1 | /* |
2 | * Copyright (c) 1997, 2017, Oracle and/or its affiliates. All rights reserved. |
3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
4 | * |
5 | * This code is free software; you can redistribute it and/or modify it |
6 | * under the terms of the GNU General Public License version 2 only, as |
7 | * published by the Free Software Foundation. |
8 | * |
9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
12 | * version 2 for more details (a copy is included in the LICENSE file that |
13 | * accompanied this code). |
14 | * |
15 | * You should have received a copy of the GNU General Public License version |
16 | * 2 along with this work; if not, write to the Free Software Foundation, |
17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
18 | * |
19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
20 | * or visit www.oracle.com if you need additional information or have any |
21 | * questions. |
22 | * |
23 | */ |
24 | |
25 | #include "precompiled.hpp" |
26 | #include "gc/shared/barrierSet.hpp" |
27 | #include "gc/shared/c2/barrierSetC2.hpp" |
28 | #include "memory/allocation.inline.hpp" |
29 | #include "memory/resourceArea.hpp" |
30 | #include "oops/compressedOops.hpp" |
31 | #include "opto/ad.hpp" |
32 | #include "opto/addnode.hpp" |
33 | #include "opto/callnode.hpp" |
34 | #include "opto/idealGraphPrinter.hpp" |
35 | #include "opto/matcher.hpp" |
36 | #include "opto/memnode.hpp" |
37 | #include "opto/movenode.hpp" |
38 | #include "opto/opcodes.hpp" |
39 | #include "opto/regmask.hpp" |
40 | #include "opto/rootnode.hpp" |
41 | #include "opto/runtime.hpp" |
42 | #include "opto/type.hpp" |
43 | #include "opto/vectornode.hpp" |
44 | #include "runtime/os.hpp" |
45 | #include "runtime/sharedRuntime.hpp" |
46 | #include "utilities/align.hpp" |
47 | |
48 | OptoReg::Name OptoReg::c_frame_pointer; |
49 | |
50 | const RegMask *Matcher::idealreg2regmask[_last_machine_leaf]; |
51 | RegMask Matcher::mreg2regmask[_last_Mach_Reg]; |
52 | RegMask Matcher::STACK_ONLY_mask; |
53 | RegMask Matcher::c_frame_ptr_mask; |
54 | const uint Matcher::_begin_rematerialize = _BEGIN_REMATERIALIZE; |
55 | const uint Matcher::_end_rematerialize = _END_REMATERIALIZE; |
56 | |
57 | //---------------------------Matcher------------------------------------------- |
58 | Matcher::Matcher() |
59 | : PhaseTransform( Phase::Ins_Select ), |
60 | _states_arena(Chunk::medium_size, mtCompiler), |
61 | _visited(&_states_arena), |
62 | _shared(&_states_arena), |
63 | _dontcare(&_states_arena), |
64 | _reduceOp(reduceOp), _leftOp(leftOp), _rightOp(rightOp), |
65 | _swallowed(swallowed), |
66 | _begin_inst_chain_rule(_BEGIN_INST_CHAIN_RULE), |
67 | _end_inst_chain_rule(_END_INST_CHAIN_RULE), |
68 | _must_clone(must_clone), |
69 | _shared_nodes(C->comp_arena()), |
70 | #ifdef ASSERT |
71 | _old2new_map(C->comp_arena()), |
72 | _new2old_map(C->comp_arena()), |
73 | #endif |
74 | _allocation_started(false), |
75 | _ruleName(ruleName), |
76 | _register_save_policy(register_save_policy), |
77 | _c_reg_save_policy(c_reg_save_policy), |
78 | _register_save_type(register_save_type) { |
79 | C->set_matcher(this); |
80 | |
81 | idealreg2spillmask [Op_RegI] = NULL; |
82 | idealreg2spillmask [Op_RegN] = NULL; |
83 | idealreg2spillmask [Op_RegL] = NULL; |
84 | idealreg2spillmask [Op_RegF] = NULL; |
85 | idealreg2spillmask [Op_RegD] = NULL; |
86 | idealreg2spillmask [Op_RegP] = NULL; |
87 | idealreg2spillmask [Op_VecS] = NULL; |
88 | idealreg2spillmask [Op_VecD] = NULL; |
89 | idealreg2spillmask [Op_VecX] = NULL; |
90 | idealreg2spillmask [Op_VecY] = NULL; |
91 | idealreg2spillmask [Op_VecZ] = NULL; |
92 | idealreg2spillmask [Op_RegFlags] = NULL; |
93 | |
94 | idealreg2debugmask [Op_RegI] = NULL; |
95 | idealreg2debugmask [Op_RegN] = NULL; |
