Thu, 03 Jan 2013 15:09:55 -0800
8005522: use fast-string instructions on x86 for zeroing
Summary: use 'rep stosb' instead of 'rep stosq' when fast-string operations are available.
Reviewed-by: twisti, roland
1 /*
2 * Copyright (c) 1997, 2012, 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 "opto/vectornode.hpp"
39 #include "runtime/atomic.hpp"
40 #include "runtime/os.hpp"
41 #ifdef TARGET_ARCH_MODEL_x86_32
42 # include "adfiles/ad_x86_32.hpp"
43 #endif
44 #ifdef TARGET_ARCH_MODEL_x86_64
45 # include "adfiles/ad_x86_64.hpp"
46 #endif
47 #ifdef TARGET_ARCH_MODEL_sparc
48 # include "adfiles/ad_sparc.hpp"
49 #endif
50 #ifdef TARGET_ARCH_MODEL_zero
51 # include "adfiles/ad_zero.hpp"
52 #endif
53 #ifdef TARGET_ARCH_MODEL_arm
54 # include "adfiles/ad_arm.hpp"
55 #endif
56 #ifdef TARGET_ARCH_MODEL_ppc
57 # include "adfiles/ad_ppc.hpp"
58 #endif
60 OptoReg::Name OptoReg::c_frame_pointer;
62 const RegMask *Matcher::idealreg2regmask[_last_machine_leaf];
63 RegMask Matcher::mreg2regmask[_last_Mach_Reg];
64 RegMask Matcher::STACK_ONLY_mask;
65 RegMask Matcher::c_frame_ptr_mask;
66 const uint Matcher::_begin_rematerialize = _BEGIN_REMATERIALIZE;
67 const uint Matcher::_end_rematerialize = _END_REMATERIALIZE;
69 //---------------------------Matcher-------------------------------------------
70 Matcher::Matcher( Node_List &proj_list ) :
71 PhaseTransform( Phase::Ins_Select ),
72 #ifdef ASSERT
73 _old2new_map(C->comp_arena()),
74 _new2old_map(C->comp_arena()),
75 #endif
76 _shared_nodes(C->comp_arena()),
77 _reduceOp(reduceOp), _leftOp(leftOp), _rightOp(rightOp),
78 _swallowed(swallowed),
79 _begin_inst_chain_rule(_BEGIN_INST_CHAIN_RULE),
80 _end_inst_chain_rule(_END_INST_CHAIN_RULE),
81 _must_clone(must_clone), _proj_list(proj_list),
82 _register_save_policy(register_save_policy),
83 _c_reg_save_policy(c_reg_save_policy),
84 _register_save_type(register_save_type),
85 _ruleName(ruleName),
86 _allocation_started(false),
87 _states_arena(Chunk::medium_size),
88 _visited(&_states_arena),
89 _shared(&_states_arena),
90 _dontcare(&_states_arena) {
91 C->set_matcher(this);
93 idealreg2spillmask [Op_RegI] = NULL;
94 idealreg2spillmask [Op_RegN] = NULL;
95 idealreg2spillmask [Op_RegL] = NULL;
96 idealreg2spillmask [Op_RegF] = NULL;
97 idealreg2spillmask [Op_RegD] = NULL;
98 idealreg2spillmask [Op_RegP] = NULL;
99 idealreg2spillmask [Op_VecS] = NULL;
100 idealreg2spillmask [Op_VecD] = NULL;
101 idealreg2spillmask [Op_VecX] = NULL;
102 idealreg2spillmask [Op_VecY] = NULL;
104 idealreg2debugmask [Op_RegI] = NULL;
105 idealreg2debugmask [Op_RegN] = NULL;
106 idealreg2debugmask [Op_RegL] = NULL;
107 idealreg2debugmask [Op_RegF] = NULL;
108 idealreg2debugmask [Op_RegD] = NULL;
109 idealreg2debugmask [Op_RegP] = NULL;
110 idealreg2debugmask [Op_VecS] = NULL;
111 idealreg2debugmask [Op_VecD] = NULL;
112 idealreg2debugmask [Op_VecX] = NULL;
113 idealreg2debugmask [Op_VecY] = 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;
121 idealreg2mhdebugmask[Op_VecS] = NULL;
122 idealreg2mhdebugmask[Op_VecD] = NULL;
123 idealreg2mhdebugmask[Op_VecX] = NULL;
124 idealreg2mhdebugmask[Op_VecY] = NULL;
126 debug_only(_mem_node = NULL;) // Ideal memory node consumed by mach node
127 }
129 //------------------------------warp_incoming_stk_arg------------------------
130 // This warps a VMReg into an OptoReg::Name
131 OptoReg::Name Matcher::warp_incoming_stk_arg( VMReg reg ) {
132 OptoReg::Name warped;
133 if( reg->is_stack() ) { // Stack slot argument?
134 warped = OptoReg::add(_old_SP, reg->reg2stack() );
135 warped = OptoReg::add(warped, C->out_preserve_stack_slots());
136 if( warped >= _in_arg_limit )
137 _in_arg_limit = OptoReg::add(warped, 1); // Bump max stack slot seen
138 if (!RegMask::can_represent_arg(warped)) {
139 // the compiler cannot represent this method's calling sequence
140 C->record_method_not_compilable_all_tiers("unsupported incoming calling sequence");
141 return OptoReg::Bad;
142 }
143 return warped;
144 }
145 return OptoReg::as_OptoReg(reg);
146 }
148 //---------------------------compute_old_SP------------------------------------
149 OptoReg::Name Compile::compute_old_SP() {
150 int fixed = fixed_slots();
151 int preserve = in_preserve_stack_slots();
152 return OptoReg::stack2reg(round_to(fixed + preserve, Matcher::stack_alignment_in_slots()));
153 }
157 #ifdef ASSERT
158 void Matcher::verify_new_nodes_only(Node* xroot) {
159 // Make sure that the new graph only references new nodes
160 ResourceMark rm;
161 Unique_Node_List worklist;
162 VectorSet visited(Thread::current()->resource_area());
163 worklist.push(xroot);
164 while (worklist.size() > 0) {
165 Node* n = worklist.pop();
166 visited <<= n->_idx;
167 assert(C->node_arena()->contains(n), "dead node");
168 for (uint j = 0; j < n->req(); j++) {
169 Node* in = n->in(j);
170 if (in != NULL) {
171 assert(C->node_arena()->contains(in), "dead node");
172 if (!visited.test(in->_idx)) {
173 worklist.push(in);
174 }
175 }
176 }
177 }
178 }
179 #endif
182 //---------------------------match---------------------------------------------
183 void Matcher::match( ) {
184 if( MaxLabelRootDepth < 100 ) { // Too small?
185 assert(false, "invalid MaxLabelRootDepth, increase it to 100 minimum");
186 MaxLabelRootDepth = 100;
187 }
188 // One-time initialization of some register masks.
189 init_spill_mask( C->root()->in(1) );
190 _return_addr_mask = return_addr();
191 #ifdef _LP64
192 // Pointers take 2 slots in 64-bit land
193 _return_addr_mask.Insert(OptoReg::add(return_addr(),1));
194 #endif
196 // Map a Java-signature return type into return register-value
197 // machine registers for 0, 1 and 2 returned values.
198 const TypeTuple *range = C->tf()->range();
199 if( range->cnt() > TypeFunc::Parms ) { // If not a void function
200 // Get ideal-register return type
201 int ireg = range->field_at(TypeFunc::Parms)->ideal_reg();
202 // Get machine return register
203 uint sop = C->start()->Opcode();
204 OptoRegPair regs = return_value(ireg, false);
206 // And mask for same
207 _return_value_mask = RegMask(regs.first());
208 if( OptoReg::is_valid(regs.second()) )
209 _return_value_mask.Insert(regs.second());
210 }
212 // ---------------
213 // Frame Layout
215 // Need the method signature to determine the incoming argument types,
216 // because the types determine which registers the incoming arguments are
217 // in, and this affects the matched code.
218 const TypeTuple *domain = C->tf()->domain();
219 uint argcnt = domain->cnt() - TypeFunc::Parms;
220 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt );
221 VMRegPair *vm_parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt );
222 _parm_regs = NEW_RESOURCE_ARRAY( OptoRegPair, argcnt );
223 _calling_convention_mask = NEW_RESOURCE_ARRAY( RegMask, argcnt );
224 uint i;
225 for( i = 0; i<argcnt; i++ ) {
226 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type();
227 }
229 // Pass array of ideal registers and length to USER code (from the AD file)
230 // that will convert this to an array of register numbers.
231 const StartNode *start = C->start();
232 start->calling_convention( sig_bt, vm_parm_regs, argcnt );
233 #ifdef ASSERT
234 // Sanity check users' calling convention. Real handy while trying to
235 // get the initial port correct.
236 { for (uint i = 0; i<argcnt; i++) {
237 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) {
238 assert(domain->field_at(i+TypeFunc::Parms)==Type::HALF, "only allowed on halve" );
239 _parm_regs[i].set_bad();
240 continue;
241 }
242 VMReg parm_reg = vm_parm_regs[i].first();
243 assert(parm_reg->is_valid(), "invalid arg?");
244 if (parm_reg->is_reg()) {
245 OptoReg::Name opto_parm_reg = OptoReg::as_OptoReg(parm_reg);
246 assert(can_be_java_arg(opto_parm_reg) ||
247 C->stub_function() == CAST_FROM_FN_PTR(address, OptoRuntime::rethrow_C) ||
248 opto_parm_reg == inline_cache_reg(),
249 "parameters in register must be preserved by runtime stubs");
250 }
251 for (uint j = 0; j < i; j++) {
252 assert(parm_reg != vm_parm_regs[j].first(),
253 "calling conv. must produce distinct regs");
254 }
255 }
256 }
257 #endif
259 // Do some initial frame layout.
261 // Compute the old incoming SP (may be called FP) as
262 // OptoReg::stack0() + locks + in_preserve_stack_slots + pad2.
263 _old_SP = C->compute_old_SP();
264 assert( is_even(_old_SP), "must be even" );
266 // Compute highest incoming stack argument as
267 // _old_SP + out_preserve_stack_slots + incoming argument size.
268 _in_arg_limit = OptoReg::add(_old_SP, C->out_preserve_stack_slots());
269 assert( is_even(_in_arg_limit), "out_preserve must be even" );
270 for( i = 0; i < argcnt; i++ ) {
271 // Permit args to have no register
272 _calling_convention_mask[i].Clear();
273 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) {
274 continue;
275 }
276 // calling_convention returns stack arguments as a count of
277 // slots beyond OptoReg::stack0()/VMRegImpl::stack0. We need to convert this to
278 // the allocators point of view, taking into account all the
279 // preserve area, locks & pad2.
281 OptoReg::Name reg1 = warp_incoming_stk_arg(vm_parm_regs[i].first());
282 if( OptoReg::is_valid(reg1))
283 _calling_convention_mask[i].Insert(reg1);
285 OptoReg::Name reg2 = warp_incoming_stk_arg(vm_parm_regs[i].second());
286 if( OptoReg::is_valid(reg2))
287 _calling_convention_mask[i].Insert(reg2);
289 // Saved biased stack-slot register number
290 _parm_regs[i].set_pair(reg2, reg1);
291 }
293 // Finally, make sure the incoming arguments take up an even number of
294 // words, in case the arguments or locals need to contain doubleword stack
295 // slots. The rest of the system assumes that stack slot pairs (in
296 // particular, in the spill area) which look aligned will in fact be
297 // aligned relative to the stack pointer in the target machine. Double
298 // stack slots will always be allocated aligned.
299 _new_SP = OptoReg::Name(round_to(_in_arg_limit, RegMask::SlotsPerLong));
301 // Compute highest outgoing stack argument as
302 // _new_SP + out_preserve_stack_slots + max(outgoing argument size).
303 _out_arg_limit = OptoReg::add(_new_SP, C->out_preserve_stack_slots());
304 assert( is_even(_out_arg_limit), "out_preserve must be even" );
306 if (!RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1))) {
307 // the compiler cannot represent this method's calling sequence
308 C->record_method_not_compilable("must be able to represent all call arguments in reg mask");
309 }
311 if (C->failing()) return; // bailed out on incoming arg failure
313 // ---------------
314 // Collect roots of matcher trees. Every node for which
315 // _shared[_idx] is cleared is guaranteed to not be shared, and thus
316 // can be a valid interior of some tree.
317 find_shared( C->root() );
318 find_shared( C->top() );
320 C->print_method("Before Matching");
322 // Create new ideal node ConP #NULL even if it does exist in old space
323 // to avoid false sharing if the corresponding mach node is not used.
324 // The corresponding mach node is only used in rare cases for derived
325 // pointers.
326 Node* new_ideal_null = ConNode::make(C, TypePtr::NULL_PTR);
328 // Swap out to old-space; emptying new-space
329 Arena *old = C->node_arena()->move_contents(C->old_arena());
331 // Save debug and profile information for nodes in old space:
332 _old_node_note_array = C->node_note_array();
333 if (_old_node_note_array != NULL) {
334 C->set_node_note_array(new(C->comp_arena()) GrowableArray<Node_Notes*>
335 (C->comp_arena(), _old_node_note_array->length(),
336 0, NULL));
337 }
339 // Pre-size the new_node table to avoid the need for range checks.
