
1466 #endif
1467 // Switch to a register-only opcode; this value must be in a register
1468 // and cannot be subsumed as part of a larger instruction.
1469 s->DFA( m->ideal_reg(), m );
1471 } else {
1472 // If match tree has no control and we do, adopt it for entire tree
1473 if( control == NULL && m->in(0) != NULL && m->req() > 1 )
1474 control = m->in(0); // Pick up control
1475 // Else match as a normal part of the match tree.
1476 control = Label_Root(m,s,control,mem);
1477 if (C->failing()) return NULL;
1478 }
1479 }
1482 // Call DFA to match this node, and return
1483 svec->DFA( n->Opcode(), n );
1485 #ifdef ASSERT
1486 uint x;
1487 for( x = 0; x < _LAST_MACH_OPER; x++ )
1488 if( svec->valid(x) )
1489 break;
1491 if (x >= _LAST_MACH_OPER) {
1492 n->dump();
1493 svec->dump();
1494 assert( false, "bad AD file" );
1495 }
1496 #endif
1497 return control;
1498 }
1501 // Con nodes reduced using the same rule can share their MachNode
1502 // which reduces the number of copies of a constant in the final
1503 // program. The register allocator is free to split uses later to
1504 // split live ranges.
1505 MachNode* Matcher::find_shared_node(Node* leaf, uint rule) {
1506 if (!leaf->is_Con() && !leaf->is_DecodeN()) return NULL;
1508 // See if this Con has already been reduced using this rule.
1509 if (_shared_nodes.Size() <= leaf->_idx) return NULL;
1510 MachNode* last = (MachNode*)_shared_nodes.at(leaf->_idx);
1511 if (last != NULL && rule == last->rule()) {
1512 // Don't expect control change for DecodeN
1513 if (leaf->is_DecodeN())
1514 return last;
1515 // Get the new space root.
1516 Node* xroot = new_node(C->root());
1517 if (xroot == NULL) {
1518 // This shouldn't happen give the order of matching.
1519 return NULL;
1520 }
1522 // Shared constants need to have their control be root so they
1523 // can be scheduled properly.
1524 Node* control = last->in(0);
1525 if (control != xroot) {
1526 if (control == NULL || control == C->root()) {
1527 last->set_req(0, xroot);
1528 } else {
1529 assert(false, "unexpected control");
1530 return NULL;
1531 }
1532 }
1533 return last;
1534 }
1535 return NULL;
1536 }
1539 //------------------------------ReduceInst-------------------------------------
1540 // Reduce a State tree (with given Control) into a tree of MachNodes.
1541 // This routine (and it's cohort ReduceOper) convert Ideal Nodes into
1542 // complicated machine Nodes. Each MachNode covers some tree of Ideal Nodes.
1543 // Each MachNode has a number of complicated MachOper operands; each
1544 // MachOper also covers a further tree of Ideal Nodes.
1546 // The root of the Ideal match tree is always an instruction, so we enter
1547 // the recursion here. After building the MachNode, we need to recurse
1548 // the tree checking for these cases:
1549 // (1) Child is an instruction -
1550 // Build the instruction (recursively), add it as an edge.
1551 // Build a simple operand (register) to hold the result of the instruction.
1552 // (2) Child is an interior part of an instruction -
1553 // Skip over it (do nothing)
1554 // (3) Child is the start of a operand -
1555 // Build the operand, place it inside the instruction
1556 // Call ReduceOper.
1557 MachNode *Matcher::ReduceInst( State *s, int rule, Node *&mem ) {
1558 assert( rule >= NUM_OPERANDS, "called with operand rule" );
1560 MachNode* shared_node = find_shared_node(s->_leaf, rule);
1561 if (shared_node != NULL) {
1562 return shared_node;
1563 }
1565 // Build the object to represent this state & prepare for recursive calls
1566 MachNode *mach = s->MachNodeGenerator( rule, C );
1567 mach->_opnds[0] = s->MachOperGenerator( _reduceOp[rule], C );
1568 assert( mach->_opnds[0] != NULL, "Missing result operand" );
1569 Node *leaf = s->_leaf;
1570 // Check for instruction or instruction chain rule
1571 if( rule >= _END_INST_CHAIN_RULE || rule < _BEGIN_INST_CHAIN_RULE ) {
1572 assert(C->node_arena()->contains(s->_leaf) || !has_new_node(s->_leaf),
1573 "duplicating node that's already been matched");
1574 // Instruction
1575 mach->add_req( leaf->in(0) ); // Set initial control
1576 // Reduce interior of complex instruction
1577 ReduceInst_Interior( s, rule, mem, mach, 1 );
1578 } else {
1579 // Instruction chain rules are data-dependent on their inputs
1580 mach->add_req(0); // Set initial control to none
1581 ReduceInst_Chain_Rule( s, rule, mem, mach );
1582 }
1584 // If a Memory was used, insert a Memory edge
1585 if( mem != (Node*)1 ) {
1586 mach->ins_req(MemNode::Memory,mem);
1587 #ifdef ASSERT
1588 // Verify adr type after matching memory operation
1589 const MachOper* oper = mach->memory_operand();
1590 if (oper != NULL && oper != (MachOper*)-1) {
1591 // It has a unique memory operand. Find corresponding ideal mem node.