96 | idealreg2debugmask [Op_RegL] = NULL; |
97 | idealreg2debugmask [Op_RegF] = NULL; |
98 | idealreg2debugmask [Op_RegD] = NULL; |
99 | idealreg2debugmask [Op_RegP] = NULL; |
100 | idealreg2debugmask [Op_VecS] = NULL; |
101 | idealreg2debugmask [Op_VecD] = NULL; |
102 | idealreg2debugmask [Op_VecX] = NULL; |
103 | idealreg2debugmask [Op_VecY] = NULL; |
104 | idealreg2debugmask [Op_VecZ] = NULL; |
105 | idealreg2debugmask [Op_RegFlags] = NULL; |
106 | |
107 | idealreg2mhdebugmask[Op_RegI] = NULL; |
108 | idealreg2mhdebugmask[Op_RegN] = NULL; |
109 | idealreg2mhdebugmask[Op_RegL] = NULL; |
110 | idealreg2mhdebugmask[Op_RegF] = NULL; |
111 | idealreg2mhdebugmask[Op_RegD] = NULL; |
112 | idealreg2mhdebugmask[Op_RegP] = NULL; |
113 | idealreg2mhdebugmask[Op_VecS] = NULL; |
114 | idealreg2mhdebugmask[Op_VecD] = NULL; |
115 | idealreg2mhdebugmask[Op_VecX] = NULL; |
116 | idealreg2mhdebugmask[Op_VecY] = NULL; |
117 | idealreg2mhdebugmask[Op_VecZ] = NULL; |
118 | idealreg2mhdebugmask[Op_RegFlags] = NULL; |
119 | |
120 | debug_only(_mem_node = NULL;) // Ideal memory node consumed by mach node |
121 | } |
122 | |
123 | //------------------------------warp_incoming_stk_arg------------------------ |
124 | // This warps a VMReg into an OptoReg::Name |
125 | OptoReg::Name Matcher::warp_incoming_stk_arg( VMReg reg ) { |
126 | OptoReg::Name warped; |
127 | if( reg->is_stack() ) { // Stack slot argument? |
128 | warped = OptoReg::add(_old_SP, reg->reg2stack() ); |
129 | warped = OptoReg::add(warped, C->out_preserve_stack_slots()); |
130 | if( warped >= _in_arg_limit ) |
131 | _in_arg_limit = OptoReg::add(warped, 1); // Bump max stack slot seen |
132 | if (!RegMask::can_represent_arg(warped)) { |
133 | // the compiler cannot represent this method's calling sequence |
134 | C->record_method_not_compilable("unsupported incoming calling sequence" ); |
135 | return OptoReg::Bad; |
136 | } |
137 | return warped; |
138 | } |
139 | return OptoReg::as_OptoReg(reg); |
140 | } |
141 | |
142 | //---------------------------compute_old_SP------------------------------------ |
143 | OptoReg::Name Compile::compute_old_SP() { |
144 | int fixed = fixed_slots(); |
145 | int preserve = in_preserve_stack_slots(); |
146 | return OptoReg::stack2reg(align_up(fixed + preserve, (int)Matcher::stack_alignment_in_slots())); |
147 | } |
148 | |
149 | |
150 | |
151 | #ifdef ASSERT |
152 | void Matcher::verify_new_nodes_only(Node* xroot) { |
153 | // Make sure that the new graph only references new nodes |
154 | ResourceMark rm; |
155 | Unique_Node_List worklist; |
156 | VectorSet visited(Thread::current()->resource_area()); |
157 | worklist.push(xroot); |
158 | while (worklist.size() > 0) { |
159 | Node* n = worklist.pop(); |
160 | visited <<= n->_idx; |
161 | assert(C->node_arena()->contains(n), "dead node" ); |
162 | for (uint j = 0; j < n->req(); j++) { |
163 | Node* in = n->in(j); |
164 | if (in != NULL) { |
165 | assert(C->node_arena()->contains(in), "dead node" ); |
166 | if (!visited.test(in->_idx)) { |
167 | worklist.push(in); |
168 | } |
169 | } |
170 | } |
171 | } |
172 | } |
173 | #endif |
174 | |
175 | |
176 | //---------------------------match--------------------------------------------- |
177 | void Matcher::match( ) { |
178 | if( MaxLabelRootDepth < 100 ) { // Too small? |
179 | assert(false, "invalid MaxLabelRootDepth, increase it to 100 minimum" ); |
180 | MaxLabelRootDepth = 100; |
181 | } |
182 | // One-time initialization of some register masks. |
183 | init_spill_mask( C->root()->in(1) ); |
184 | _return_addr_mask = return_addr(); |
185 | #ifdef _LP64 |
186 | // Pointers take 2 slots in 64-bit land |
187 | _return_addr_mask.Insert(OptoReg::add(return_addr(),1)); |
188 | #endif |
189 | |
190 | // Map a Java-signature return type into return register-value |