340 grow_new_node_array(C->unique());
342 // Reset node counter so MachNodes start with _idx at 0
343 int nodes = C->unique(); // save value
344 C->set_unique(0);
345 C->reset_dead_node_list();
347 // Recursively match trees from old space into new space.
348 // Correct leaves of new-space Nodes; they point to old-space.
349 _visited.Clear(); // Clear visit bits for xform call
350 C->set_cached_top_node(xform( C->top(), nodes ));
351 if (!C->failing()) {
352 Node* xroot = xform( C->root(), 1 );
353 if (xroot == NULL) {
354 Matcher::soft_match_failure(); // recursive matching process failed
355 C->record_method_not_compilable("instruction match failed");
356 } else {
357 // During matching shared constants were attached to C->root()
358 // because xroot wasn't available yet, so transfer the uses to
359 // the xroot.
360 for( DUIterator_Fast jmax, j = C->root()->fast_outs(jmax); j < jmax; j++ ) {
361 Node* n = C->root()->fast_out(j);
362 if (C->node_arena()->contains(n)) {
363 assert(n->in(0) == C->root(), "should be control user");
364 n->set_req(0, xroot);
365 --j;
366 --jmax;
367 }
368 }
370 // Generate new mach node for ConP #NULL
371 assert(new_ideal_null != NULL, "sanity");
372 _mach_null = match_tree(new_ideal_null);
373 // Don't set control, it will confuse GCM since there are no uses.
374 // The control will be set when this node is used first time
375 // in find_base_for_derived().
376 assert(_mach_null != NULL, "");
378 C->set_root(xroot->is_Root() ? xroot->as_Root() : NULL);
380 #ifdef ASSERT
381 verify_new_nodes_only(xroot);
382 #endif
383 }
384 }
385 if (C->top() == NULL || C->root() == NULL) {
386 C->record_method_not_compilable("graph lost"); // %%% cannot happen?
387 }
388 if (C->failing()) {
389 // delete old;
390 old->destruct_contents();
391 return;
392 }
393 assert( C->top(), "" );
394 assert( C->root(), "" );
395 validate_null_checks();
397 // Now smoke old-space
398 NOT_DEBUG( old->destruct_contents() );
400 // ------------------------
401 // Set up save-on-entry registers
402 Fixup_Save_On_Entry( );
403 }
406 //------------------------------Fixup_Save_On_Entry----------------------------
407 // The stated purpose of this routine is to take care of save-on-entry
408 // registers. However, the overall goal of the Match phase is to convert into
409 // machine-specific instructions which have RegMasks to guide allocation.
410 // So what this procedure really does is put a valid RegMask on each input
411 // to the machine-specific variations of all Return, TailCall and Halt
412 // instructions. It also adds edgs to define the save-on-entry values (and of
413 // course gives them a mask).
415 static RegMask *init_input_masks( uint size, RegMask &ret_adr, RegMask &fp ) {
416 RegMask *rms = NEW_RESOURCE_ARRAY( RegMask, size );
417 // Do all the pre-defined register masks
418 rms[TypeFunc::Control ] = RegMask::Empty;
419 rms[TypeFunc::I_O ] = RegMask::Empty;
420 rms[TypeFunc::Memory ] = RegMask::Empty;
421 rms[TypeFunc::ReturnAdr] = ret_adr;
422 rms[TypeFunc::FramePtr ] = fp;
423 return rms;
424 }
426 //---------------------------init_first_stack_mask-----------------------------
427 // Create the initial stack mask used by values spilling to the stack.
428 // Disallow any debug info in outgoing argument areas by setting the
429 // initial mask accordingly.
430 void Matcher::init_first_stack_mask() {
432 // Allocate storage for spill masks as masks for the appropriate load type.
433 RegMask *rms = (RegMask*)C->comp_arena()->Amalloc_D(sizeof(RegMask) * (3*6+4));
435 idealreg2spillmask [Op_RegN] = &rms[0];
436 idealreg2spillmask [Op_RegI] = &rms[1];
437 idealreg2spillmask [Op_RegL] = &rms[2];
438 idealreg2spillmask [Op_RegF] = &rms[3];
439 idealreg2spillmask [Op_RegD] = &rms[4];
440 idealreg2spillmask [Op_RegP] = &rms[5];
442 idealreg2debugmask [Op_RegN] = &rms[6];
443 idealreg2debugmask [Op_RegI] = &rms[7];
444 idealreg2debugmask [Op_RegL] = &rms[8];
445 idealreg2debugmask [Op_RegF] = &rms[9];
446 idealreg2debugmask [Op_RegD] = &rms[10];
447 idealreg2debugmask [Op_RegP] = &rms[11];
449 idealreg2mhdebugmask[Op_RegN] = &rms[12];
450 idealreg2mhdebugmask[Op_RegI] = &rms[13];
451 idealreg2mhdebugmask[Op_RegL] = &rms[14];
452 idealreg2mhdebugmask[Op_RegF] = &rms[15];
453 idealreg2mhdebugmask[Op_RegD] = &rms[16];
454 idealreg2mhdebugmask[Op_RegP] = &rms[17];
456 idealreg2spillmask [Op_VecS] = &rms[18];
457 idealreg2spillmask [Op_VecD] = &rms[19];
458 idealreg2spillmask [Op_VecX] = &rms[20];
459 idealreg2spillmask [Op_VecY] = &rms[21];
461 OptoReg::Name i;
463 // At first, start with the empty mask
464 C->FIRST_STACK_mask().Clear();
466 // Add in the incoming argument area
467 OptoReg::Name init = OptoReg::add(_old_SP, C->out_preserve_stack_slots());
468 for (i = init; i < _in_arg_limit; i = OptoReg::add(i,1))
469 C->FIRST_STACK_mask().Insert(i);
471 // Add in all bits past the outgoing argument area
472 guarantee(RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1)),
473 "must be able to represent all call arguments in reg mask");
474 init = _out_arg_limit;
475 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1))
476 C->FIRST_STACK_mask().Insert(i);
478 // Finally, set the "infinite stack" bit.
479 C->FIRST_STACK_mask().set_AllStack();
481 // Make spill masks. Registers for their class, plus FIRST_STACK_mask.
482 RegMask aligned_stack_mask = C->FIRST_STACK_mask();
483 // Keep spill masks aligned.
484 aligned_stack_mask.clear_to_pairs();
485 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack");
487 *idealreg2spillmask[Op_RegP] = *idealreg2regmask[Op_RegP];
488 #ifdef _LP64
489 *idealreg2spillmask[Op_RegN] = *idealreg2regmask[Op_RegN];
490 idealreg2spillmask[Op_RegN]->OR(C->FIRST_STACK_mask());
491 idealreg2spillmask[Op_RegP]->OR(aligned_stack_mask);
492 #else
493 idealreg2spillmask[Op_RegP]->OR(C->FIRST_STACK_mask());
494 #endif
495 *idealreg2spillmask[Op_RegI] = *idealreg2regmask[Op_RegI];
496 idealreg2spillmask[Op_RegI]->OR(C->FIRST_STACK_mask());
497 *idealreg2spillmask[Op_RegL] = *idealreg2regmask[Op_RegL];
498 idealreg2spillmask[Op_RegL]->OR(aligned_stack_mask);
499 *idealreg2spillmask[Op_RegF] = *idealreg2regmask[Op_RegF];
500 idealreg2spillmask[Op_RegF]->OR(C->FIRST_STACK_mask());
501 *idealreg2spillmask[Op_RegD] = *idealreg2regmask[Op_RegD];
502 idealreg2spillmask[Op_RegD]->OR(aligned_stack_mask);
504 if (Matcher::vector_size_supported(T_BYTE,4)) {
505 *idealreg2spillmask[Op_VecS] = *idealreg2regmask[Op_VecS];
506 idealreg2spillmask[Op_VecS]->OR(C->FIRST_STACK_mask());
507 }
508 if (Matcher::vector_size_supported(T_FLOAT,2)) {
509 *idealreg2spillmask[Op_VecD] = *idealreg2regmask[Op_VecD];
510 idealreg2spillmask[Op_VecD]->OR(aligned_stack_mask);
511 }
512 if (Matcher::vector_size_supported(T_FLOAT,4)) {
513 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecX);
514 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack");
515 *idealreg2spillmask[Op_VecX] = *idealreg2regmask[Op_VecX];
516 idealreg2spillmask[Op_VecX]->OR(aligned_stack_mask);
517 }
518 if (Matcher::vector_size_supported(T_FLOAT,8)) {
519 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecY);
520 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack");
521 *idealreg2spillmask[Op_VecY] = *idealreg2regmask[Op_VecY];
522 idealreg2spillmask[Op_VecY]->OR(aligned_stack_mask);
523 }
524 if (UseFPUForSpilling) {
525 // This mask logic assumes that the spill operations are
526 // symmetric and that the registers involved are the same size.
527 // On sparc for instance we may have to use 64 bit moves will
528 // kill 2 registers when used with F0-F31.
529 idealreg2spillmask[Op_RegI]->OR(*idealreg2regmask[Op_RegF]);
530 idealreg2spillmask[Op_RegF]->OR(*idealreg2regmask[Op_RegI]);
531 #ifdef _LP64
532 idealreg2spillmask[Op_RegN]->OR(*idealreg2regmask[Op_RegF]);
533 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]);
534 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]);
535 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegD]);
536 #else
537 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegF]);
538 #ifdef ARM
539 // ARM has support for moving 64bit values between a pair of
540 // integer registers and a double register
541 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]);
542 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]);
543 #endif
544 #endif
545 }
547 // Make up debug masks. Any spill slot plus callee-save registers.
548 // Caller-save registers are assumed to be trashable by the various
549 // inline-cache fixup routines.
550 *idealreg2debugmask [Op_RegN]= *idealreg2spillmask[Op_RegN];
551 *idealreg2debugmask [Op_RegI]= *idealreg2spillmask[Op_RegI];
552 *idealreg2debugmask [Op_RegL]= *idealreg2spillmask[Op_RegL];
553 *idealreg2debugmask [Op_RegF]= *idealreg2spillmask[Op_RegF];
554 *idealreg2debugmask [Op_RegD]= *idealreg2spillmask[Op_RegD];
555 *idealreg2debugmask [Op_RegP]= *idealreg2spillmask[Op_RegP];
557 *idealreg2mhdebugmask[Op_RegN]= *idealreg2spillmask[Op_RegN];
558 *idealreg2mhdebugmask[Op_RegI]= *idealreg2spillmask[Op_RegI];
559 *idealreg2mhdebugmask[Op_RegL]= *idealreg2spillmask[Op_RegL];
560 *idealreg2mhdebugmask[Op_RegF]= *idealreg2spillmask[Op_RegF];
561 *idealreg2mhdebugmask[Op_RegD]= *idealreg2spillmask[Op_RegD];
562 *idealreg2mhdebugmask[Op_RegP]= *idealreg2spillmask[Op_RegP];
564 // Prevent stub compilations from attempting to reference
565 // callee-saved registers from debug info
566 bool exclude_soe = !Compile::current()->is_method_compilation();
568 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) {
569 // registers the caller has to save do not work
570 if( _register_save_policy[i] == 'C' ||
571 _register_save_policy[i] == 'A' ||
572 (_register_save_policy[i] == 'E' && exclude_soe) ) {
573 idealreg2debugmask [Op_RegN]->Remove(i);
574 idealreg2debugmask [Op_RegI]->Remove(i); // Exclude save-on-call
575 idealreg2debugmask [Op_RegL]->Remove(i); // registers from debug
576 idealreg2debugmask [Op_RegF]->Remove(i); // masks
577 idealreg2debugmask [Op_RegD]->Remove(i);
578 idealreg2debugmask [Op_RegP]->Remove(i);
580 idealreg2mhdebugmask[Op_RegN]->Remove(i);
581 idealreg2mhdebugmask[Op_RegI]->Remove(i);
582 idealreg2mhdebugmask[Op_RegL]->Remove(i);
583 idealreg2mhdebugmask[Op_RegF]->Remove(i);
584 idealreg2mhdebugmask[Op_RegD]->Remove(i);
585 idealreg2mhdebugmask[Op_RegP]->Remove(i);
586 }
587 }
589 // Subtract the register we use to save the SP for MethodHandle
590 // invokes to from the debug mask.
591 const RegMask save_mask = method_handle_invoke_SP_save_mask();
592 idealreg2mhdebugmask[Op_RegN]->SUBTRACT(save_mask);
593 idealreg2mhdebugmask[Op_RegI]->SUBTRACT(save_mask);
594 idealreg2mhdebugmask[Op_RegL]->SUBTRACT(save_mask);
595 idealreg2mhdebugmask[Op_RegF]->SUBTRACT(save_mask);
596 idealreg2mhdebugmask[Op_RegD]->SUBTRACT(save_mask);
597 idealreg2mhdebugmask[Op_RegP]->SUBTRACT(save_mask);
598 }
600 //---------------------------is_save_on_entry----------------------------------
601 bool Matcher::is_save_on_entry( int reg ) {
602 return
603 _register_save_policy[reg] == 'E' ||
604 _register_save_policy[reg] == 'A' || // Save-on-entry register?