1592 Node* m = NULL;
1593 if (leaf->is_Mem()) {
1594 m = leaf;
1595 } else {
1596 m = _mem_node;
1597 assert(m != NULL && m->is_Mem(), "expecting memory node");
1598 }
1599 const Type* mach_at = mach->adr_type();
1600 // DecodeN node consumed by an address may have different type
1601 // then its input. Don't compare types for such case.
1602 if (m->adr_type() != mach_at &&
1603 (m->in(MemNode::Address)->is_DecodeN() ||
1604 m->in(MemNode::Address)->is_AddP() &&
1605 m->in(MemNode::Address)->in(AddPNode::Address)->is_DecodeN() ||
1606 m->in(MemNode::Address)->is_AddP() &&
1607 m->in(MemNode::Address)->in(AddPNode::Address)->is_AddP() &&
1608 m->in(MemNode::Address)->in(AddPNode::Address)->in(AddPNode::Address)->is_DecodeN())) {
1609 mach_at = m->adr_type();
1610 }
1611 if (m->adr_type() != mach_at) {
1612 m->dump();
1613 tty->print_cr("mach:");
1614 mach->dump(1);
1615 }
1616 assert(m->adr_type() == mach_at, "matcher should not change adr type");
1617 }
1618 #endif
1619 }
1621 // If the _leaf is an AddP, insert the base edge
1622 if( leaf->is_AddP() )
1623 mach->ins_req(AddPNode::Base,leaf->in(AddPNode::Base));
1625 uint num_proj = _proj_list.size();
1627 // Perform any 1-to-many expansions required
1628 MachNode *ex = mach->Expand(s,_proj_list, mem);
1629 if( ex != mach ) {
1630 assert(ex->ideal_reg() == mach->ideal_reg(), "ideal types should match");
1631 if( ex->in(1)->is_Con() )
1632 ex->in(1)->set_req(0, C->root());
1633 // Remove old node from the graph
1634 for( uint i=0; i<mach->req(); i++ ) {
1635 mach->set_req(i,NULL);
1636 }
1637 #ifdef ASSERT
1638 _new2old_map.map(ex->_idx, s->_leaf);
1639 #endif
1640 }
1642 // PhaseChaitin::fixup_spills will sometimes generate spill code
1643 // via the matcher. By the time, nodes have been wired into the CFG,
1644 // and any further nodes generated by expand rules will be left hanging
1645 // in space, and will not get emitted as output code. Catch this.
1646 // Also, catch any new register allocation constraints ("projections")
1647 // generated belatedly during spill code generation.
1648 if (_allocation_started) {
1649 guarantee(ex == mach, "no expand rules during spill generation");
1650 guarantee(_proj_list.size() == num_proj, "no allocation during spill generation");
1651 }
1653 if (leaf->is_Con() || leaf->is_DecodeN()) {
1654 // Record the con for sharing
1655 _shared_nodes.map(leaf->_idx, ex);
1656 }
1658 return ex;
1659 }
1661 void Matcher::ReduceInst_Chain_Rule( State *s, int rule, Node *&mem, MachNode *mach ) {
1662 // 'op' is what I am expecting to receive
1663 int op = _leftOp[rule];
1664 // Operand type to catch childs result
1665 // This is what my child will give me.
1666 int opnd_class_instance = s->_rule[op];
1667 // Choose between operand class or not.
1668 // This is what I will receive.
1669 int catch_op = (FIRST_OPERAND_CLASS <= op && op < NUM_OPERANDS) ? opnd_class_instance : op;
1670 // New rule for child. Chase operand classes to get the actual rule.
1671 int newrule = s->_rule[catch_op];
1673 if( newrule < NUM_OPERANDS ) {
1674 // Chain from operand or operand class, may be output of shared node
1675 assert( 0 <= opnd_class_instance && opnd_class_instance < NUM_OPERANDS,
1676 "Bad AD file: Instruction chain rule must chain from operand");
1677 // Insert operand into array of operands for this instruction
1678 mach->_opnds[1] = s->MachOperGenerator( opnd_class_instance, C );
1680 ReduceOper( s, newrule, mem, mach );
1681 } else {
1682 // Chain from the result of an instruction
1683 assert( newrule >= _LAST_MACH_OPER, "Do NOT chain from internal operand");
1684 mach->_opnds[1] = s->MachOperGenerator( _reduceOp[catch_op], C );
1685 Node *mem1 = (Node*)1;
1686 debug_only(Node *save_mem_node = _mem_node;)
1687 mach->add_req( ReduceInst(s, newrule, mem1) );
1688 debug_only(_mem_node = save_mem_node;)
1689 }
1690 return;
1691 }
1694 uint Matcher::ReduceInst_Interior( State *s, int rule, Node *&mem, MachNode *mach, uint num_opnds ) {
1695 if( s->_leaf->is_Load() ) {
1696 Node *mem2 = s->_leaf->in(MemNode::Memory);
1697 assert( mem == (Node*)1 || mem == mem2, "multiple Memories being matched at once?" );
1698 debug_only( if( mem == (Node*)1 ) _mem_node = s->_leaf;)
1699 mem = mem2;
1700 }
1701 if( s->_leaf->in(0) != NULL && s->_leaf->req() > 1) {
1702 if( mach->in(0) == NULL )
1703 mach->set_req(0, s->_leaf->in(0));
1704 }
1706 // Now recursively walk the state tree & add operand list.