191 | // machine registers for 0, 1 and 2 returned values. |
192 | const TypeTuple *range = C->tf()->range(); |
193 | if( range->cnt() > TypeFunc::Parms ) { // If not a void function |
194 | // Get ideal-register return type |
195 | uint ireg = range->field_at(TypeFunc::Parms)->ideal_reg(); |
196 | // Get machine return register |
197 | uint sop = C->start()->Opcode(); |
198 | OptoRegPair regs = return_value(ireg, false); |
199 | |
200 | // And mask for same |
201 | _return_value_mask = RegMask(regs.first()); |
202 | if( OptoReg::is_valid(regs.second()) ) |
203 | _return_value_mask.Insert(regs.second()); |
204 | } |
205 | |
206 | // --------------- |
207 | // Frame Layout |
208 | |
209 | // Need the method signature to determine the incoming argument types, |
210 | // because the types determine which registers the incoming arguments are |
211 | // in, and this affects the matched code. |
212 | const TypeTuple *domain = C->tf()->domain(); |
213 | uint argcnt = domain->cnt() - TypeFunc::Parms; |
214 | BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt ); |
215 | VMRegPair *vm_parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt ); |
216 | _parm_regs = NEW_RESOURCE_ARRAY( OptoRegPair, argcnt ); |
217 | _calling_convention_mask = NEW_RESOURCE_ARRAY( RegMask, argcnt ); |
218 | uint i; |
219 | for( i = 0; i<argcnt; i++ ) { |
220 | sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type(); |
221 | } |
222 | |
223 | // Pass array of ideal registers and length to USER code (from the AD file) |
224 | // that will convert this to an array of register numbers. |
225 | const StartNode *start = C->start(); |
226 | start->calling_convention( sig_bt, vm_parm_regs, argcnt ); |
227 | #ifdef ASSERT |
228 | // Sanity check users' calling convention. Real handy while trying to |
229 | // get the initial port correct. |
230 | { for (uint i = 0; i<argcnt; i++) { |
231 | if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) { |
232 | assert(domain->field_at(i+TypeFunc::Parms)==Type::HALF, "only allowed on halve" ); |
233 | _parm_regs[i].set_bad(); |
234 | continue; |
235 | } |
236 | VMReg parm_reg = vm_parm_regs[i].first(); |
237 | assert(parm_reg->is_valid(), "invalid arg?" ); |
238 | if (parm_reg->is_reg()) { |
239 | OptoReg::Name opto_parm_reg = OptoReg::as_OptoReg(parm_reg); |
240 | assert(can_be_java_arg(opto_parm_reg) || |
241 | C->stub_function() == CAST_FROM_FN_PTR(address, OptoRuntime::rethrow_C) || |
242 | opto_parm_reg == inline_cache_reg(), |
243 | "parameters in register must be preserved by runtime stubs" ); |
244 | } |
245 | for (uint j = 0; j < i; j++) { |
246 | assert(parm_reg != vm_parm_regs[j].first(), |
247 | "calling conv. must produce distinct regs" ); |
248 | } |
249 | } |
250 | } |
251 | #endif |
252 | |
253 | // Do some initial frame layout. |
254 | |
255 | // Compute the old incoming SP (may be called FP) as |
256 | // OptoReg::stack0() + locks + in_preserve_stack_slots + pad2. |
257 | _old_SP = C->compute_old_SP(); |
258 | assert( is_even(_old_SP), "must be even" ); |
259 | |
260 | // Compute highest incoming stack argument as |
261 | // _old_SP + out_preserve_stack_slots + incoming argument size. |
262 | _in_arg_limit = OptoReg::add(_old_SP, C->out_preserve_stack_slots()); |
263 | assert( is_even(_in_arg_limit), "out_preserve must be even" ); |
264 | for( i = 0; i < argcnt; i++ ) { |
265 | // Permit args to have no register |
266 | _calling_convention_mask[i].Clear(); |
267 | if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) { |
268 | continue; |
269 | } |
270 | // calling_convention returns stack arguments as a count of |
271 | // slots beyond OptoReg::stack0()/VMRegImpl::stack0. We need to convert this to |
272 | // the allocators point of view, taking into account all the |