605 // Also save argument registers in the trampolining stubs
606 (C->save_argument_registers() && is_spillable_arg(reg));
607 }
609 //---------------------------Fixup_Save_On_Entry-------------------------------
610 void Matcher::Fixup_Save_On_Entry( ) {
611 init_first_stack_mask();
613 Node *root = C->root(); // Short name for root
614 // Count number of save-on-entry registers.
615 uint soe_cnt = number_of_saved_registers();
616 uint i;
618 // Find the procedure Start Node
619 StartNode *start = C->start();
620 assert( start, "Expect a start node" );
622 // Save argument registers in the trampolining stubs
623 if( C->save_argument_registers() )
624 for( i = 0; i < _last_Mach_Reg; i++ )
625 if( is_spillable_arg(i) )
626 soe_cnt++;
628 // Input RegMask array shared by all Returns.
629 // The type for doubles and longs has a count of 2, but
630 // there is only 1 returned value
631 uint ret_edge_cnt = TypeFunc::Parms + ((C->tf()->range()->cnt() == TypeFunc::Parms) ? 0 : 1);
632 RegMask *ret_rms = init_input_masks( ret_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
633 // Returns have 0 or 1 returned values depending on call signature.
634 // Return register is specified by return_value in the AD file.
635 if (ret_edge_cnt > TypeFunc::Parms)
636 ret_rms[TypeFunc::Parms+0] = _return_value_mask;
638 // Input RegMask array shared by all Rethrows.
639 uint reth_edge_cnt = TypeFunc::Parms+1;
640 RegMask *reth_rms = init_input_masks( reth_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
641 // Rethrow takes exception oop only, but in the argument 0 slot.
642 reth_rms[TypeFunc::Parms] = mreg2regmask[find_receiver(false)];
643 #ifdef _LP64
644 // Need two slots for ptrs in 64-bit land
645 reth_rms[TypeFunc::Parms].Insert(OptoReg::add(OptoReg::Name(find_receiver(false)),1));
646 #endif
648 // Input RegMask array shared by all TailCalls
649 uint tail_call_edge_cnt = TypeFunc::Parms+2;
650 RegMask *tail_call_rms = init_input_masks( tail_call_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
652 // Input RegMask array shared by all TailJumps
653 uint tail_jump_edge_cnt = TypeFunc::Parms+2;
654 RegMask *tail_jump_rms = init_input_masks( tail_jump_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
656 // TailCalls have 2 returned values (target & moop), whose masks come
657 // from the usual MachNode/MachOper mechanism. Find a sample
658 // TailCall to extract these masks and put the correct masks into
659 // the tail_call_rms array.
660 for( i=1; i < root->req(); i++ ) {
661 MachReturnNode *m = root->in(i)->as_MachReturn();
662 if( m->ideal_Opcode() == Op_TailCall ) {
663 tail_call_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0);
664 tail_call_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1);
665 break;
666 }
667 }
669 // TailJumps have 2 returned values (target & ex_oop), whose masks come
670 // from the usual MachNode/MachOper mechanism. Find a sample
671 // TailJump to extract these masks and put the correct masks into
672 // the tail_jump_rms array.
673 for( i=1; i < root->req(); i++ ) {
674 MachReturnNode *m = root->in(i)->as_MachReturn();
675 if( m->ideal_Opcode() == Op_TailJump ) {
676 tail_jump_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0);
677 tail_jump_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1);
678 break;
679 }
680 }
682 // Input RegMask array shared by all Halts
683 uint halt_edge_cnt = TypeFunc::Parms;
684 RegMask *halt_rms = init_input_masks( halt_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
686 // Capture the return input masks into each exit flavor
687 for( i=1; i < root->req(); i++ ) {
688 MachReturnNode *exit = root->in(i)->as_MachReturn();
689 switch( exit->ideal_Opcode() ) {
690 case Op_Return : exit->_in_rms = ret_rms; break;
691 case Op_Rethrow : exit->_in_rms = reth_rms; break;
692 case Op_TailCall : exit->_in_rms = tail_call_rms; break;
693 case Op_TailJump : exit->_in_rms = tail_jump_rms; break;
694 case Op_Halt : exit->_in_rms = halt_rms; break;
695 default : ShouldNotReachHere();
696 }
697 }
699 // Next unused projection number from Start.
700 int proj_cnt = C->tf()->domain()->cnt();
702 // Do all the save-on-entry registers. Make projections from Start for
703 // them, and give them a use at the exit points. To the allocator, they
704 // look like incoming register arguments.
705 for( i = 0; i < _last_Mach_Reg; i++ ) {
706 if( is_save_on_entry(i) ) {
708 // Add the save-on-entry to the mask array
709 ret_rms [ ret_edge_cnt] = mreg2regmask[i];
710 reth_rms [ reth_edge_cnt] = mreg2regmask[i];
711 tail_call_rms[tail_call_edge_cnt] = mreg2regmask[i];
712 tail_jump_rms[tail_jump_edge_cnt] = mreg2regmask[i];
713 // Halts need the SOE registers, but only in the stack as debug info.
714 // A just-prior uncommon-trap or deoptimization will use the SOE regs.
715 halt_rms [ halt_edge_cnt] = *idealreg2spillmask[_register_save_type[i]];
717 Node *mproj;
719 // Is this a RegF low half of a RegD? Double up 2 adjacent RegF's
720 // into a single RegD.
721 if( (i&1) == 0 &&
722 _register_save_type[i ] == Op_RegF &&
723 _register_save_type[i+1] == Op_RegF &&
724 is_save_on_entry(i+1) ) {
725 // Add other bit for double
726 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1));
727 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1));
728 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1));
729 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1));
730 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1));
731 mproj = new (C) MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegD );
732 proj_cnt += 2; // Skip 2 for doubles
733 }
734 else if( (i&1) == 1 && // Else check for high half of double
735 _register_save_type[i-1] == Op_RegF &&
736 _register_save_type[i ] == Op_RegF &&
737 is_save_on_entry(i-1) ) {
738 ret_rms [ ret_edge_cnt] = RegMask::Empty;
739 reth_rms [ reth_edge_cnt] = RegMask::Empty;
740 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty;
741 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty;
742 halt_rms [ halt_edge_cnt] = RegMask::Empty;
743 mproj = C->top();
744 }
745 // Is this a RegI low half of a RegL? Double up 2 adjacent RegI's
746 // into a single RegL.
747 else if( (i&1) == 0 &&
748 _register_save_type[i ] == Op_RegI &&
749 _register_save_type[i+1] == Op_RegI &&
750 is_save_on_entry(i+1) ) {
751 // Add other bit for long
752 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1));
753 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1));
754 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1));
755 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1));
756 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1));
757 mproj = new (C) MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegL );
758 proj_cnt += 2; // Skip 2 for longs
759 }
760 else if( (i&1) == 1 && // Else check for high half of long
761 _register_save_type[i-1] == Op_RegI &&
762 _register_save_type[i ] == Op_RegI &&
763 is_save_on_entry(i-1) ) {
764 ret_rms [ ret_edge_cnt] = RegMask::Empty;
765 reth_rms [ reth_edge_cnt] = RegMask::Empty;
766 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty;
767 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty;
768 halt_rms [ halt_edge_cnt] = RegMask::Empty;
769 mproj = C->top();
770 } else {
771 // Make a projection for it off the Start
772 mproj = new (C) MachProjNode( start, proj_cnt++, ret_rms[ret_edge_cnt], _register_save_type[i] );
773 }
775 ret_edge_cnt ++;
776 reth_edge_cnt ++;
777 tail_call_edge_cnt ++;
778 tail_jump_edge_cnt ++;
779 halt_edge_cnt ++;
781 // Add a use of the SOE register to all exit paths
782 for( uint j=1; j < root->req(); j++ )
783 root->in(j)->add_req(mproj);
784 } // End of if a save-on-entry register
785 } // End of for all machine registers
786 }
788 //------------------------------init_spill_mask--------------------------------
789 void Matcher::init_spill_mask( Node *ret ) {
790 if( idealreg2regmask[Op_RegI] ) return; // One time only init
792 OptoReg::c_frame_pointer = c_frame_pointer();
793 c_frame_ptr_mask = c_frame_pointer();
794 #ifdef _LP64
795 // pointers are twice as big
796 c_frame_ptr_mask.Insert(OptoReg::add(c_frame_pointer(),1));
797 #endif
799 // Start at OptoReg::stack0()
800 STACK_ONLY_mask.Clear();
801 OptoReg::Name init = OptoReg::stack2reg(0);
802 // STACK_ONLY_mask is all stack bits
803 OptoReg::Name i;
804 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1))
805 STACK_ONLY_mask.Insert(i);
806 // Also set the "infinite stack" bit.
807 STACK_ONLY_mask.set_AllStack();
809 // Copy the register names over into the shared world
810 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) {
811 // SharedInfo::regName[i] = regName[i];
812 // Handy RegMasks per machine register
813 mreg2regmask[i].Insert(i);
814 }
816 // Grab the Frame Pointer
817 Node *fp = ret->in(TypeFunc::FramePtr);
818 Node *mem = ret->in(TypeFunc::Memory);
819 const TypePtr* atp = TypePtr::BOTTOM;
820 // Share frame pointer while making spill ops
821 set_shared(fp);
823 // Compute generic short-offset Loads
824 #ifdef _LP64
825 MachNode *spillCP = match_tree(new (C) LoadNNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM));
826 #endif
827 MachNode *spillI = match_tree(new (C) LoadINode(NULL,mem,fp,atp));
828 MachNode *spillL = match_tree(new (C) LoadLNode(NULL,mem,fp,atp));
829 MachNode *spillF = match_tree(new (C) LoadFNode(NULL,mem,fp,atp));
830 MachNode *spillD = match_tree(new (C) LoadDNode(NULL,mem,fp,atp));
831 MachNode *spillP = match_tree(new (C) LoadPNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM));
832 assert(spillI != NULL && spillL != NULL && spillF != NULL &&
833 spillD != NULL && spillP != NULL, "");
835 // Get the ADLC notion of the right regmask, for each basic type.
836 #ifdef _LP64
837 idealreg2regmask[Op_RegN] = &spillCP->out_RegMask();
838 #endif
839 idealreg2regmask[Op_RegI] = &spillI->out_RegMask();
840 idealreg2regmask[Op_RegL] = &spillL->out_RegMask();
841 idealreg2regmask[Op_RegF] = &spillF->out_RegMask();
842 idealreg2regmask[Op_RegD] = &spillD->out_RegMask();
843 idealreg2regmask[Op_RegP] = &spillP->out_RegMask();
845 // Vector regmasks.
846 if (Matcher::vector_size_supported(T_BYTE,4)) {
847 TypeVect::VECTS = TypeVect::make(T_BYTE, 4);
848 MachNode *spillVectS = match_tree(new (C) LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTS));
849 idealreg2regmask[Op_VecS] = &spillVectS->out_RegMask();
850 }
851 if (Matcher::vector_size_supported(T_FLOAT,2)) {
852 MachNode *spillVectD = match_tree(new (C) LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTD));
853 idealreg2regmask[Op_VecD] = &spillVectD->out_RegMask();
854 }
855 if (Matcher::vector_size_supported(T_FLOAT,4)) {
856 MachNode *spillVectX = match_tree(new (C) LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTX));
857 idealreg2regmask[Op_VecX] = &spillVectX->out_RegMask();
858 }
859 if (Matcher::vector_size_supported(T_FLOAT,8)) {
860 MachNode *spillVectY = match_tree(new (C) LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTY));
861 idealreg2regmask[Op_VecY] = &spillVectY->out_RegMask();
862 }
863 }
865 #ifdef ASSERT
866 static void match_alias_type(Compile* C, Node* n, Node* m) {
867 if (!VerifyAliases) return; // do not go looking for trouble by default
868 const TypePtr* nat = n->adr_type();
869 const TypePtr* mat = m->adr_type();
870 int nidx = C->get_alias_index(nat);
871 int midx = C->get_alias_index(mat);
872 // Detune the assert for cases like (AndI 0xFF (LoadB p)).
873 if (nidx == Compile::AliasIdxTop && midx >= Compile::AliasIdxRaw) {
874 for (uint i = 1; i < n->req(); i++) {
875 Node* n1 = n->in(i);
876 const TypePtr* n1at = n1->adr_type();
877 if (n1at != NULL) {
878 nat = n1at;
879 nidx = C->get_alias_index(n1at);
880 }
881 }
882 }
883 // %%% Kludgery. Instead, fix ideal adr_type methods for all these cases:
884 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxRaw) {
885 switch (n->Opcode()) {
886 case Op_PrefetchRead:
887 case Op_PrefetchWrite:
888 case Op_PrefetchAllocation:
889 nidx = Compile::AliasIdxRaw;
890 nat = TypeRawPtr::BOTTOM;
891 break;
892 }
893 }
894 if (nidx == Compile::AliasIdxRaw && midx == Compile::AliasIdxTop) {
895 switch (n->Opcode()) {
896 case Op_ClearArray:
897 midx = Compile::AliasIdxRaw;
898 mat = TypeRawPtr::BOTTOM;
899 break;
900 }
901 }
902 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxBot) {
903 switch (n->Opcode()) {
904 case Op_Return:
905 case Op_Rethrow:
906 case Op_Halt:
907 case Op_TailCall:
908 case Op_TailJump:
909 nidx = Compile::AliasIdxBot;
910 nat = TypePtr::BOTTOM;
911 break;
912 }
913 }
914 if (nidx == Compile::AliasIdxBot && midx == Compile::AliasIdxTop) {
915 switch (n->Opcode()) {
916 case Op_StrComp:
917 case Op_StrEquals:
918 case Op_StrIndexOf:
919 case Op_AryEq:
920 case Op_MemBarVolatile:
921 case Op_MemBarCPUOrder: // %%% these ideals should have narrower adr_type?