1707 for( uint i=0; i<2; i++ ) { // binary tree
1708 State *newstate = s->_kids[i];
1709 if( newstate == NULL ) break; // Might only have 1 child
1710 // 'op' is what I am expecting to receive
1711 int op;
1712 if( i == 0 ) {
1713 op = _leftOp[rule];
1714 } else {
1715 op = _rightOp[rule];
1716 }
1717 // Operand type to catch childs result
1718 // This is what my child will give me.
1719 int opnd_class_instance = newstate->_rule[op];
1720 // Choose between operand class or not.
1721 // This is what I will receive.
1722 int catch_op = (op >= FIRST_OPERAND_CLASS && op < NUM_OPERANDS) ? opnd_class_instance : op;
1723 // New rule for child. Chase operand classes to get the actual rule.
1724 int newrule = newstate->_rule[catch_op];
1726 if( newrule < NUM_OPERANDS ) { // Operand/operandClass or internalOp/instruction?
1727 // Operand/operandClass
1728 // Insert operand into array of operands for this instruction
1729 mach->_opnds[num_opnds++] = newstate->MachOperGenerator( opnd_class_instance, C );
1730 ReduceOper( newstate, newrule, mem, mach );
1732 } else { // Child is internal operand or new instruction
1733 if( newrule < _LAST_MACH_OPER ) { // internal operand or instruction?
1734 // internal operand --> call ReduceInst_Interior
1735 // Interior of complex instruction. Do nothing but recurse.
1736 num_opnds = ReduceInst_Interior( newstate, newrule, mem, mach, num_opnds );
1737 } else {
1738 // instruction --> call build operand( ) to catch result
1739 // --> ReduceInst( newrule )
1740 mach->_opnds[num_opnds++] = s->MachOperGenerator( _reduceOp[catch_op], C );
1741 Node *mem1 = (Node*)1;
1742 debug_only(Node *save_mem_node = _mem_node;)
1743 mach->add_req( ReduceInst( newstate, newrule, mem1 ) );
1744 debug_only(_mem_node = save_mem_node;)
1745 }
1746 }
1747 assert( mach->_opnds[num_opnds-1], "" );
1748 }
1749 return num_opnds;
1750 }
1752 // This routine walks the interior of possible complex operands.
1753 // At each point we check our children in the match tree:
1754 // (1) No children -
1755 // We are a leaf; add _leaf field as an input to the MachNode
1756 // (2) Child is an internal operand -
1757 // Skip over it ( do nothing )
1758 // (3) Child is an instruction -
1759 // Call ReduceInst recursively and
1760 // and instruction as an input to the MachNode
1761 void Matcher::ReduceOper( State *s, int rule, Node *&mem, MachNode *mach ) {
1762 assert( rule < _LAST_MACH_OPER, "called with operand rule" );
1763 State *kid = s->_kids[0];
1764 assert( kid == NULL || s->_leaf->in(0) == NULL, "internal operands have no control" );
1766 // Leaf? And not subsumed?
1767 if( kid == NULL && !_swallowed[rule] ) {
1768 mach->add_req( s->_leaf ); // Add leaf pointer
1769 return; // Bail out
1770 }
1772 if( s->_leaf->is_Load() ) {
1773 assert( mem == (Node*)1, "multiple Memories being matched at once?" );
1774 mem = s->_leaf->in(MemNode::Memory);
1775 debug_only(_mem_node = s->_leaf;)
1776 }
1777 if( s->_leaf->in(0) && s->_leaf->req() > 1) {
1778 if( !mach->in(0) )
1779 mach->set_req(0,s->_leaf->in(0));
1780 else {
1781 assert( s->_leaf->in(0) == mach->in(0), "same instruction, differing controls?" );
1782 }
1783 }
1785 for( uint i=0; kid != NULL && i<2; kid = s->_kids[1], i++ ) { // binary tree
1786 int newrule;
1787 if( i == 0 )
1788 newrule = kid->_rule[_leftOp[rule]];
1789 else
1790 newrule = kid->_rule[_rightOp[rule]];
1792 if( newrule < _LAST_MACH_OPER ) { // Operand or instruction?
1793 // Internal operand; recurse but do nothing else
1794 ReduceOper( kid, newrule, mem, mach );
1796 } else { // Child is a new instruction
1797 // Reduce the instruction, and add a direct pointer from this
1798 // machine instruction to the newly reduced one.
1799 Node *mem1 = (Node*)1;
1800 debug_only(Node *save_mem_node = _mem_node;)
1801 mach->add_req( ReduceInst( kid, newrule, mem1 ) );
1802 debug_only(_mem_node = save_mem_node;)
1803 }
1804 }
1805 }
1808 // -------------------------------------------------------------------------
1809 // Java-Java calling convention
1810 // (what you use when Java calls Java)
1812 //------------------------------find_receiver----------------------------------
1813 // For a given signature, return the OptoReg for parameter 0.