273 | // preserve area, locks & pad2. |
274 | |
275 | OptoReg::Name reg1 = warp_incoming_stk_arg(vm_parm_regs[i].first()); |
276 | if( OptoReg::is_valid(reg1)) |
277 | _calling_convention_mask[i].Insert(reg1); |
278 | |
279 | OptoReg::Name reg2 = warp_incoming_stk_arg(vm_parm_regs[i].second()); |
280 | if( OptoReg::is_valid(reg2)) |
281 | _calling_convention_mask[i].Insert(reg2); |
282 | |
283 | // Saved biased stack-slot register number |
284 | _parm_regs[i].set_pair(reg2, reg1); |
285 | } |
286 | |
287 | // Finally, make sure the incoming arguments take up an even number of |
288 | // words, in case the arguments or locals need to contain doubleword stack |
289 | // slots. The rest of the system assumes that stack slot pairs (in |
290 | // particular, in the spill area) which look aligned will in fact be |
291 | // aligned relative to the stack pointer in the target machine. Double |
292 | // stack slots will always be allocated aligned. |
293 | _new_SP = OptoReg::Name(align_up(_in_arg_limit, (int)RegMask::SlotsPerLong)); |
294 | |
295 | // Compute highest outgoing stack argument as |
296 | // _new_SP + out_preserve_stack_slots + max(outgoing argument size). |
297 | _out_arg_limit = OptoReg::add(_new_SP, C->out_preserve_stack_slots()); |
298 | assert( is_even(_out_arg_limit), "out_preserve must be even" ); |
299 | |
300 | if (!RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1))) { |
301 | // the compiler cannot represent this method's calling sequence |
302 | C->record_method_not_compilable("must be able to represent all call arguments in reg mask" ); |
303 | } |
304 | |
305 | if (C->failing()) return; // bailed out on incoming arg failure |
306 | |
307 | // --------------- |
308 | // Collect roots of matcher trees. Every node for which |
309 | // _shared[_idx] is cleared is guaranteed to not be shared, and thus |
310 | // can be a valid interior of some tree. |
311 | find_shared( C->root() ); |
312 | find_shared( C->top() ); |
313 | |
314 | C->print_method(PHASE_BEFORE_MATCHING); |
315 | |
316 | // Create new ideal node ConP #NULL even if it does exist in old space |
317 | // to avoid false sharing if the corresponding mach node is not used. |
318 | // The corresponding mach node is only used in rare cases for derived |
319 | // pointers. |
320 | Node* new_ideal_null = ConNode::make(TypePtr::NULL_PTR); |
321 | |
322 | // Swap out to old-space; emptying new-space |
323 | Arena *old = C->node_arena()->move_contents(C->old_arena()); |
324 | |
325 | // Save debug and profile information for nodes in old space: |
326 | _old_node_note_array = C->node_note_array(); |
327 | if (_old_node_note_array != NULL) { |
328 | C->set_node_note_array(new(C->comp_arena()) GrowableArray<Node_Notes*> |
329 | (C->comp_arena(), _old_node_note_array->length(), |
330 | 0, NULL)); |
331 | } |
332 | |
333 | // Pre-size the new_node table to avoid the need for range checks. |
334 | grow_new_node_array(C->unique()); |
335 | |
336 | // Reset node counter so MachNodes start with _idx at 0 |
337 | int live_nodes = C->live_nodes(); |
338 | C->set_unique(0); |
339 | C->reset_dead_node_list(); |
340 | |
341 | // Recursively match trees from old space into new space. |
342 | // Correct leaves of new-space Nodes; they point to old-space. |
343 | _visited.Clear(); // Clear visit bits for xform call |
344 | C->set_cached_top_node(xform( C->top(), live_nodes )); |
345 | if (!C->failing()) { |
346 | Node* xroot = xform( C->root(), 1 ); |
347 | if (xroot == NULL) { |
348 | Matcher::soft_match_failure(); // recursive matching process failed |
349 | C->record_method_not_compilable("instruction match failed" ); |
350 | } else { |
351 | // During matching shared constants were attached to C->root() |
352 | // because xroot wasn't available yet, so transfer the uses to |