922 nidx = Compile::AliasIdxTop;
923 nat = NULL;
924 break;
925 }
926 }
927 if (nidx != midx) {
928 if (PrintOpto || (PrintMiscellaneous && (WizardMode || Verbose))) {
929 tty->print_cr("==== Matcher alias shift %d => %d", nidx, midx);
930 n->dump();
931 m->dump();
932 }
933 assert(C->subsume_loads() && C->must_alias(nat, midx),
934 "must not lose alias info when matching");
935 }
936 }
937 #endif
940 //------------------------------MStack-----------------------------------------
941 // State and MStack class used in xform() and find_shared() iterative methods.
942 enum Node_State { Pre_Visit, // node has to be pre-visited
943 Visit, // visit node
944 Post_Visit, // post-visit node
945 Alt_Post_Visit // alternative post-visit path
946 };
948 class MStack: public Node_Stack {
949 public:
950 MStack(int size) : Node_Stack(size) { }
952 void push(Node *n, Node_State ns) {
953 Node_Stack::push(n, (uint)ns);
954 }
955 void push(Node *n, Node_State ns, Node *parent, int indx) {
956 ++_inode_top;
957 if ((_inode_top + 1) >= _inode_max) grow();
958 _inode_top->node = parent;
959 _inode_top->indx = (uint)indx;
960 ++_inode_top;
961 _inode_top->node = n;
962 _inode_top->indx = (uint)ns;
963 }
964 Node *parent() {
965 pop();
966 return node();
967 }
968 Node_State state() const {
969 return (Node_State)index();
970 }
971 void set_state(Node_State ns) {
972 set_index((uint)ns);
973 }
974 };
977 //------------------------------xform------------------------------------------
978 // Given a Node in old-space, Match him (Label/Reduce) to produce a machine
979 // Node in new-space. Given a new-space Node, recursively walk his children.
980 Node *Matcher::transform( Node *n ) { ShouldNotCallThis(); return n; }
981 Node *Matcher::xform( Node *n, int max_stack ) {
982 // Use one stack to keep both: child's node/state and parent's node/index
983 MStack mstack(max_stack * 2 * 2); // C->unique() * 2 * 2
984 mstack.push(n, Visit, NULL, -1); // set NULL as parent to indicate root
986 while (mstack.is_nonempty()) {
987 n = mstack.node(); // Leave node on stack
988 Node_State nstate = mstack.state();
989 if (nstate == Visit) {
990 mstack.set_state(Post_Visit);
991 Node *oldn = n;
992 // Old-space or new-space check
993 if (!C->node_arena()->contains(n)) {
994 // Old space!
995 Node* m;
996 if (has_new_node(n)) { // Not yet Label/Reduced
997 m = new_node(n);
998 } else {
999 if (!is_dontcare(n)) { // Matcher can match this guy
1000 // Calls match special. They match alone with no children.
1001 // Their children, the incoming arguments, match normally.
1002 m = n->is_SafePoint() ? match_sfpt(n->as_SafePoint()):match_tree(n);
1003 if (C->failing()) return NULL;
1004 if (m == NULL) { Matcher::soft_match_failure(); return NULL; }
1005 } else { // Nothing the matcher cares about
1006 if( n->is_Proj() && n->in(0)->is_Multi()) { // Projections?
1007 // Convert to machine-dependent projection
1008 m = n->in(0)->as_Multi()->match( n->as_Proj(), this );
1009 #ifdef ASSERT
1010 _new2old_map.map(m->_idx, n);
1011 #endif
1012 if (m->in(0) != NULL) // m might be top
1013 collect_null_checks(m, n);
1014 } else { // Else just a regular 'ol guy
1015 m = n->clone(); // So just clone into new-space
1016 #ifdef ASSERT
1017 _new2old_map.map(m->_idx, n);
1018 #endif
1019 // Def-Use edges will be added incrementally as Uses
1020 // of this node are matched.
1021 assert(m->outcnt() == 0, "no Uses of this clone yet");
1022 }
1023 }
1025 set_new_node(n, m); // Map old to new
1026 if (_old_node_note_array != NULL) {
1027 Node_Notes* nn = C->locate_node_notes(_old_node_note_array,
1028 n->_idx);
1029 C->set_node_notes_at(m->_idx, nn);
1030 }
1031 debug_only(match_alias_type(C, n, m));
1032 }
1033 n = m; // n is now a new-space node
1034 mstack.set_node(n);
1035 }
1037 // New space!
1038 if (_visited.test_set(n->_idx)) continue; // while(mstack.is_nonempty())
1040 int i;
1041 // Put precedence edges on stack first (match them last).
1042 for (i = oldn->req(); (uint)i < oldn->len(); i++) {
1043 Node *m = oldn->in(i);
1044 if (m == NULL) break;
1045 // set -1 to call add_prec() instead of set_req() during Step1
1046 mstack.push(m, Visit, n, -1);
1047 }
1049 // For constant debug info, I'd rather have unmatched constants.
1050 int cnt = n->req();
1051 JVMState* jvms = n->jvms();
1052 int debug_cnt = jvms ? jvms->debug_start() : cnt;
1054 // Now do only debug info. Clone constants rather than matching.
1055 // Constants are represented directly in the debug info without
1056 // the need for executable machine instructions.
1057 // Monitor boxes are also represented directly.
1058 for (i = cnt - 1; i >= debug_cnt; --i) { // For all debug inputs do
1059 Node *m = n->in(i); // Get input
1060 int op = m->Opcode();
1061 assert((op == Op_BoxLock) == jvms->is_monitor_use(i), "boxes only at monitor sites");
1062 if( op == Op_ConI || op == Op_ConP || op == Op_ConN || op == Op_ConNKlass ||
1063 op == Op_ConF || op == Op_ConD || op == Op_ConL
1064 // || op == Op_BoxLock // %%%% enable this and remove (+++) in chaitin.cpp
1065 ) {
1066 m = m->clone();
1067 #ifdef ASSERT
1068 _new2old_map.map(m->_idx, n);
1069 #endif
1070 mstack.push(m, Post_Visit, n, i); // Don't need to visit
1071 mstack.push(m->in(0), Visit, m, 0);
1072 } else {
1073 mstack.push(m, Visit, n, i);
1074 }
1075 }
1077 // And now walk his children, and convert his inputs to new-space.
1078 for( ; i >= 0; --i ) { // For all normal inputs do
1079 Node *m = n->in(i); // Get input
1080 if(m != NULL)
1081 mstack.push(m, Visit, n, i);
1082 }
1084 }
1085 else if (nstate == Post_Visit) {
1086 // Set xformed input
1087 Node *p = mstack.parent();
1088 if (p != NULL) { // root doesn't have parent
1089 int i = (int)mstack.index();
1090 if (i >= 0)
1091 p->set_req(i, n); // required input
1092 else if (i == -1)
1093 p->add_prec(n); // precedence input
1094 else
1095 ShouldNotReachHere();
1096 }
1097 mstack.pop(); // remove processed node from stack
1098 }
1099 else {
1100 ShouldNotReachHere();
1101 }
1102 } // while (mstack.is_nonempty())
1103 return n; // Return new-space Node
1104 }
1106 //------------------------------warp_outgoing_stk_arg------------------------
1107 OptoReg::Name Matcher::warp_outgoing_stk_arg( VMReg reg, OptoReg::Name begin_out_arg_area, OptoReg::Name &out_arg_limit_per_call ) {
1108 // Convert outgoing argument location to a pre-biased stack offset
1109 if (reg->is_stack()) {
1110 OptoReg::Name warped = reg->reg2stack();
1111 // Adjust the stack slot offset to be the register number used
1112 // by the allocator.
1113 warped = OptoReg::add(begin_out_arg_area, warped);
1114 // Keep track of the largest numbered stack slot used for an arg.
1115 // Largest used slot per call-site indicates the amount of stack
1116 // that is killed by the call.
1117 if( warped >= out_arg_limit_per_call )
1118 out_arg_limit_per_call = OptoReg::add(warped,1);
1119 if (!RegMask::can_represent_arg(warped)) {
1120 C->record_method_not_compilable_all_tiers("unsupported calling sequence");
1121 return OptoReg::Bad;
1122 }
1123 return warped;
1124 }
1125 return OptoReg::as_OptoReg(reg);
1126 }
1129 //------------------------------match_sfpt-------------------------------------
1130 // Helper function to match call instructions. Calls match special.
1131 // They match alone with no children. Their children, the incoming
1132 // arguments, match normally.
1133 MachNode *Matcher::match_sfpt( SafePointNode *sfpt ) {
1134 MachSafePointNode *msfpt = NULL;
1135 MachCallNode *mcall = NULL;
1136 uint cnt;
1137 // Split out case for SafePoint vs Call
1138 CallNode *call;
1139 const TypeTuple *domain;
1140 ciMethod* method = NULL;
1141 bool is_method_handle_invoke = false; // for special kill effects
1142 if( sfpt->is_Call() ) {
1143 call = sfpt->as_Call();
1144 domain = call->tf()->domain();
1145 cnt = domain->cnt();
1147 // Match just the call, nothing else
1148 MachNode *m = match_tree(call);
1149 if (C->failing()) return NULL;
1150 if( m == NULL ) { Matcher::soft_match_failure(); return NULL; }
1152 // Copy data from the Ideal SafePoint to the machine version
1153 mcall = m->as_MachCall();
1155 mcall->set_tf( call->tf());
1156 mcall->set_entry_point(call->entry_point());
1157 mcall->set_cnt( call->cnt());
1159 if( mcall->is_MachCallJava() ) {
1160 MachCallJavaNode *mcall_java = mcall->as_MachCallJava();
1161 const CallJavaNode *call_java = call->as_CallJava();
1162 method = call_java->method();
1163 mcall_java->_method = method;
1164 mcall_java->_bci = call_java->_bci;
1165 mcall_java->_optimized_virtual = call_java->is_optimized_virtual();
1166 is_method_handle_invoke = call_java->is_method_handle_invoke();
1167 mcall_java->_method_handle_invoke = is_method_handle_invoke;
1168 if (is_method_handle_invoke) {
1169 C->set_has_method_handle_invokes(true);
1170 }
1171 if( mcall_java->is_MachCallStaticJava() )
1172 mcall_java->as_MachCallStaticJava()->_name =
1173 call_java->as_CallStaticJava()->_name;
1174 if( mcall_java->is_MachCallDynamicJava() )
1175 mcall_java->as_MachCallDynamicJava()->_vtable_index =
1176 call_java->as_CallDynamicJava()->_vtable_index;
1177 }
1178 else if( mcall->is_MachCallRuntime() ) {
1179 mcall->as_MachCallRuntime()->_name = call->as_CallRuntime()->_name;
1180 }
1181 msfpt = mcall;
1182 }
1183 // This is a non-call safepoint
1184 else {
1185 call = NULL;
1186 domain = NULL;
1187 MachNode *mn = match_tree(sfpt);
1188 if (C->failing()) return NULL;
1189 msfpt = mn->as_MachSafePoint();
1190 cnt = TypeFunc::Parms;
1191 }
1193 // Advertise the correct memory effects (for anti-dependence computation).
1194 msfpt->set_adr_type(sfpt->adr_type());
1196 // Allocate a private array of RegMasks. These RegMasks are not shared.
1197 msfpt->_in_rms = NEW_RESOURCE_ARRAY( RegMask, cnt );
1198 // Empty them all.
1199 memset( msfpt->_in_rms, 0, sizeof(RegMask)*cnt );
1201 // Do all the pre-defined non-Empty register masks
1202 msfpt->_in_rms[TypeFunc::ReturnAdr] = _return_addr_mask;
1203 msfpt->_in_rms[TypeFunc::FramePtr ] = c_frame_ptr_mask;
1205 // Place first outgoing argument can possibly be put.
1206 OptoReg::Name begin_out_arg_area = OptoReg::add(_new_SP, C->out_preserve_stack_slots());
1207 assert( is_even(begin_out_arg_area), "" );
1208 // Compute max outgoing register number per call site.
1209 OptoReg::Name out_arg_limit_per_call = begin_out_arg_area;
1210 // Calls to C may hammer extra stack slots above and beyond any arguments.
1211 // These are usually backing store for register arguments for varargs.
1212 if( call != NULL && call->is_CallRuntime() )
1213 out_arg_limit_per_call = OptoReg::add(out_arg_limit_per_call,C->varargs_C_out_slots_killed());
1216 // Do the normal argument list (parameters) register masks
1217 int argcnt = cnt - TypeFunc::Parms;
1218 if( argcnt > 0 ) { // Skip it all if we have no args
1219 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt );
1220 VMRegPair *parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt );
1221 int i;
1222 for( i = 0; i < argcnt; i++ ) {
1223 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type();
1224 }
1225 // V-call to pick proper calling convention
1226 call->calling_convention( sig_bt, parm_regs, argcnt );
1228 #ifdef ASSERT
1229 // Sanity check users' calling convention. Really handy during
1230 // the initial porting effort. Fairly expensive otherwise.