1814 OptoReg::Name Matcher::find_receiver( bool is_outgoing ) {
1815 VMRegPair regs;
1816 BasicType sig_bt = T_OBJECT;
1817 calling_convention(&sig_bt, ®s, 1, is_outgoing);
1818 // Return argument 0 register. In the LP64 build pointers
1819 // take 2 registers, but the VM wants only the 'main' name.
1820 return OptoReg::as_OptoReg(regs.first());
1821 }
1823 // A method-klass-holder may be passed in the inline_cache_reg
1824 // and then expanded into the inline_cache_reg and a method_oop register
1825 // defined in ad_<arch>.cpp
1828 //------------------------------find_shared------------------------------------
1829 // Set bits if Node is shared or otherwise a root
1830 void Matcher::find_shared( Node *n ) {
1831 // Allocate stack of size C->unique() * 2 to avoid frequent realloc
1832 MStack mstack(C->unique() * 2);
1833 // Mark nodes as address_visited if they are inputs to an address expression
1834 VectorSet address_visited(Thread::current()->resource_area());
1835 mstack.push(n, Visit); // Don't need to pre-visit root node
1836 while (mstack.is_nonempty()) {
1837 n = mstack.node(); // Leave node on stack
1838 Node_State nstate = mstack.state();
1839 uint nop = n->Opcode();
1840 if (nstate == Pre_Visit) {
1841 if (address_visited.test(n->_idx)) { // Visited in address already?
1842 // Flag as visited and shared now.
1843 set_visited(n);
1844 }
1845 if (is_visited(n)) { // Visited already?
1846 // Node is shared and has no reason to clone. Flag it as shared.
1847 // This causes it to match into a register for the sharing.
1848 set_shared(n); // Flag as shared and
1849 mstack.pop(); // remove node from stack
1850 continue;
1851 }
1852 nstate = Visit; // Not already visited; so visit now
1853 }
1854 if (nstate == Visit) {
1855 mstack.set_state(Post_Visit);
1856 set_visited(n); // Flag as visited now
1857 bool mem_op = false;
1859 switch( nop ) { // Handle some opcodes special
1860 case Op_Phi: // Treat Phis as shared roots
1861 case Op_Parm:
1862 case Op_Proj: // All handled specially during matching
1863 case Op_SafePointScalarObject:
1864 set_shared(n);
1865 set_dontcare(n);
1866 break;
1867 case Op_If:
1868 case Op_CountedLoopEnd:
1869 mstack.set_state(Alt_Post_Visit); // Alternative way
1870 // Convert (If (Bool (CmpX A B))) into (If (Bool) (CmpX A B)). Helps
1871 // with matching cmp/branch in 1 instruction. The Matcher needs the
1872 // Bool and CmpX side-by-side, because it can only get at constants
1873 // that are at the leaves of Match trees, and the Bool's condition acts
1874 // as a constant here.
1875 mstack.push(n->in(1), Visit); // Clone the Bool
1876 mstack.push(n->in(0), Pre_Visit); // Visit control input
1877 continue; // while (mstack.is_nonempty())
1878 case Op_ConvI2D: // These forms efficiently match with a prior
1879 case Op_ConvI2F: // Load but not a following Store
1880 if( n->in(1)->is_Load() && // Prior load
1881 n->outcnt() == 1 && // Not already shared
1882 n->unique_out()->is_Store() ) // Following store
1883 set_shared(n); // Force it to be a root
1884 break;
1885 case Op_ReverseBytesI:
1886 case Op_ReverseBytesL:
1887 if( n->in(1)->is_Load() && // Prior load
1888 n->outcnt() == 1 ) // Not already shared
1889 set_shared(n); // Force it to be a root
1890 break;
1891 case Op_BoxLock: // Cant match until we get stack-regs in ADLC
1892 case Op_IfFalse:
1893 case Op_IfTrue:
1894 case Op_MachProj:
1895 case Op_MergeMem:
1896 case Op_Catch:
1897 case Op_CatchProj:
1898 case Op_CProj:
1899 case Op_JumpProj:
1900 case Op_JProj:
1901 case Op_NeverBranch:
1902 set_dontcare(n);
1903 break;
1904 case Op_Jump:
1905 mstack.push(n->in(1), Visit); // Switch Value
1906 mstack.push(n->in(0), Pre_Visit); // Visit Control input
1907 continue; // while (mstack.is_nonempty())
1908 case Op_StrComp:
1909 case Op_StrEquals:
1910 case Op_StrIndexOf:
1911 case Op_AryEq:
1912 set_shared(n); // Force result into register (it will be anyways)
1913 break;
1914 case Op_ConP: { // Convert pointers above the centerline to NUL
1915 TypeNode *tn = n->as_Type(); // Constants derive from type nodes
1916 const TypePtr* tp = tn->type()->is_ptr();
1917 if (tp->_ptr == TypePtr::AnyNull) {
1918 tn->set_type(TypePtr::NULL_PTR);
1919 }
1920 break;
1921 }
1922 case Op_ConN: { // Convert narrow pointers above the centerline to NUL
1923 TypeNode *tn = n->as_Type(); // Constants derive from type nodes
1924 const TypePtr* tp = tn->type()->make_ptr();
1925 if (tp && tp->_ptr == TypePtr::AnyNull) {
1926 tn->set_type(TypeNarrowOop::NULL_PTR);
1927 }
1928 break;
1929 }
1930 case Op_Binary: // These are introduced in the Post_Visit state.