353 | // the xroot. |
354 | for( DUIterator_Fast jmax, j = C->root()->fast_outs(jmax); j < jmax; j++ ) { |
355 | Node* n = C->root()->fast_out(j); |
356 | if (C->node_arena()->contains(n)) { |
357 | assert(n->in(0) == C->root(), "should be control user" ); |
358 | n->set_req(0, xroot); |
359 | --j; |
360 | --jmax; |
361 | } |
362 | } |
363 | |
364 | // Generate new mach node for ConP #NULL |
365 | assert(new_ideal_null != NULL, "sanity" ); |
366 | _mach_null = match_tree(new_ideal_null); |
367 | // Don't set control, it will confuse GCM since there are no uses. |
368 | // The control will be set when this node is used first time |
369 | // in find_base_for_derived(). |
370 | assert(_mach_null != NULL, "" ); |
371 | |
372 | C->set_root(xroot->is_Root() ? xroot->as_Root() : NULL); |
373 | |
374 | #ifdef ASSERT |
375 | verify_new_nodes_only(xroot); |
376 | #endif |
377 | } |
378 | } |
379 | if (C->top() == NULL || C->root() == NULL) { |
380 | C->record_method_not_compilable("graph lost" ); // %%% cannot happen? |
381 | } |
382 | if (C->failing()) { |
383 | // delete old; |
384 | old->destruct_contents(); |
385 | return; |
386 | } |
387 | assert( C->top(), "" ); |
388 | assert( C->root(), "" ); |
389 | validate_null_checks(); |
390 | |
391 | // Now smoke old-space |
392 | NOT_DEBUG( old->destruct_contents() ); |
393 | |
394 | // ------------------------ |
395 | // Set up save-on-entry registers |
396 | Fixup_Save_On_Entry( ); |
397 | } |
398 | |
399 | |
400 | //------------------------------Fixup_Save_On_Entry---------------------------- |
401 | // The stated purpose of this routine is to take care of save-on-entry |
402 | // registers. However, the overall goal of the Match phase is to convert into |
403 | // machine-specific instructions which have RegMasks to guide allocation. |
404 | // So what this procedure really does is put a valid RegMask on each input |
405 | // to the machine-specific variations of all Return, TailCall and Halt |
406 | // instructions. It also adds edgs to define the save-on-entry values (and of |
407 | // course gives them a mask). |
408 | |
409 | static RegMask *init_input_masks( uint size, RegMask &ret_adr, RegMask &fp ) { |
410 | RegMask *rms = NEW_RESOURCE_ARRAY( RegMask, size ); |
411 | // Do all the pre-defined register masks |
412 | rms[TypeFunc::Control ] = RegMask::Empty; |
413 | rms[TypeFunc::I_O ] = RegMask::Empty; |
414 | rms[TypeFunc::Memory ] = RegMask::Empty; |
415 | rms[TypeFunc::ReturnAdr] = ret_adr; |
416 | rms[TypeFunc::FramePtr ] = fp; |
417 | return rms; |
418 | } |
419 | |
420 | #define NOF_STACK_MASKS (3*6+5) |
421 | |
422 | // Create the initial stack mask used by values spilling to the stack. |
423 | // Disallow any debug info in outgoing argument areas by setting the |
424 | // initial mask accordingly. |
425 | void Matcher::init_first_stack_mask() { |
426 | |
427 | // Allocate storage for spill masks as masks for the appropriate load type. |
428 | RegMask *rms = (RegMask*)C->comp_arena()->Amalloc_D(sizeof(RegMask) * NOF_STACK_MASKS); |
429 | |
430 | // Initialize empty placeholder masks into the newly allocated arena |
431 | for (int i = 0; i < NOF_STACK_MASKS; i++) { |
432 | new (rms + i) RegMask(); |
433 | } |
434 | |
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]; |
441 | |
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]; |
448 | |
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]; |
455 | |
456 | idealreg2spillmask [Op_VecS] = &rms[18]; |
457 | idealreg2spillmask [Op_VecD] = &rms[19]; |
458 | idealreg2spillmask [Op_VecX] = &rms[20]; |
459 | idealreg2spillmask [Op_VecY] = &rms[21]; |
460 | idealreg2spillmask [Op_VecZ] = &rms[22]; |
461 | |
462 | OptoReg::Name i; |
463 | |
464 | // At first, start with the empty mask |