1231 { for (int i = 0; i<argcnt; i++) {
1232 if( !parm_regs[i].first()->is_valid() &&
1233 !parm_regs[i].second()->is_valid() ) continue;
1234 VMReg reg1 = parm_regs[i].first();
1235 VMReg reg2 = parm_regs[i].second();
1236 for (int j = 0; j < i; j++) {
1237 if( !parm_regs[j].first()->is_valid() &&
1238 !parm_regs[j].second()->is_valid() ) continue;
1239 VMReg reg3 = parm_regs[j].first();
1240 VMReg reg4 = parm_regs[j].second();
1241 if( !reg1->is_valid() ) {
1242 assert( !reg2->is_valid(), "valid halvsies" );
1243 } else if( !reg3->is_valid() ) {
1244 assert( !reg4->is_valid(), "valid halvsies" );
1245 } else {
1246 assert( reg1 != reg2, "calling conv. must produce distinct regs");
1247 assert( reg1 != reg3, "calling conv. must produce distinct regs");
1248 assert( reg1 != reg4, "calling conv. must produce distinct regs");
1249 assert( reg2 != reg3, "calling conv. must produce distinct regs");
1250 assert( reg2 != reg4 || !reg2->is_valid(), "calling conv. must produce distinct regs");
1251 assert( reg3 != reg4, "calling conv. must produce distinct regs");
1252 }
1253 }
1254 }
1255 }
1256 #endif
1258 // Visit each argument. Compute its outgoing register mask.
1259 // Return results now can have 2 bits returned.
1260 // Compute max over all outgoing arguments both per call-site
1261 // and over the entire method.
1262 for( i = 0; i < argcnt; i++ ) {
1263 // Address of incoming argument mask to fill in
1264 RegMask *rm = &mcall->_in_rms[i+TypeFunc::Parms];
1265 if( !parm_regs[i].first()->is_valid() &&
1266 !parm_regs[i].second()->is_valid() ) {
1267 continue; // Avoid Halves
1268 }
1269 // Grab first register, adjust stack slots and insert in mask.
1270 OptoReg::Name reg1 = warp_outgoing_stk_arg(parm_regs[i].first(), begin_out_arg_area, out_arg_limit_per_call );
1271 if (OptoReg::is_valid(reg1))
1272 rm->Insert( reg1 );
1273 // Grab second register (if any), adjust stack slots and insert in mask.
1274 OptoReg::Name reg2 = warp_outgoing_stk_arg(parm_regs[i].second(), begin_out_arg_area, out_arg_limit_per_call );
1275 if (OptoReg::is_valid(reg2))
1276 rm->Insert( reg2 );
1277 } // End of for all arguments
1279 // Compute number of stack slots needed to restore stack in case of
1280 // Pascal-style argument popping.
1281 mcall->_argsize = out_arg_limit_per_call - begin_out_arg_area;
1282 }
1284 if (is_method_handle_invoke) {
1285 // Kill some extra stack space in case method handles want to do
1286 // a little in-place argument insertion.
1287 // FIXME: Is this still necessary?
1288 int regs_per_word = NOT_LP64(1) LP64_ONLY(2); // %%% make a global const!
1289 out_arg_limit_per_call += Method::extra_stack_entries() * regs_per_word;
1290 // Do not update mcall->_argsize because (a) the extra space is not
1291 // pushed as arguments and (b) _argsize is dead (not used anywhere).
1292 }
1294 // Compute the max stack slot killed by any call. These will not be
1295 // available for debug info, and will be used to adjust FIRST_STACK_mask
1296 // after all call sites have been visited.
1297 if( _out_arg_limit < out_arg_limit_per_call)
1298 _out_arg_limit = out_arg_limit_per_call;
1300 if (mcall) {
1301 // Kill the outgoing argument area, including any non-argument holes and
1302 // any legacy C-killed slots. Use Fat-Projections to do the killing.
1303 // Since the max-per-method covers the max-per-call-site and debug info
1304 // is excluded on the max-per-method basis, debug info cannot land in
1305 // this killed area.
1306 uint r_cnt = mcall->tf()->range()->cnt();
1307 MachProjNode *proj = new (C) MachProjNode( mcall, r_cnt+10000, RegMask::Empty, MachProjNode::fat_proj );
1308 if (!RegMask::can_represent_arg(OptoReg::Name(out_arg_limit_per_call-1))) {
1309 C->record_method_not_compilable_all_tiers("unsupported outgoing calling sequence");
1310 } else {
1311 for (int i = begin_out_arg_area; i < out_arg_limit_per_call; i++)
1312 proj->_rout.Insert(OptoReg::Name(i));
1313 }
1314 if( proj->_rout.is_NotEmpty() )
1315 _proj_list.push(proj);
1316 }
1317 // Transfer the safepoint information from the call to the mcall
1318 // Move the JVMState list
1319 msfpt->set_jvms(sfpt->jvms());
1320 for (JVMState* jvms = msfpt->jvms(); jvms; jvms = jvms->caller()) {
1321 jvms->set_map(sfpt);
1322 }
1324 // Debug inputs begin just after the last incoming parameter
1325 assert( (mcall == NULL) || (mcall->jvms() == NULL) ||
1326 (mcall->jvms()->debug_start() + mcall->_jvmadj == mcall->tf()->domain()->cnt()), "" );
1328 // Move the OopMap
1329 msfpt->_oop_map = sfpt->_oop_map;
1331 // Registers killed by the call are set in the local scheduling pass
1332 // of Global Code Motion.
1333 return msfpt;
1334 }
1336 //---------------------------match_tree----------------------------------------
1337 // Match a Ideal Node DAG - turn it into a tree; Label & Reduce. Used as part
1338 // of the whole-sale conversion from Ideal to Mach Nodes. Also used for
1339 // making GotoNodes while building the CFG and in init_spill_mask() to identify
1340 // a Load's result RegMask for memoization in idealreg2regmask[]
1341 MachNode *Matcher::match_tree( const Node *n ) {
1342 assert( n->Opcode() != Op_Phi, "cannot match" );
1343 assert( !n->is_block_start(), "cannot match" );
1344 // Set the mark for all locally allocated State objects.
1345 // When this call returns, the _states_arena arena will be reset
1346 // freeing all State objects.
1347 ResourceMark rm( &_states_arena );
1349 LabelRootDepth = 0;
1351 // StoreNodes require their Memory input to match any LoadNodes
1352 Node *mem = n->is_Store() ? n->in(MemNode::Memory) : (Node*)1 ;
1353 #ifdef ASSERT
1354 Node* save_mem_node = _mem_node;
1355 _mem_node = n->is_Store() ? (Node*)n : NULL;
1356 #endif
1357 // State object for root node of match tree
1358 // Allocate it on _states_arena - stack allocation can cause stack overflow.
1359 State *s = new (&_states_arena) State;
1360 s->_kids[0] = NULL;
1361 s->_kids[1] = NULL;
1362 s->_leaf = (Node*)n;
1363 // Label the input tree, allocating labels from top-level arena
1364 Label_Root( n, s, n->in(0), mem );
1365 if (C->failing()) return NULL;
1367 // The minimum cost match for the whole tree is found at the root State
1368 uint mincost = max_juint;
1369 uint cost = max_juint;
1370 uint i;
1371 for( i = 0; i < NUM_OPERANDS; i++ ) {
1372 if( s->valid(i) && // valid entry and
1373 s->_cost[i] < cost && // low cost and
1374 s->_rule[i] >= NUM_OPERANDS ) // not an operand
1375 cost = s->_cost[mincost=i];
1376 }
1377 if (mincost == max_juint) {
1378 #ifndef PRODUCT
1379 tty->print("No matching rule for:");
1380 s->dump();
1381 #endif
1382 Matcher::soft_match_failure();
1383 return NULL;
1384 }
1385 // Reduce input tree based upon the state labels to machine Nodes
1386 MachNode *m = ReduceInst( s, s->_rule[mincost], mem );
1387 #ifdef ASSERT
1388 _old2new_map.map(n->_idx, m);
1389 _new2old_map.map(m->_idx, (Node*)n);
1390 #endif
1392 // Add any Matcher-ignored edges
1393 uint cnt = n->req();
1394 uint start = 1;
1395 if( mem != (Node*)1 ) start = MemNode::Memory+1;
1396 if( n->is_AddP() ) {
1397 assert( mem == (Node*)1, "" );
1398 start = AddPNode::Base+1;
1399 }
1400 for( i = start; i < cnt; i++ ) {
1401 if( !n->match_edge(i) ) {
1402 if( i < m->req() )
1403 m->ins_req( i, n->in(i) );
1404 else
1405 m->add_req( n->in(i) );
1406 }
1407 }
1409 debug_only( _mem_node = save_mem_node; )
1410 return m;
1411 }
1414 //------------------------------match_into_reg---------------------------------
1415 // Choose to either match this Node in a register or part of the current
1416 // match tree. Return true for requiring a register and false for matching
1417 // as part of the current match tree.
1418 static bool match_into_reg( const Node *n, Node *m, Node *control, int i, bool shared ) {
1420 const Type *t = m->bottom_type();
1422 if (t->singleton()) {
1423 // Never force constants into registers. Allow them to match as
1424 // constants or registers. Copies of the same value will share
1425 // the same register. See find_shared_node.
1426 return false;
1427 } else { // Not a constant
1428 // Stop recursion if they have different Controls.
1429 Node* m_control = m->in(0);
1430 // Control of load's memory can post-dominates load's control.
1431 // So use it since load can't float above its memory.
1432 Node* mem_control = (m->is_Load()) ? m->in(MemNode::Memory)->in(0) : NULL;
1433 if (control && m_control && control != m_control && control != mem_control) {
1435 // Actually, we can live with the most conservative control we
1436 // find, if it post-dominates the others. This allows us to
1437 // pick up load/op/store trees where the load can float a little
1438 // above the store.
1439 Node *x = control;
1440 const uint max_scan = 6; // Arbitrary scan cutoff
1441 uint j;
1442 for (j=0; j<max_scan; j++) {
1443 if (x->is_Region()) // Bail out at merge points
1444 return true;
1445 x = x->in(0);
1446 if (x == m_control) // Does 'control' post-dominate
1447 break; // m->in(0)? If so, we can use it
1448 if (x == mem_control) // Does 'control' post-dominate
1449 break; // mem_control? If so, we can use it
1450 }
1451 if (j == max_scan) // No post-domination before scan end?
1452 return true; // Then break the match tree up
1453 }
1454 if ((m->is_DecodeN() && Matcher::narrow_oop_use_complex_address()) ||
1455 (m->is_DecodeNKlass() && Matcher::narrow_klass_use_complex_address())) {
1456 // These are commonly used in address expressions and can
1457 // efficiently fold into them on X64 in some cases.
1458 return false;
1459 }
1460 }
1462 // Not forceable cloning. If shared, put it into a register.
1463 return shared;
1464 }
1467 //------------------------------Instruction Selection--------------------------
1468 // Label method walks a "tree" of nodes, using the ADLC generated DFA to match
1469 // ideal nodes to machine instructions. Trees are delimited by shared Nodes,
1470 // things the Matcher does not match (e.g., Memory), and things with different
1471 // Controls (hence forced into different blocks). We pass in the Control
1472 // selected for this entire State tree.
1474 // The Matcher works on Trees, but an Intel add-to-memory requires a DAG: the
1475 // Store and the Load must have identical Memories (as well as identical
1476 // pointers). Since the Matcher does not have anything for Memory (and
1477 // does not handle DAGs), I have to match the Memory input myself. If the
1478 // Tree root is a Store, I require all Loads to have the identical memory.
1479 Node *Matcher::Label_Root( const Node *n, State *svec, Node *control, const Node *mem){
1480 // Since Label_Root is a recursive function, its possible that we might run
1481 // out of stack space. See bugs 6272980 & 6227033 for more info.
1482 LabelRootDepth++;
1483 if (LabelRootDepth > MaxLabelRootDepth) {
1484 C->record_method_not_compilable_all_tiers("Out of stack space, increase MaxLabelRootDepth");
1485 return NULL;
1486 }
1487 uint care = 0; // Edges matcher cares about
1488 uint cnt = n->req();
1489 uint i = 0;
1491 // Examine children for memory state
1492 // Can only subsume a child into your match-tree if that child's memory state
1493 // is not modified along the path to another input.
1494 // It is unsafe even if the other inputs are separate roots.
1495 Node *input_mem = NULL;
1496 for( i = 1; i < cnt; i++ ) {
1497 if( !n->match_edge(i) ) continue;
1498 Node *m = n->in(i); // Get ith input
1499 assert( m, "expect non-null children" );
1500 if( m->is_Load() ) {
1501 if( input_mem == NULL ) {
1502 input_mem = m->in(MemNode::Memory);
1503 } else if( input_mem != m->in(MemNode::Memory) ) {
1504 input_mem = NodeSentinel;
1505 }
1506 }
1507 }
1509 for( i = 1; i < cnt; i++ ){// For my children
1510 if( !n->match_edge(i) ) continue;
1511 Node *m = n->in(i); // Get ith input
1512 // Allocate states out of a private arena
1513 State *s = new (&_states_arena) State;
1514 svec->_kids[care++] = s;
1515 assert( care <= 2, "binary only for now" );
1517 // Recursively label the State tree.