1931 ShouldNotReachHere();
1932 break;
1933 case Op_ClearArray:
1934 case Op_SafePoint:
1935 mem_op = true;
1936 break;
1937 default:
1938 if( n->is_Store() ) {
1939 // Do match stores, despite no ideal reg
1940 mem_op = true;
1941 break;
1942 }
1943 if( n->is_Mem() ) { // Loads and LoadStores
1944 mem_op = true;
1945 // Loads must be root of match tree due to prior load conflict
1946 if( C->subsume_loads() == false )
1947 set_shared(n);
1948 }
1949 // Fall into default case
1950 if( !n->ideal_reg() )
1951 set_dontcare(n); // Unmatchable Nodes
1952 } // end_switch
1954 for(int i = n->req() - 1; i >= 0; --i) { // For my children
1955 Node *m = n->in(i); // Get ith input
1956 if (m == NULL) continue; // Ignore NULLs
1957 uint mop = m->Opcode();
1959 // Must clone all producers of flags, or we will not match correctly.
1960 // Suppose a compare setting int-flags is shared (e.g., a switch-tree)
1961 // then it will match into an ideal Op_RegFlags. Alas, the fp-flags
1962 // are also there, so we may match a float-branch to int-flags and
1963 // expect the allocator to haul the flags from the int-side to the
1964 // fp-side. No can do.
1965 if( _must_clone[mop] ) {
1966 mstack.push(m, Visit);
1967 continue; // for(int i = ...)
1968 }
1970 if( mop == Op_AddP && m->in(AddPNode::Base)->Opcode() == Op_DecodeN ) {
1971 // Bases used in addresses must be shared but since
1972 // they are shared through a DecodeN they may appear
1973 // to have a single use so force sharing here.
1974 set_shared(m->in(AddPNode::Base)->in(1));
1975 }
1977 // Clone addressing expressions as they are "free" in memory access instructions
1978 if( mem_op && i == MemNode::Address && mop == Op_AddP ) {
1979 // Some inputs for address expression are not put on stack
1980 // to avoid marking them as shared and forcing them into register
1981 // if they are used only in address expressions.
1982 // But they should be marked as shared if there are other uses
1983 // besides address expressions.
1985 Node *off = m->in(AddPNode::Offset);
1986 if( off->is_Con() &&
1987 // When there are other uses besides address expressions
1988 // put it on stack and mark as shared.
1989 !is_visited(m) ) {
1990 address_visited.test_set(m->_idx); // Flag as address_visited
1991 Node *adr = m->in(AddPNode::Address);
1993 // Intel, ARM and friends can handle 2 adds in addressing mode
1994 if( clone_shift_expressions && adr->is_AddP() &&
1995 // AtomicAdd is not an addressing expression.
1996 // Cheap to find it by looking for screwy base.
1997 !adr->in(AddPNode::Base)->is_top() &&
1998 // Are there other uses besides address expressions?
1999 !is_visited(adr) ) {
2000 address_visited.set(adr->_idx); // Flag as address_visited
2001 Node *shift = adr->in(AddPNode::Offset);
2002 // Check for shift by small constant as well
2003 if( shift->Opcode() == Op_LShiftX && shift->in(2)->is_Con() &&
2004 shift->in(2)->get_int() <= 3 &&
2005 // Are there other uses besides address expressions?
2006 !is_visited(shift) ) {
2007 address_visited.set(shift->_idx); // Flag as address_visited
2008 mstack.push(shift->in(2), Visit);
2009 Node *conv = shift->in(1);
2010 #ifdef _LP64
2011 // Allow Matcher to match the rule which bypass
2012 // ConvI2L operation for an array index on LP64
2013 // if the index value is positive.
2014 if( conv->Opcode() == Op_ConvI2L &&
2015 conv->as_Type()->type()->is_long()->_lo >= 0 &&
2016 // Are there other uses besides address expressions?
2017 !is_visited(conv) ) {
2018 address_visited.set(conv->_idx); // Flag as address_visited
2019 mstack.push(conv->in(1), Pre_Visit);
2020 } else
2021 #endif
2022 mstack.push(conv, Pre_Visit);
2023 } else {
2024 mstack.push(shift, Pre_Visit);
2025 }
2026 mstack.push(adr->in(AddPNode::Address), Pre_Visit);
2027 mstack.push(adr->in(AddPNode::Base), Pre_Visit);
2028 } else { // Sparc, Alpha, PPC and friends
2029 mstack.push(adr, Pre_Visit);
2030 }
2032 // Clone X+offset as it also folds into most addressing expressions
2033 mstack.push(off, Visit);
2034 mstack.push(m->in(AddPNode::Base), Pre_Visit);
2035 continue; // for(int i = ...)
2036 } // if( off->is_Con() )
2037 } // if( mem_op &&
2038 mstack.push(m, Pre_Visit);
2039 } // for(int i = ...)