465 | C->FIRST_STACK_mask().Clear(); |
466 | |
467 | // Add in the incoming argument area |
468 | OptoReg::Name init_in = OptoReg::add(_old_SP, C->out_preserve_stack_slots()); |
469 | for (i = init_in; i < _in_arg_limit; i = OptoReg::add(i,1)) { |
470 | C->FIRST_STACK_mask().Insert(i); |
471 | } |
472 | // Add in all bits past the outgoing argument area |
473 | guarantee(RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1)), |
474 | "must be able to represent all call arguments in reg mask" ); |
475 | OptoReg::Name init = _out_arg_limit; |
476 | for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) { |
477 | C->FIRST_STACK_mask().Insert(i); |
478 | } |
479 | // Finally, set the "infinite stack" bit. |
480 | C->FIRST_STACK_mask().set_AllStack(); |
481 | |
482 | // Make spill masks. Registers for their class, plus FIRST_STACK_mask. |
483 | RegMask aligned_stack_mask = C->FIRST_STACK_mask(); |
484 | // Keep spill masks aligned. |
485 | aligned_stack_mask.clear_to_pairs(); |
486 | assert(aligned_stack_mask.is_AllStack(), "should be infinite stack" ); |
487 | |
488 | *idealreg2spillmask[Op_RegP] = *idealreg2regmask[Op_RegP]; |
489 | #ifdef _LP64 |
490 | *idealreg2spillmask[Op_RegN] = *idealreg2regmask[Op_RegN]; |
491 | idealreg2spillmask[Op_RegN]->OR(C->FIRST_STACK_mask()); |
492 | idealreg2spillmask[Op_RegP]->OR(aligned_stack_mask); |
493 | #else |
494 | idealreg2spillmask[Op_RegP]->OR(C->FIRST_STACK_mask()); |
495 | #endif |
496 | *idealreg2spillmask[Op_RegI] = *idealreg2regmask[Op_RegI]; |
497 | idealreg2spillmask[Op_RegI]->OR(C->FIRST_STACK_mask()); |
498 | *idealreg2spillmask[Op_RegL] = *idealreg2regmask[Op_RegL]; |
499 | idealreg2spillmask[Op_RegL]->OR(aligned_stack_mask); |
500 | *idealreg2spillmask[Op_RegF] = *idealreg2regmask[Op_RegF]; |
501 | idealreg2spillmask[Op_RegF]->OR(C->FIRST_STACK_mask()); |
502 | *idealreg2spillmask[Op_RegD] = *idealreg2regmask[Op_RegD]; |
503 | idealreg2spillmask[Op_RegD]->OR(aligned_stack_mask); |
504 | |
505 | if (Matcher::vector_size_supported(T_BYTE,4)) { |
506 | *idealreg2spillmask[Op_VecS] = *idealreg2regmask[Op_VecS]; |
507 | idealreg2spillmask[Op_VecS]->OR(C->FIRST_STACK_mask()); |
508 | } |
509 | if (Matcher::vector_size_supported(T_FLOAT,2)) { |
510 | // For VecD we need dual alignment and 8 bytes (2 slots) for spills. |
511 | // RA guarantees such alignment since it is needed for Double and Long values. |
512 | *idealreg2spillmask[Op_VecD] = *idealreg2regmask[Op_VecD]; |
513 | idealreg2spillmask[Op_VecD]->OR(aligned_stack_mask); |
514 | } |
515 | if (Matcher::vector_size_supported(T_FLOAT,4)) { |
516 | // For VecX we need quadro alignment and 16 bytes (4 slots) for spills. |
517 | // |
518 | // RA can use input arguments stack slots for spills but until RA |
519 | // we don't know frame size and offset of input arg stack slots. |
520 | // |
521 | // Exclude last input arg stack slots to avoid spilling vectors there |
522 | // otherwise vector spills could stomp over stack slots in caller frame. |
523 | OptoReg::Name in = OptoReg::add(_in_arg_limit, -1); |
524 | for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecX); k++) { |
525 | aligned_stack_mask.Remove(in); |
526 | in = OptoReg::add(in, -1); |
527 | } |
528 | aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecX); |
529 | assert(aligned_stack_mask.is_AllStack(), "should be infinite stack" ); |
530 | *idealreg2spillmask[Op_VecX] = *idealreg2regmask[Op_VecX]; |
531 | idealreg2spillmask[Op_VecX]->OR(aligned_stack_mask); |
532 | } |
533 | if (Matcher::vector_size_supported(T_FLOAT,8)) { |
534 | // For VecY we need octo alignment and 32 bytes (8 slots) for spills. |
535 | OptoReg::Name in = OptoReg::add(_in_arg_limit, -1); |
536 | for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecY); k++) { |