1518 s->_kids[0] = NULL;
1519 s->_kids[1] = NULL;
1520 s->_leaf = m;
1522 // Check for leaves of the State Tree; things that cannot be a part of
1523 // the current tree. If it finds any, that value is matched as a
1524 // register operand. If not, then the normal matching is used.
1525 if( match_into_reg(n, m, control, i, is_shared(m)) ||
1526 //
1527 // Stop recursion if this is LoadNode and the root of this tree is a
1528 // StoreNode and the load & store have different memories.
1529 ((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) ||
1530 // Can NOT include the match of a subtree when its memory state
1531 // is used by any of the other subtrees
1532 (input_mem == NodeSentinel) ) {
1533 #ifndef PRODUCT
1534 // Print when we exclude matching due to different memory states at input-loads
1535 if( PrintOpto && (Verbose && WizardMode) && (input_mem == NodeSentinel)
1536 && !((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) ) {
1537 tty->print_cr("invalid input_mem");
1538 }
1539 #endif
1540 // Switch to a register-only opcode; this value must be in a register
1541 // and cannot be subsumed as part of a larger instruction.
1542 s->DFA( m->ideal_reg(), m );
1544 } else {
1545 // If match tree has no control and we do, adopt it for entire tree
1546 if( control == NULL && m->in(0) != NULL && m->req() > 1 )
1547 control = m->in(0); // Pick up control
1548 // Else match as a normal part of the match tree.
1549 control = Label_Root(m,s,control,mem);
1550 if (C->failing()) return NULL;
1551 }
1552 }
1555 // Call DFA to match this node, and return
1556 svec->DFA( n->Opcode(), n );
1558 #ifdef ASSERT
1559 uint x;
1560 for( x = 0; x < _LAST_MACH_OPER; x++ )
1561 if( svec->valid(x) )
1562 break;
1564 if (x >= _LAST_MACH_OPER) {
1565 n->dump();
1566 svec->dump();
1567 assert( false, "bad AD file" );
1568 }
1569 #endif
1570 return control;
1571 }
1574 // Con nodes reduced using the same rule can share their MachNode
1575 // which reduces the number of copies of a constant in the final
1576 // program. The register allocator is free to split uses later to
1577 // split live ranges.
1578 MachNode* Matcher::find_shared_node(Node* leaf, uint rule) {
1579 if (!leaf->is_Con() && !leaf->is_DecodeNarrowPtr()) return NULL;
1581 // See if this Con has already been reduced using this rule.
1582 if (_shared_nodes.Size() <= leaf->_idx) return NULL;
1583 MachNode* last = (MachNode*)_shared_nodes.at(leaf->_idx);
1584 if (last != NULL && rule == last->rule()) {
1585 // Don't expect control change for DecodeN
1586 if (leaf->is_DecodeNarrowPtr())
1587 return last;
1588 // Get the new space root.
1589 Node* xroot = new_node(C->root());
1590 if (xroot == NULL) {
1591 // This shouldn't happen give the order of matching.
1592 return NULL;
1593 }
1595 // Shared constants need to have their control be root so they
1596 // can be scheduled properly.
1597 Node* control = last->in(0);
1598 if (control != xroot) {
1599 if (control == NULL || control == C->root()) {
1600 last->set_req(0, xroot);
1601 } else {
1602 assert(false, "unexpected control");
1603 return NULL;
1604 }
1605 }
1606 return last;
1607 }
1608 return NULL;
1609 }
1612 //------------------------------ReduceInst-------------------------------------
1613 // Reduce a State tree (with given Control) into a tree of MachNodes.
1614 // This routine (and it's cohort ReduceOper) convert Ideal Nodes into
1615 // complicated machine Nodes. Each MachNode covers some tree of Ideal Nodes.
1616 // Each MachNode has a number of complicated MachOper operands; each
1617 // MachOper also covers a further tree of Ideal Nodes.
1619 // The root of the Ideal match tree is always an instruction, so we enter
1620 // the recursion here. After building the MachNode, we need to recurse
1621 // the tree checking for these cases:
1622 // (1) Child is an instruction -
1623 // Build the instruction (recursively), add it as an edge.
1624 // Build a simple operand (register) to hold the result of the instruction.
1625 // (2) Child is an interior part of an instruction -
1626 // Skip over it (do nothing)
1627 // (3) Child is the start of a operand -
1628 // Build the operand, place it inside the instruction
1629 // Call ReduceOper.
1630 MachNode *Matcher::ReduceInst( State *s, int rule, Node *&mem ) {
1631 assert( rule >= NUM_OPERANDS, "called with operand rule" );
1633 MachNode* shared_node = find_shared_node(s->_leaf, rule);
1634 if (shared_node != NULL) {
1635 return shared_node;
1636 }
1638 // Build the object to represent this state & prepare for recursive calls
1639 MachNode *mach = s->MachNodeGenerator( rule, C );
1640 mach->_opnds[0] = s->MachOperGenerator( _reduceOp[rule], C );
1641 assert( mach->_opnds[0] != NULL, "Missing result operand" );
1642 Node *leaf = s->_leaf;
1643 // Check for instruction or instruction chain rule
1644 if( rule >= _END_INST_CHAIN_RULE || rule < _BEGIN_INST_CHAIN_RULE ) {
1645 assert(C->node_arena()->contains(s->_leaf) || !has_new_node(s->_leaf),
1646 "duplicating node that's already been matched");
1647 // Instruction
1648 mach->add_req( leaf->in(0) ); // Set initial control
1649 // Reduce interior of complex instruction
1650 ReduceInst_Interior( s, rule, mem, mach, 1 );
1651 } else {
1652 // Instruction chain rules are data-dependent on their inputs
1653 mach->add_req(0); // Set initial control to none
1654 ReduceInst_Chain_Rule( s, rule, mem, mach );
1655 }
1657 // If a Memory was used, insert a Memory edge
1658 if( mem != (Node*)1 ) {
1659 mach->ins_req(MemNode::Memory,mem);
1660 #ifdef ASSERT
1661 // Verify adr type after matching memory operation
1662 const MachOper* oper = mach->memory_operand();
1663 if (oper != NULL && oper != (MachOper*)-1) {
1664 // It has a unique memory operand. Find corresponding ideal mem node.
1665 Node* m = NULL;
1666 if (leaf->is_Mem()) {
1667 m = leaf;
1668 } else {
1669 m = _mem_node;
1670 assert(m != NULL && m->is_Mem(), "expecting memory node");
1671 }
1672 const Type* mach_at = mach->adr_type();
1673 // DecodeN node consumed by an address may have different type
1674 // then its input. Don't compare types for such case.
1675 if (m->adr_type() != mach_at &&
1676 (m->in(MemNode::Address)->is_DecodeNarrowPtr() ||
1677 m->in(MemNode::Address)->is_AddP() &&
1678 m->in(MemNode::Address)->in(AddPNode::Address)->is_DecodeNarrowPtr() ||
1679 m->in(MemNode::Address)->is_AddP() &&
1680 m->in(MemNode::Address)->in(AddPNode::Address)->is_AddP() &&
1681 m->in(MemNode::Address)->in(AddPNode::Address)->in(AddPNode::Address)->is_DecodeNarrowPtr())) {
1682 mach_at = m->adr_type();
1683 }
1684 if (m->adr_type() != mach_at) {
1685 m->dump();
1686 tty->print_cr("mach:");
1687 mach->dump(1);
1688 }
1689 assert(m->adr_type() == mach_at, "matcher should not change adr type");
1690 }
1691 #endif
1692 }
1694 // If the _leaf is an AddP, insert the base edge
1695 if( leaf->is_AddP() )
1696 mach->ins_req(AddPNode::Base,leaf->in(AddPNode::Base));
1698 uint num_proj = _proj_list.size();
1700 // Perform any 1-to-many expansions required
1701 MachNode *ex = mach->Expand(s,_proj_list, mem);
1702 if( ex != mach ) {
1703 assert(ex->ideal_reg() == mach->ideal_reg(), "ideal types should match");
1704 if( ex->in(1)->is_Con() )
1705 ex->in(1)->set_req(0, C->root());
1706 // Remove old node from the graph
1707 for( uint i=0; i<mach->req(); i++ ) {
1708 mach->set_req(i,NULL);
1709 }
1710 #ifdef ASSERT
1711 _new2old_map.map(ex->_idx, s->_leaf);
1712 #endif
1713 }
1715 // PhaseChaitin::fixup_spills will sometimes generate spill code
1716 // via the matcher. By the time, nodes have been wired into the CFG,
1717 // and any further nodes generated by expand rules will be left hanging
1718 // in space, and will not get emitted as output code. Catch this.
1719 // Also, catch any new register allocation constraints ("projections")
1720 // generated belatedly during spill code generation.
1721 if (_allocation_started) {
1722 guarantee(ex == mach, "no expand rules during spill generation");
1723 guarantee(_proj_list.size() == num_proj, "no allocation during spill generation");
1724 }
1726 if (leaf->is_Con() || leaf->is_DecodeNarrowPtr()) {
1727 // Record the con for sharing
1728 _shared_nodes.map(leaf->_idx, ex);
1729 }
1731 return ex;
1732 }
1734 void Matcher::ReduceInst_Chain_Rule( State *s, int rule, Node *&mem, MachNode *mach ) {
1735 // 'op' is what I am expecting to receive
1736 int op = _leftOp[rule];
1737 // Operand type to catch childs result
1738 // This is what my child will give me.
1739 int opnd_class_instance = s->_rule[op];
1740 // Choose between operand class or not.
1741 // This is what I will receive.
1742 int catch_op = (FIRST_OPERAND_CLASS <= op && op < NUM_OPERANDS) ? opnd_class_instance : op;
1743 // New rule for child. Chase operand classes to get the actual rule.
1744 int newrule = s->_rule[catch_op];
1746 if( newrule < NUM_OPERANDS ) {
1747 // Chain from operand or operand class, may be output of shared node
1748 assert( 0 <= opnd_class_instance && opnd_class_instance < NUM_OPERANDS,
1749 "Bad AD file: Instruction chain rule must chain from operand");
1750 // Insert operand into array of operands for this instruction
1751 mach->_opnds[1] = s->MachOperGenerator( opnd_class_instance, C );
1753 ReduceOper( s, newrule, mem, mach );
1754 } else {
1755 // Chain from the result of an instruction
1756 assert( newrule >= _LAST_MACH_OPER, "Do NOT chain from internal operand");
1757 mach->_opnds[1] = s->MachOperGenerator( _reduceOp[catch_op], C );
1758 Node *mem1 = (Node*)1;
1759 debug_only(Node *save_mem_node = _mem_node;)
1760 mach->add_req( ReduceInst(s, newrule, mem1) );
1761 debug_only(_mem_node = save_mem_node;)
1762 }
1763 return;
1764 }
1767 uint Matcher::ReduceInst_Interior( State *s, int rule, Node *&mem, MachNode *mach, uint num_opnds ) {
1768 if( s->_leaf->is_Load() ) {
1769 Node *mem2 = s->_leaf->in(MemNode::Memory);
1770 assert( mem == (Node*)1 || mem == mem2, "multiple Memories being matched at once?" );
1771 debug_only( if( mem == (Node*)1 ) _mem_node = s->_leaf;)
1772 mem = mem2;
1773 }
1774 if( s->_leaf->in(0) != NULL && s->_leaf->req() > 1) {
1775 if( mach->in(0) == NULL )
1776 mach->set_req(0, s->_leaf->in(0));
1777 }
1779 // Now recursively walk the state tree & add operand list.
1780 for( uint i=0; i<2; i++ ) { // binary tree
1781 State *newstate = s->_kids[i];
1782 if( newstate == NULL ) break; // Might only have 1 child
1783 // 'op' is what I am expecting to receive
1784 int op;
1785 if( i == 0 ) {
1786 op = _leftOp[rule];
1787 } else {
1788 op = _rightOp[rule];
1789 }
1790 // Operand type to catch childs result
1791 // This is what my child will give me.
1792 int opnd_class_instance = newstate->_rule[op];
1793 // Choose between operand class or not.
1794 // This is what I will receive.
1795 int catch_op = (op >= FIRST_OPERAND_CLASS && op < NUM_OPERANDS) ? opnd_class_instance : op;
1796 // New rule for child. Chase operand classes to get the actual rule.
1797 int newrule = newstate->_rule[catch_op];
1799 if( newrule < NUM_OPERANDS ) { // Operand/operandClass or internalOp/instruction?
1800 // Operand/operandClass
1801 // Insert operand into array of operands for this instruction
1802 mach->_opnds[num_opnds++] = newstate->MachOperGenerator( opnd_class_instance, C );
1803 ReduceOper( newstate, newrule, mem, mach );
1805 } else { // Child is internal operand or new instruction
1806 if( newrule < _LAST_MACH_OPER ) { // internal operand or instruction?