2040 }
2041 else if (nstate == Alt_Post_Visit) {
2042 mstack.pop(); // Remove node from stack
2043 // We cannot remove the Cmp input from the Bool here, as the Bool may be
2044 // shared and all users of the Bool need to move the Cmp in parallel.
2045 // This leaves both the Bool and the If pointing at the Cmp. To
2046 // prevent the Matcher from trying to Match the Cmp along both paths
2047 // BoolNode::match_edge always returns a zero.
2049 // We reorder the Op_If in a pre-order manner, so we can visit without
2050 // accidentally sharing the Cmp (the Bool and the If make 2 users).
2051 n->add_req( n->in(1)->in(1) ); // Add the Cmp next to the Bool
2052 }
2053 else if (nstate == Post_Visit) {
2054 mstack.pop(); // Remove node from stack
2056 // Now hack a few special opcodes
2057 switch( n->Opcode() ) { // Handle some opcodes special
2058 case Op_StorePConditional:
2059 case Op_StoreIConditional:
2060 case Op_StoreLConditional:
2061 case Op_CompareAndSwapI:
2062 case Op_CompareAndSwapL:
2063 case Op_CompareAndSwapP:
2064 case Op_CompareAndSwapN: { // Convert trinary to binary-tree
2065 Node *newval = n->in(MemNode::ValueIn );
2066 Node *oldval = n->in(LoadStoreNode::ExpectedIn);
2067 Node *pair = new (C, 3) BinaryNode( oldval, newval );
2068 n->set_req(MemNode::ValueIn,pair);
2069 n->del_req(LoadStoreNode::ExpectedIn);
2070 break;
2071 }
2072 case Op_CMoveD: // Convert trinary to binary-tree
2073 case Op_CMoveF:
2074 case Op_CMoveI:
2075 case Op_CMoveL:
2076 case Op_CMoveN:
2077 case Op_CMoveP: {
2078 // Restructure into a binary tree for Matching. It's possible that
2079 // we could move this code up next to the graph reshaping for IfNodes
2080 // or vice-versa, but I do not want to debug this for Ladybird.
2081 // 10/2/2000 CNC.
2082 Node *pair1 = new (C, 3) BinaryNode(n->in(1),n->in(1)->in(1));
2083 n->set_req(1,pair1);
2084 Node *pair2 = new (C, 3) BinaryNode(n->in(2),n->in(3));
2085 n->set_req(2,pair2);
2086 n->del_req(3);
2087 break;
2088 }
2089 case Op_LoopLimit: {
2090 Node *pair1 = new (C, 3) BinaryNode(n->in(1),n->in(2));
2091 n->set_req(1,pair1);
2092 n->set_req(2,n->in(3));
2093 n->del_req(3);
2094 break;
2095 }
2096 case Op_StrEquals: {
2097 Node *pair1 = new (C, 3) BinaryNode(n->in(2),n->in(3));
2098 n->set_req(2,pair1);
2099 n->set_req(3,n->in(4));
2100 n->del_req(4);
2101 break;
2102 }
2103 case Op_StrComp:
2104 case Op_StrIndexOf: {
2105 Node *pair1 = new (C, 3) BinaryNode(n->in(2),n->in(3));
2106 n->set_req(2,pair1);
2107 Node *pair2 = new (C, 3) BinaryNode(n->in(4),n->in(5));
2108 n->set_req(3,pair2);
2109 n->del_req(5);
2110 n->del_req(4);
2111 break;
2112 }
2113 default:
2114 break;
2115 }
2116 }
2117 else {
2118 ShouldNotReachHere();
2119 }
2120 } // end of while (mstack.is_nonempty())
2121 }
2123 #ifdef ASSERT
2124 // machine-independent root to machine-dependent root
2125 void Matcher::dump_old2new_map() {
2126 _old2new_map.dump();
2127 }
2128 #endif
2130 //---------------------------collect_null_checks-------------------------------
2131 // Find null checks in the ideal graph; write a machine-specific node for
2132 // it. Used by later implicit-null-check handling. Actually collects
2133 // either an IfTrue or IfFalse for the common NOT-null path, AND the ideal
2134 // value being tested.
2135 void Matcher::collect_null_checks( Node *proj, Node *orig_proj ) {
2136 Node *iff = proj->in(0);
2137 if( iff->Opcode() == Op_If ) {
2138 // During matching If's have Bool & Cmp side-by-side
2139 BoolNode *b = iff->in(1)->as_Bool();
2140 Node *cmp = iff->in(2);
2141 int opc = cmp->Opcode();
2142 if (opc != Op_CmpP && opc != Op_CmpN) return;
2144 const Type* ct = cmp->in(2)->bottom_type();
2145 if (ct == TypePtr::NULL_PTR ||
2146 (opc == Op_CmpN && ct == TypeNarrowOop::NULL_PTR)) {
2148 bool push_it = false;
2149 if( proj->Opcode() == Op_IfTrue ) {
2150 extern int all_null_checks_found;
2151 all_null_checks_found++;
2152 if( b->_test._test == BoolTest::ne ) {
2153 push_it = true;
2154 }
2155 } else {
2156 assert( proj->Opcode() == Op_IfFalse, "" );
2157 if( b->_test._test == BoolTest::eq ) {
2158 push_it = true;
2159 }
2160 }
2161 if( push_it ) {
2162 _null_check_tests.push(proj);
2163 Node* val = cmp->in(1);
2164 #ifdef _LP64
2165 if (val->bottom_type()->isa_narrowoop() &&
2166 !Matcher::narrow_oop_use_complex_address()) {
2167 //
2168 // Look for DecodeN node which should be pinned to orig_proj.