537 | aligned_stack_mask.Remove(in); |
538 | in = OptoReg::add(in, -1); |
539 | } |
540 | aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecY); |
541 | assert(aligned_stack_mask.is_AllStack(), "should be infinite stack" ); |
542 | *idealreg2spillmask[Op_VecY] = *idealreg2regmask[Op_VecY]; |
543 | idealreg2spillmask[Op_VecY]->OR(aligned_stack_mask); |
544 | } |
545 | if (Matcher::vector_size_supported(T_FLOAT,16)) { |
546 | // For VecZ we need enough alignment and 64 bytes (16 slots) for spills. |
547 | OptoReg::Name in = OptoReg::add(_in_arg_limit, -1); |
548 | for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecZ); k++) { |
549 | aligned_stack_mask.Remove(in); |
550 | in = OptoReg::add(in, -1); |
551 | } |
552 | aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecZ); |
553 | assert(aligned_stack_mask.is_AllStack(), "should be infinite stack" ); |
554 | *idealreg2spillmask[Op_VecZ] = *idealreg2regmask[Op_VecZ]; |
555 | idealreg2spillmask[Op_VecZ]->OR(aligned_stack_mask); |
556 | } |
557 | if (UseFPUForSpilling) { |
558 | // This mask logic assumes that the spill operations are |
559 | // symmetric and that the registers involved are the same size. |
560 | // On sparc for instance we may have to use 64 bit moves will |
561 | // kill 2 registers when used with F0-F31. |
562 | idealreg2spillmask[Op_RegI]->OR(*idealreg2regmask[Op_RegF]); |
563 | idealreg2spillmask[Op_RegF]->OR(*idealreg2regmask[Op_RegI]); |
564 | #ifdef _LP64 |
565 | idealreg2spillmask[Op_RegN]->OR(*idealreg2regmask[Op_RegF]); |
566 | idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]); |
567 | idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]); |
568 | idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegD]); |
569 | #else |
570 | idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegF]); |
571 | #ifdef ARM |
572 | // ARM has support for moving 64bit values between a pair of |
573 | // integer registers and a double register |
574 | idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]); |
575 | idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]); |
576 | #endif |
577 | #endif |
578 | } |
579 | |
580 | // Make up debug masks. Any spill slot plus callee-save registers. |
581 | // Caller-save registers are assumed to be trashable by the various |
582 | // inline-cache fixup routines. |
583 | *idealreg2debugmask [Op_RegN]= *idealreg2spillmask[Op_RegN]; |
584 | *idealreg2debugmask [Op_RegI]= *idealreg2spillmask[Op_RegI]; |
585 | *idealreg2debugmask [Op_RegL]= *idealreg2spillmask[Op_RegL]; |
586 | *idealreg2debugmask [Op_RegF]= *idealreg2spillmask[Op_RegF]; |
587 | *idealreg2debugmask [Op_RegD]= *idealreg2spillmask[Op_RegD]; |
588 | *idealreg2debugmask [Op_RegP]= *idealreg2spillmask[Op_RegP]; |
589 | |
590 | *idealreg2mhdebugmask[Op_RegN]= *idealreg2spillmask[Op_RegN]; |
591 | *idealreg2mhdebugmask[Op_RegI]= *idealreg2spillmask[Op_RegI]; |
592 | *idealreg2mhdebugmask[Op_RegL]= *idealreg2spillmask[Op_RegL]; |
593 | *idealreg2mhdebugmask[Op_RegF]= *idealreg2spillmask[Op_RegF]; |
594 | *idealreg2mhdebugmask[Op_RegD]= *idealreg2spillmask[Op_RegD]; |
595 | *idealreg2mhdebugmask[Op_RegP]= *idealreg2spillmask[Op_RegP]; |
596 | |
597 | // Prevent stub compilations from attempting to reference |
598 | // callee-saved registers from debug info |
599 | bool exclude_soe = !Compile::current()->is_method_compilation(); |
600 | |
601 | for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) { |
602 | // registers the caller has to save do not work |
603 | if( _register_save_policy[i] == 'C' || |
604 | _register_save_policy[i] == 'A' || |
605 | (_register_save_policy[i] == 'E' && exclude_soe) ) { |
606 | idealreg2debugmask [Op_RegN]->Remove(i); |
607 | idealreg2debugmask [Op_RegI]->Remove(i); // Exclude save-on-call |