1807 // internal operand --> call ReduceInst_Interior
1808 // Interior of complex instruction. Do nothing but recurse.
1809 num_opnds = ReduceInst_Interior( newstate, newrule, mem, mach, num_opnds );
1810 } else {
1811 // instruction --> call build operand( ) to catch result
1812 // --> ReduceInst( newrule )
1813 mach->_opnds[num_opnds++] = s->MachOperGenerator( _reduceOp[catch_op], C );
1814 Node *mem1 = (Node*)1;
1815 debug_only(Node *save_mem_node = _mem_node;)
1816 mach->add_req( ReduceInst( newstate, newrule, mem1 ) );
1817 debug_only(_mem_node = save_mem_node;)
1818 }
1819 }
1820 assert( mach->_opnds[num_opnds-1], "" );
1821 }
1822 return num_opnds;
1823 }
1825 // This routine walks the interior of possible complex operands.
1826 // At each point we check our children in the match tree:
1827 // (1) No children -
1828 // We are a leaf; add _leaf field as an input to the MachNode
1829 // (2) Child is an internal operand -
1830 // Skip over it ( do nothing )
1831 // (3) Child is an instruction -
1832 // Call ReduceInst recursively and
1833 // and instruction as an input to the MachNode
1834 void Matcher::ReduceOper( State *s, int rule, Node *&mem, MachNode *mach ) {
1835 assert( rule < _LAST_MACH_OPER, "called with operand rule" );
1836 State *kid = s->_kids[0];
1837 assert( kid == NULL || s->_leaf->in(0) == NULL, "internal operands have no control" );
1839 // Leaf? And not subsumed?
1840 if( kid == NULL && !_swallowed[rule] ) {
1841 mach->add_req( s->_leaf ); // Add leaf pointer
1842 return; // Bail out
1843 }
1845 if( s->_leaf->is_Load() ) {
1846 assert( mem == (Node*)1, "multiple Memories being matched at once?" );
1847 mem = s->_leaf->in(MemNode::Memory);
1848 debug_only(_mem_node = s->_leaf;)
1849 }
1850 if( s->_leaf->in(0) && s->_leaf->req() > 1) {
1851 if( !mach->in(0) )
1852 mach->set_req(0,s->_leaf->in(0));
1853 else {
1854 assert( s->_leaf->in(0) == mach->in(0), "same instruction, differing controls?" );
1855 }
1856 }
1858 for( uint i=0; kid != NULL && i<2; kid = s->_kids[1], i++ ) { // binary tree
1859 int newrule;
1860 if( i == 0 )
1861 newrule = kid->_rule[_leftOp[rule]];
1862 else
1863 newrule = kid->_rule[_rightOp[rule]];
1865 if( newrule < _LAST_MACH_OPER ) { // Operand or instruction?
1866 // Internal operand; recurse but do nothing else
1867 ReduceOper( kid, newrule, mem, mach );
1869 } else { // Child is a new instruction
1870 // Reduce the instruction, and add a direct pointer from this
1871 // machine instruction to the newly reduced one.
1872 Node *mem1 = (Node*)1;
1873 debug_only(Node *save_mem_node = _mem_node;)
1874 mach->add_req( ReduceInst( kid, newrule, mem1 ) );
1875 debug_only(_mem_node = save_mem_node;)
1876 }
1877 }
1878 }
1881 // -------------------------------------------------------------------------
1882 // Java-Java calling convention
1883 // (what you use when Java calls Java)
1885 //------------------------------find_receiver----------------------------------
1886 // For a given signature, return the OptoReg for parameter 0.
1887 OptoReg::Name Matcher::find_receiver( bool is_outgoing ) {
1888 VMRegPair regs;
1889 BasicType sig_bt = T_OBJECT;
1890 calling_convention(&sig_bt, ®s, 1, is_outgoing);
1891 // Return argument 0 register. In the LP64 build pointers
1892 // take 2 registers, but the VM wants only the 'main' name.
1893 return OptoReg::as_OptoReg(regs.first());
1894 }
1896 // A method-klass-holder may be passed in the inline_cache_reg
1897 // and then expanded into the inline_cache_reg and a method_oop register
1898 // defined in ad_<arch>.cpp
1901 //------------------------------find_shared------------------------------------
1902 // Set bits if Node is shared or otherwise a root
1903 void Matcher::find_shared( Node *n ) {
1904 // Allocate stack of size C->unique() * 2 to avoid frequent realloc
1905 MStack mstack(C->unique() * 2);
1906 // Mark nodes as address_visited if they are inputs to an address expression
1907 VectorSet address_visited(Thread::current()->resource_area());
1908 mstack.push(n, Visit); // Don't need to pre-visit root node
1909 while (mstack.is_nonempty()) {
1910 n = mstack.node(); // Leave node on stack
1911 Node_State nstate = mstack.state();
1912 uint nop = n->Opcode();
1913 if (nstate == Pre_Visit) {
1914 if (address_visited.test(n->_idx)) { // Visited in address already?
1915 // Flag as visited and shared now.
1916 set_visited(n);
1917 }
1918 if (is_visited(n)) { // Visited already?
1919 // Node is shared and has no reason to clone. Flag it as shared.
1920 // This causes it to match into a register for the sharing.
1921 set_shared(n); // Flag as shared and
1922 mstack.pop(); // remove node from stack
1923 continue;
1924 }
1925 nstate = Visit; // Not already visited; so visit now
1926 }
1927 if (nstate == Visit) {
1928 mstack.set_state(Post_Visit);
1929 set_visited(n); // Flag as visited now
1930 bool mem_op = false;
1932 switch( nop ) { // Handle some opcodes special
1933 case Op_Phi: // Treat Phis as shared roots
1934 case Op_Parm:
1935 case Op_Proj: // All handled specially during matching
1936 case Op_SafePointScalarObject:
1937 set_shared(n);
1938 set_dontcare(n);
1939 break;
1940 case Op_If:
1941 case Op_CountedLoopEnd:
1942 mstack.set_state(Alt_Post_Visit); // Alternative way
1943 // Convert (If (Bool (CmpX A B))) into (If (Bool) (CmpX A B)). Helps
1944 // with matching cmp/branch in 1 instruction. The Matcher needs the
1945 // Bool and CmpX side-by-side, because it can only get at constants
1946 // that are at the leaves of Match trees, and the Bool's condition acts
1947 // as a constant here.
1948 mstack.push(n->in(1), Visit); // Clone the Bool
1949 mstack.push(n->in(0), Pre_Visit); // Visit control input
1950 continue; // while (mstack.is_nonempty())
1951 case Op_ConvI2D: // These forms efficiently match with a prior
1952 case Op_ConvI2F: // Load but not a following Store
1953 if( n->in(1)->is_Load() && // Prior load
1954 n->outcnt() == 1 && // Not already shared
1955 n->unique_out()->is_Store() ) // Following store
1956 set_shared(n); // Force it to be a root
1957 break;
1958 case Op_ReverseBytesI:
1959 case Op_ReverseBytesL:
1960 if( n->in(1)->is_Load() && // Prior load
1961 n->outcnt() == 1 ) // Not already shared
1962 set_shared(n); // Force it to be a root
1963 break;
1964 case Op_BoxLock: // Cant match until we get stack-regs in ADLC
1965 case Op_IfFalse:
1966 case Op_IfTrue:
1967 case Op_MachProj:
1968 case Op_MergeMem:
1969 case Op_Catch:
1970 case Op_CatchProj:
1971 case Op_CProj:
1972 case Op_JumpProj:
1973 case Op_JProj:
1974 case Op_NeverBranch:
1975 set_dontcare(n);
1976 break;
1977 case Op_Jump:
1978 mstack.push(n->in(1), Pre_Visit); // Switch Value (could be shared)
1979 mstack.push(n->in(0), Pre_Visit); // Visit Control input
1980 continue; // while (mstack.is_nonempty())
1981 case Op_StrComp:
1982 case Op_StrEquals:
1983 case Op_StrIndexOf:
1984 case Op_AryEq:
1985 set_shared(n); // Force result into register (it will be anyways)
1986 break;
1987 case Op_ConP: { // Convert pointers above the centerline to NUL
1988 TypeNode *tn = n->as_Type(); // Constants derive from type nodes
1989 const TypePtr* tp = tn->type()->is_ptr();
1990 if (tp->_ptr == TypePtr::AnyNull) {
1991 tn->set_type(TypePtr::NULL_PTR);
1992 }
1993 break;
1994 }
1995 case Op_ConN: { // Convert narrow pointers above the centerline to NUL
1996 TypeNode *tn = n->as_Type(); // Constants derive from type nodes
1997 const TypePtr* tp = tn->type()->make_ptr();
1998 if (tp && tp->_ptr == TypePtr::AnyNull) {
1999 tn->set_type(TypeNarrowOop::NULL_PTR);
2000 }
2001 break;
2002 }
2003 case Op_Binary: // These are introduced in the Post_Visit state.
2004 ShouldNotReachHere();
2005 break;
2006 case Op_ClearArray:
2007 case Op_SafePoint:
2008 mem_op = true;
2009 break;
2010 default:
2011 if( n->is_Store() ) {
2012 // Do match stores, despite no ideal reg
2013 mem_op = true;
2014 break;
2015 }
2016 if( n->is_Mem() ) { // Loads and LoadStores
2017 mem_op = true;
2018 // Loads must be root of match tree due to prior load conflict
2019 if( C->subsume_loads() == false )
2020 set_shared(n);
2021 }
2022 // Fall into default case
2023 if( !n->ideal_reg() )
2024 set_dontcare(n); // Unmatchable Nodes
2025 } // end_switch
2027 for(int i = n->req() - 1; i >= 0; --i) { // For my children
2028 Node *m = n->in(i); // Get ith input
2029 if (m == NULL) continue; // Ignore NULLs
2030 uint mop = m->Opcode();
2032 // Must clone all producers of flags, or we will not match correctly.
2033 // Suppose a compare setting int-flags is shared (e.g., a switch-tree)
2034 // then it will match into an ideal Op_RegFlags. Alas, the fp-flags
2035 // are also there, so we may match a float-branch to int-flags and
2036 // expect the allocator to haul the flags from the int-side to the
2037 // fp-side. No can do.
2038 if( _must_clone[mop] ) {
2039 mstack.push(m, Visit);
2040 continue; // for(int i = ...)
2041 }
2043 if( mop == Op_AddP && m->in(AddPNode::Base)->is_DecodeNarrowPtr()) {
2044 // Bases used in addresses must be shared but since
2045 // they are shared through a DecodeN they may appear
2046 // to have a single use so force sharing here.
2047 set_shared(m->in(AddPNode::Base)->in(1));
2048 }
2050 // Clone addressing expressions as they are "free" in memory access instructions
2051 if( mem_op && i == MemNode::Address && mop == Op_AddP ) {
2052 // Some inputs for address expression are not put on stack
2053 // to avoid marking them as shared and forcing them into register
2054 // if they are used only in address expressions.
2055 // But they should be marked as shared if there are other uses
2056 // besides address expressions.
2058 Node *off = m->in(AddPNode::Offset);
2059 if( off->is_Con() &&
2060 // When there are other uses besides address expressions
2061 // put it on stack and mark as shared.
2062 !is_visited(m) ) {
2063 address_visited.test_set(m->_idx); // Flag as address_visited
2064 Node *adr = m->in(AddPNode::Address);
2066 // Intel, ARM and friends can handle 2 adds in addressing mode
2067 if( clone_shift_expressions && adr->is_AddP() &&
2068 // AtomicAdd is not an addressing expression.
2069 // Cheap to find it by looking for screwy base.
2070 !adr->in(AddPNode::Base)->is_top() &&
2071 // Are there other uses besides address expressions?
2072 !is_visited(adr) ) {
2073 address_visited.set(adr->_idx); // Flag as address_visited
2074 Node *shift = adr->in(AddPNode::Offset);
2075 // Check for shift by small constant as well
2076 if( shift->Opcode() == Op_LShiftX && shift->in(2)->is_Con() &&
2077 shift->in(2)->get_int() <= 3 &&
2078 // Are there other uses besides address expressions?
2079 !is_visited(shift) ) {
2080 address_visited.set(shift->_idx); // Flag as address_visited
2081 mstack.push(shift->in(2), Visit);
2082 Node *conv = shift->in(1);
2083 #ifdef _LP64
2084 // Allow Matcher to match the rule which bypass
2085 // ConvI2L operation for an array index on LP64
2086 // if the index value is positive.
2087 if( conv->Opcode() == Op_ConvI2L &&
2088 conv->as_Type()->type()->is_long()->_lo >= 0 &&
2089 // Are there other uses besides address expressions?
2090 !is_visited(conv) ) {
2091 address_visited.set(conv->_idx); // Flag as address_visited
2092 mstack.push(conv->in(1), Pre_Visit);
2093 } else
2094 #endif
2095 mstack.push(conv, Pre_Visit);
2096 } else {
2097 mstack.push(shift, Pre_Visit);
2098 }
2099 mstack.push(adr->in(AddPNode::Address), Pre_Visit);
2100 mstack.push(adr->in(AddPNode::Base), Pre_Visit);
2101 } else { // Sparc, Alpha, PPC and friends
2102 mstack.push(adr, Pre_Visit);
2103 }
2105 // Clone X+offset as it also folds into most addressing expressions
2106 mstack.push(off, Visit);
2107 mstack.push(m->in(AddPNode::Base), Pre_Visit);
2108 continue; // for(int i = ...)