2169 // On platforms (Sparc) which can not handle 2 adds
2170 // in addressing mode we have to keep a DecodeN node and
2171 // use it to do implicit NULL check in address.
2172 //
2173 // DecodeN node was pinned to non-null path (orig_proj) during
2174 // CastPP transformation in final_graph_reshaping_impl().
2175 //
2176 uint cnt = orig_proj->outcnt();
2177 for (uint i = 0; i < orig_proj->outcnt(); i++) {
2178 Node* d = orig_proj->raw_out(i);
2179 if (d->is_DecodeN() && d->in(1) == val) {
2180 val = d;
2181 val->set_req(0, NULL); // Unpin now.
2182 // Mark this as special case to distinguish from
2183 // a regular case: CmpP(DecodeN, NULL).
2184 val = (Node*)(((intptr_t)val) | 1);
2185 break;
2186 }
2187 }
2188 }
2189 #endif
2190 _null_check_tests.push(val);
2191 }
2192 }
2193 }
2194 }
2196 //---------------------------validate_null_checks------------------------------
2197 // Its possible that the value being NULL checked is not the root of a match
2198 // tree. If so, I cannot use the value in an implicit null check.
2199 void Matcher::validate_null_checks( ) {
2200 uint cnt = _null_check_tests.size();
2201 for( uint i=0; i < cnt; i+=2 ) {
2202 Node *test = _null_check_tests[i];
2203 Node *val = _null_check_tests[i+1];
2204 bool is_decoden = ((intptr_t)val) & 1;
2205 val = (Node*)(((intptr_t)val) & ~1);
2206 if (has_new_node(val)) {
2207 Node* new_val = new_node(val);
2208 if (is_decoden) {
2209 assert(val->is_DecodeN() && val->in(0) == NULL, "sanity");
2210 // Note: new_val may have a control edge if
2211 // the original ideal node DecodeN was matched before
2212 // it was unpinned in Matcher::collect_null_checks().
2213 // Unpin the mach node and mark it.
2214 new_val->set_req(0, NULL);
2215 new_val = (Node*)(((intptr_t)new_val) | 1);
2216 }
2217 // Is a match-tree root, so replace with the matched value
2218 _null_check_tests.map(i+1, new_val);
2219 } else {
2220 // Yank from candidate list
2221 _null_check_tests.map(i+1,_null_check_tests[--cnt]);
2222 _null_check_tests.map(i,_null_check_tests[--cnt]);
2223 _null_check_tests.pop();
2224 _null_check_tests.pop();
2225 i-=2;
2226 }
2227 }
2228 }
2231 // Used by the DFA in dfa_sparc.cpp. Check for a prior FastLock
2232 // acting as an Acquire and thus we don't need an Acquire here. We
2233 // retain the Node to act as a compiler ordering barrier.
2234 bool Matcher::prior_fast_lock( const Node *acq ) {
2235 Node *r = acq->in(0);
2236 if( !r->is_Region() || r->req() <= 1 ) return false;
2237 Node *proj = r->in(1);
2238 if( !proj->is_Proj() ) return false;
2239 Node *call = proj->in(0);
2240 if( !call->is_Call() || call->as_Call()->entry_point() != OptoRuntime::complete_monitor_locking_Java() )
2241 return false;
2243 return true;
2244 }
2246 // Used by the DFA in dfa_sparc.cpp. Check for a following FastUnLock
2247 // acting as a Release and thus we don't need a Release here. We
2248 // retain the Node to act as a compiler ordering barrier.
2249 bool Matcher::post_fast_unlock( const Node *rel ) {
2250 Compile *C = Compile::current();
2251 assert( rel->Opcode() == Op_MemBarRelease, "" );
2252 const MemBarReleaseNode *mem = (const MemBarReleaseNode*)rel;
2253 DUIterator_Fast imax, i = mem->fast_outs(imax);
2254 Node *ctrl = NULL;
2255 while( true ) {
2256 ctrl = mem->fast_out(i); // Throw out-of-bounds if proj not found
2257 assert( ctrl->is_Proj(), "only projections here" );
2258 ProjNode *proj = (ProjNode*)ctrl;
2259 if( proj->_con == TypeFunc::Control &&
2260 !C->node_arena()->contains(ctrl) ) // Unmatched old-space only
2261 break;
2262 i++;
2263 }
2264 Node *iff = NULL;
2265 for( DUIterator_Fast jmax, j = ctrl->fast_outs(jmax); j < jmax; j++ ) {
2266 Node *x = ctrl->fast_out(j);
2267 if( x->is_If() && x->req() > 1 &&
2268 !C->node_arena()->contains(x) ) { // Unmatched old-space only
2269 iff = x;
2270 break;
2271 }
2272 }
2273 if( !iff ) return false;
2274 Node *bol = iff->in(1);
2275 // The iff might be some random subclass of If or bol might be Con-Top
2276 if (!bol->is_Bool()) return false;
2277 assert( bol->req() > 1, "" );
2278 return (bol->in(1)->Opcode() == Op_FastUnlock);
2279 }
2281 // Used by the DFA in dfa_xxx.cpp. Check for a following barrier or
2282 // atomic instruction acting as a store_load barrier without any
2283 // intervening volatile load, and thus we don't need a barrier here.
2284 // We retain the Node to act as a compiler ordering barrier.
2285 bool Matcher::post_store_load_barrier(const Node *vmb) {
2286 Compile *C = Compile::current();
2287 assert( vmb->is_MemBar(), "" );
2288 assert( vmb->Opcode() != Op_MemBarAcquire, "" );
2289 const MemBarNode *mem = (const MemBarNode*)vmb;
2291 // Get the Proj node, ctrl, that can be used to iterate forward
2292 Node *ctrl = NULL;
2293 DUIterator_Fast imax, i = mem->fast_outs(imax);
2294 while( true ) {
2295 ctrl = mem->fast_out(i); // Throw out-of-bounds if proj not found
2296 assert( ctrl->is_Proj(), "only projections here" );
2297 ProjNode *proj = (ProjNode*)ctrl;
2298 if( proj->_con == TypeFunc::Control &&
2299 !C->node_arena()->contains(ctrl) ) // Unmatched old-space only
2300 break;
2301 i++;
2302 }
2304 for( DUIterator_Fast jmax, j = ctrl->fast_outs(jmax); j < jmax; j++ ) {
2305 Node *x = ctrl->fast_out(j);
2306 int xop = x->Opcode();
2308 // We don't need current barrier if we see another or a lock
2309 // before seeing volatile load.
2310 //
2311 // Op_Fastunlock previously appeared in the Op_* list below.
2312 // With the advent of 1-0 lock operations we're no longer guaranteed
2313 // that a monitor exit operation contains a serializing instruction.
2315 if (xop == Op_MemBarVolatile ||
2316 xop == Op_FastLock ||
2317 xop == Op_CompareAndSwapL ||
2318 xop == Op_CompareAndSwapP ||
2319 xop == Op_CompareAndSwapN ||
2320 xop == Op_CompareAndSwapI)
2321 return true;
2323 if (x->is_MemBar()) {
2324 // We must retain this membar if there is an upcoming volatile
2325 // load, which will be preceded by acquire membar.
2326 if (xop == Op_MemBarAcquire)
2327 return false;
2328 // For other kinds of barriers, check by pretending we
2329 // are them, and seeing if we can be removed.
2330 else
2331 return post_store_load_barrier((const MemBarNode*)x);
2332 }
2334 // Delicate code to detect case of an upcoming fastlock block
2335 if( x->is_If() && x->req() > 1 &&
2336 !C->node_arena()->contains(x) ) { // Unmatched old-space only
2337 Node *iff = x;
2338 Node *bol = iff->in(1);
2339 // The iff might be some random subclass of If or bol might be Con-Top
2340 if (!bol->is_Bool()) return false;
2341 assert( bol->req() > 1, "" );
2342 return (bol->in(1)->Opcode() == Op_FastUnlock);
2343 }
2344 // probably not necessary to check for these
2345 if (x->is_Call() || x->is_SafePoint() || x->is_block_proj())
2346 return false;
2347 }
2348 return false;
2349 }
2351 //=============================================================================
2352 //---------------------------State---------------------------------------------
2353 State::State(void) {
2354 #ifdef ASSERT
2355 _id = 0;
2356 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe);
2357 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d);
2358 //memset(_cost, -1, sizeof(_cost));
2359 //memset(_rule, -1, sizeof(_rule));
2360 #endif
2361 memset(_valid, 0, sizeof(_valid));
2362 }
2364 #ifdef ASSERT
2365 State::~State() {
2366 _id = 99;
2367 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe);
2368 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d);
2369 memset(_cost, -3, sizeof(_cost));
2370 memset(_rule, -3, sizeof(_rule));
2371 }
2372 #endif
2374 #ifndef PRODUCT
2375 //---------------------------dump----------------------------------------------
2376 void State::dump() {
2377 tty->print("\n");
2378 dump(0);
2379 }
2381 void State::dump(int depth) {
2382 for( int j = 0; j < depth; j++ )
2383 tty->print(" ");
2384 tty->print("--N: ");
2385 _leaf->dump();
2386 uint i;
2387 for( i = 0; i < _LAST_MACH_OPER; i++ )
2388 // Check for valid entry
2389 if( valid(i) ) {
2390 for( int j = 0; j < depth; j++ )
2391 tty->print(" ");
2392 assert(_cost[i] != max_juint, "cost must be a valid value");
2393 assert(_rule[i] < _last_Mach_Node, "rule[i] must be valid rule");
2394 tty->print_cr("%s %d %s",
2395 ruleName[i], _cost[i], ruleName[_rule[i]] );
2396 }
2397 tty->print_cr("");
2399 for( i=0; i<2; i++ )
2400 if( _kids[i] )
2401 _kids[i]->dump(depth+1);
2402 }
2403 #endif