2109 } // if( off->is_Con() )
2110 } // if( mem_op &&
2111 mstack.push(m, Pre_Visit);
2112 } // for(int i = ...)
2113 }
2114 else if (nstate == Alt_Post_Visit) {
2115 mstack.pop(); // Remove node from stack
2116 // We cannot remove the Cmp input from the Bool here, as the Bool may be
2117 // shared and all users of the Bool need to move the Cmp in parallel.
2118 // This leaves both the Bool and the If pointing at the Cmp. To
2119 // prevent the Matcher from trying to Match the Cmp along both paths
2120 // BoolNode::match_edge always returns a zero.
2122 // We reorder the Op_If in a pre-order manner, so we can visit without
2123 // accidentally sharing the Cmp (the Bool and the If make 2 users).
2124 n->add_req( n->in(1)->in(1) ); // Add the Cmp next to the Bool
2125 }
2126 else if (nstate == Post_Visit) {
2127 mstack.pop(); // Remove node from stack
2129 // Now hack a few special opcodes
2130 switch( n->Opcode() ) { // Handle some opcodes special
2131 case Op_StorePConditional:
2132 case Op_StoreIConditional:
2133 case Op_StoreLConditional:
2134 case Op_CompareAndSwapI:
2135 case Op_CompareAndSwapL:
2136 case Op_CompareAndSwapP:
2137 case Op_CompareAndSwapN: { // Convert trinary to binary-tree
2138 Node *newval = n->in(MemNode::ValueIn );
2139 Node *oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
2140 Node *pair = new (C) BinaryNode( oldval, newval );
2141 n->set_req(MemNode::ValueIn,pair);
2142 n->del_req(LoadStoreConditionalNode::ExpectedIn);
2143 break;
2144 }
2145 case Op_CMoveD: // Convert trinary to binary-tree
2146 case Op_CMoveF:
2147 case Op_CMoveI:
2148 case Op_CMoveL:
2149 case Op_CMoveN:
2150 case Op_CMoveP: {
2151 // Restructure into a binary tree for Matching. It's possible that
2152 // we could move this code up next to the graph reshaping for IfNodes
2153 // or vice-versa, but I do not want to debug this for Ladybird.
2154 // 10/2/2000 CNC.
2155 Node *pair1 = new (C) BinaryNode(n->in(1),n->in(1)->in(1));
2156 n->set_req(1,pair1);
2157 Node *pair2 = new (C) BinaryNode(n->in(2),n->in(3));
2158 n->set_req(2,pair2);
2159 n->del_req(3);
2160 break;
2161 }
2162 case Op_LoopLimit: {
2163 Node *pair1 = new (C) BinaryNode(n->in(1),n->in(2));
2164 n->set_req(1,pair1);
2165 n->set_req(2,n->in(3));
2166 n->del_req(3);
2167 break;
2168 }
2169 case Op_StrEquals: {
2170 Node *pair1 = new (C) BinaryNode(n->in(2),n->in(3));
2171 n->set_req(2,pair1);
2172 n->set_req(3,n->in(4));
2173 n->del_req(4);
2174 break;
2175 }
2176 case Op_StrComp:
2177 case Op_StrIndexOf: {
2178 Node *pair1 = new (C) BinaryNode(n->in(2),n->in(3));
2179 n->set_req(2,pair1);
2180 Node *pair2 = new (C) BinaryNode(n->in(4),n->in(5));
2181 n->set_req(3,pair2);
2182 n->del_req(5);
2183 n->del_req(4);
2184 break;
2185 }
2186 default:
2187 break;
2188 }
2189 }
2190 else {
2191 ShouldNotReachHere();
2192 }
2193 } // end of while (mstack.is_nonempty())
2194 }
2196 #ifdef ASSERT
2197 // machine-independent root to machine-dependent root
2198 void Matcher::dump_old2new_map() {
2199 _old2new_map.dump();
2200 }
2201 #endif
2203 //---------------------------collect_null_checks-------------------------------
2204 // Find null checks in the ideal graph; write a machine-specific node for
2205 // it. Used by later implicit-null-check handling. Actually collects
2206 // either an IfTrue or IfFalse for the common NOT-null path, AND the ideal
2207 // value being tested.
2208 void Matcher::collect_null_checks( Node *proj, Node *orig_proj ) {
2209 Node *iff = proj->in(0);
2210 if( iff->Opcode() == Op_If ) {
2211 // During matching If's have Bool & Cmp side-by-side
2212 BoolNode *b = iff->in(1)->as_Bool();
2213 Node *cmp = iff->in(2);
2214 int opc = cmp->Opcode();
2215 if (opc != Op_CmpP && opc != Op_CmpN) return;
2217 const Type* ct = cmp->in(2)->bottom_type();
2218 if (ct == TypePtr::NULL_PTR ||
2219 (opc == Op_CmpN && ct == TypeNarrowOop::NULL_PTR)) {
2221 bool push_it = false;
2222 if( proj->Opcode() == Op_IfTrue ) {
2223 extern int all_null_checks_found;
2224 all_null_checks_found++;
2225 if( b->_test._test == BoolTest::ne ) {
2226 push_it = true;
2227 }
2228 } else {
2229 assert( proj->Opcode() == Op_IfFalse, "" );
2230 if( b->_test._test == BoolTest::eq ) {
2231 push_it = true;
2232 }
2233 }
2234 if( push_it ) {
2235 _null_check_tests.push(proj);
2236 Node* val = cmp->in(1);
2237 #ifdef _LP64
2238 if (val->bottom_type()->isa_narrowoop() &&
2239 !Matcher::narrow_oop_use_complex_address()) {
2240 //
2241 // Look for DecodeN node which should be pinned to orig_proj.
2242 // On platforms (Sparc) which can not handle 2 adds
2243 // in addressing mode we have to keep a DecodeN node and
2244 // use it to do implicit NULL check in address.
2245 //
2246 // DecodeN node was pinned to non-null path (orig_proj) during
2247 // CastPP transformation in final_graph_reshaping_impl().
2248 //
2249 uint cnt = orig_proj->outcnt();
2250 for (uint i = 0; i < orig_proj->outcnt(); i++) {
2251 Node* d = orig_proj->raw_out(i);
2252 if (d->is_DecodeN() && d->in(1) == val) {
2253 val = d;
2254 val->set_req(0, NULL); // Unpin now.
2255 // Mark this as special case to distinguish from
2256 // a regular case: CmpP(DecodeN, NULL).
2257 val = (Node*)(((intptr_t)val) | 1);
2258 break;
2259 }
2260 }
2261 }
2262 #endif
2263 _null_check_tests.push(val);
2264 }
2265 }
2266 }
2267 }
2269 //---------------------------validate_null_checks------------------------------
2270 // Its possible that the value being NULL checked is not the root of a match
2271 // tree. If so, I cannot use the value in an implicit null check.
2272 void Matcher::validate_null_checks( ) {
2273 uint cnt = _null_check_tests.size();
2274 for( uint i=0; i < cnt; i+=2 ) {
2275 Node *test = _null_check_tests[i];
2276 Node *val = _null_check_tests[i+1];
2277 bool is_decoden = ((intptr_t)val) & 1;
2278 val = (Node*)(((intptr_t)val) & ~1);
2279 if (has_new_node(val)) {
2280 Node* new_val = new_node(val);
2281 if (is_decoden) {
2282 assert(val->is_DecodeNarrowPtr() && val->in(0) == NULL, "sanity");
2283 // Note: new_val may have a control edge if
2284 // the original ideal node DecodeN was matched before
2285 // it was unpinned in Matcher::collect_null_checks().
2286 // Unpin the mach node and mark it.
2287 new_val->set_req(0, NULL);
2288 new_val = (Node*)(((intptr_t)new_val) | 1);
2289 }
2290 // Is a match-tree root, so replace with the matched value
2291 _null_check_tests.map(i+1, new_val);
2292 } else {
2293 // Yank from candidate list
2294 _null_check_tests.map(i+1,_null_check_tests[--cnt]);
2295 _null_check_tests.map(i,_null_check_tests[--cnt]);
2296 _null_check_tests.pop();
2297 _null_check_tests.pop();
2298 i-=2;
2299 }
2300 }
2301 }
2303 // Used by the DFA in dfa_xxx.cpp. Check for a following barrier or
2304 // atomic instruction acting as a store_load barrier without any
2305 // intervening volatile load, and thus we don't need a barrier here.
2306 // We retain the Node to act as a compiler ordering barrier.
2307 bool Matcher::post_store_load_barrier(const Node *vmb) {
2308 Compile *C = Compile::current();
2309 assert( vmb->is_MemBar(), "" );
2310 assert( vmb->Opcode() != Op_MemBarAcquire, "" );
2311 const MemBarNode *mem = (const MemBarNode*)vmb;
2313 // Get the Proj node, ctrl, that can be used to iterate forward
2314 Node *ctrl = NULL;
2315 DUIterator_Fast imax, i = mem->fast_outs(imax);
2316 while( true ) {
2317 ctrl = mem->fast_out(i); // Throw out-of-bounds if proj not found
2318 assert( ctrl->is_Proj(), "only projections here" );
2319 ProjNode *proj = (ProjNode*)ctrl;
2320 if( proj->_con == TypeFunc::Control &&
2321 !C->node_arena()->contains(ctrl) ) // Unmatched old-space only
2322 break;
2323 i++;
2324 }
2326 for( DUIterator_Fast jmax, j = ctrl->fast_outs(jmax); j < jmax; j++ ) {
2327 Node *x = ctrl->fast_out(j);
2328 int xop = x->Opcode();
2330 // We don't need current barrier if we see another or a lock
2331 // before seeing volatile load.
2332 //
2333 // Op_Fastunlock previously appeared in the Op_* list below.
2334 // With the advent of 1-0 lock operations we're no longer guaranteed
2335 // that a monitor exit operation contains a serializing instruction.
2337 if (xop == Op_MemBarVolatile ||
2338 xop == Op_FastLock ||
2339 xop == Op_CompareAndSwapL ||
2340 xop == Op_CompareAndSwapP ||
2341 xop == Op_CompareAndSwapN ||
2342 xop == Op_CompareAndSwapI)
2343 return true;
2345 if (x->is_MemBar()) {
2346 // We must retain this membar if there is an upcoming volatile
2347 // load, which will be preceded by acquire membar.
2348 if (xop == Op_MemBarAcquire)
2349 return false;
2350 // For other kinds of barriers, check by pretending we
2351 // are them, and seeing if we can be removed.
2352 else
2353 return post_store_load_barrier((const MemBarNode*)x);
2354 }
2356 // Delicate code to detect case of an upcoming fastlock block
2357 if( x->is_If() && x->req() > 1 &&
2358 !C->node_arena()->contains(x) ) { // Unmatched old-space only
2359 Node *iff = x;
2360 Node *bol = iff->in(1);
2361 // The iff might be some random subclass of If or bol might be Con-Top
2362 if (!bol->is_Bool()) return false;
2363 assert( bol->req() > 1, "" );
2364 return (bol->in(1)->Opcode() == Op_FastUnlock);
2365 }
2366 // probably not necessary to check for these
2367 if (x->is_Call() || x->is_SafePoint() || x->is_block_proj())
2368 return false;
2369 }
2370 return false;
2371 }
2373 //=============================================================================
2374 //---------------------------State---------------------------------------------
2375 State::State(void) {
2376 #ifdef ASSERT
2377 _id = 0;
2378 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe);
2379 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d);
2380 //memset(_cost, -1, sizeof(_cost));
2381 //memset(_rule, -1, sizeof(_rule));
2382 #endif
2383 memset(_valid, 0, sizeof(_valid));
2384 }
2386 #ifdef ASSERT
2387 State::~State() {
2388 _id = 99;
2389 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe);
2390 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d);
2391 memset(_cost, -3, sizeof(_cost));
2392 memset(_rule, -3, sizeof(_rule));
2393 }
2394 #endif
2396 #ifndef PRODUCT
2397 //---------------------------dump----------------------------------------------
2398 void State::dump() {
2399 tty->print("\n");
2400 dump(0);
2401 }
2403 void State::dump(int depth) {
2404 for( int j = 0; j < depth; j++ )
2405 tty->print(" ");
2406 tty->print("--N: ");
2407 _leaf->dump();
2408 uint i;
2409 for( i = 0; i < _LAST_MACH_OPER; i++ )
2410 // Check for valid entry
2411 if( valid(i) ) {
2412 for( int j = 0; j < depth; j++ )
2413 tty->print(" ");
2414 assert(_cost[i] != max_juint, "cost must be a valid value");
2415 assert(_rule[i] < _last_Mach_Node, "rule[i] must be valid rule");
2416 tty->print_cr("%s %d %s",
2417 ruleName[i], _cost[i], ruleName[_rule[i]] );
2418 }
2419 tty->print_cr("");
2421 for( i=0; i<2; i++ )
2422 if( _kids[i] )
2423 _kids[i]->dump(depth+1);
2424 }
2425 #endif