608 | idealreg2debugmask [Op_RegL]->Remove(i); // registers from debug |
609 | idealreg2debugmask [Op_RegF]->Remove(i); // masks |
610 | idealreg2debugmask [Op_RegD]->Remove(i); |
611 | idealreg2debugmask [Op_RegP]->Remove(i); |
612 | |
613 | idealreg2mhdebugmask[Op_RegN]->Remove(i); |
614 | idealreg2mhdebugmask[Op_RegI]->Remove(i); |
615 | idealreg2mhdebugmask[Op_RegL]->Remove(i); |
616 | idealreg2mhdebugmask[Op_RegF]->Remove(i); |
617 | idealreg2mhdebugmask[Op_RegD]->Remove(i); |
618 | idealreg2mhdebugmask[Op_RegP]->Remove(i); |
619 | } |
620 | } |
621 | |
622 | // Subtract the register we use to save the SP for MethodHandle |
623 | // invokes to from the debug mask. |
624 | const RegMask save_mask = method_handle_invoke_SP_save_mask(); |
625 | idealreg2mhdebugmask[Op_RegN]->SUBTRACT(save_mask); |
626 | idealreg2mhdebugmask[Op_RegI]->SUBTRACT(save_mask); |
627 | idealreg2mhdebugmask[Op_RegL]->SUBTRACT(save_mask); |
628 | idealreg2mhdebugmask[Op_RegF]->SUBTRACT(save_mask); |
629 | idealreg2mhdebugmask[Op_RegD]->SUBTRACT(save_mask); |
630 | idealreg2mhdebugmask[Op_RegP]->SUBTRACT(save_mask); |
631 | } |
632 | |
633 | //---------------------------is_save_on_entry---------------------------------- |
634 | bool Matcher::is_save_on_entry( int reg ) { |
635 | return |
636 | _register_save_policy[reg] == 'E' || |
637 | _register_save_policy[reg] == 'A' || // Save-on-entry register? |
638 | // Also save argument registers in the trampolining stubs |
639 | (C->save_argument_registers() && is_spillable_arg(reg)); |
640 | } |
641 | |
642 | //---------------------------Fixup_Save_On_Entry------------------------------- |
643 | void Matcher::Fixup_Save_On_Entry( ) { |
644 | init_first_stack_mask(); |
645 | |
646 | Node *root = C->root(); // Short name for root |
647 | // Count number of save-on-entry registers. |
648 | uint soe_cnt = number_of_saved_registers(); |
649 | uint i; |
650 | |
651 | // Find the procedure Start Node |
652 | StartNode *start = C->start(); |
653 | assert( start, "Expect a start node" ); |
654 | |
655 | // Save argument registers in the trampolining stubs |
656 | if( C->save_argument_registers() ) |
657 | for( i = 0; i < _last_Mach_Reg; i++ ) |
658 | if( is_spillable_arg(i) ) |
659 | soe_cnt++; |
660 | |
661 | // Input RegMask array shared by all Returns. |
662 | // The type for doubles and longs has a count of 2, but |
663 | // there is only 1 returned value |
664 | uint ret_edge_cnt = TypeFunc::Parms + ((C->tf()->range()->cnt() == TypeFunc::Parms) ? 0 : 1); |
665 | RegMask *ret_rms = init_input_masks( ret_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); |
666 | // Returns have 0 or 1 returned values depending on call signature. |
667 | // Return register is specified by return_value in the AD file. |
668 | if (ret_edge_cnt > TypeFunc::Parms) |
669 | ret_rms[TypeFunc::Parms+0] = _return_value_mask; |
670 | |
671 | // Input RegMask array shared by all Rethrows. |
672 | uint reth_edge_cnt = TypeFunc::Parms+1; |
673 | RegMask *reth_rms = init_input_masks( reth_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); |
674 | // Rethrow takes exception oop only, but in the argument 0 slot. |
675 | OptoReg::Name reg = find_receiver(false); |
676 | if (reg >= 0) { |
677 | reth_rms[TypeFunc::Parms] = mreg2regmask[reg]; |
678 | #ifdef _LP64 |
679 | // Need two slots for ptrs in 64-bit land |
680 | reth_rms[TypeFunc::Parms].Insert(OptoReg::add(OptoReg::Name(reg), 1)); |
681 | #endif |
682 | } |
683 | |
684 | // Input RegMask array shared by all TailCalls |
685 | uint tail_call_edge_cnt = TypeFunc::Parms+2; |
686 | RegMask *tail_call_rms = init_input_masks( tail_call_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); |
687 | |
688 | // Input RegMask array shared by all TailJumps |
689 | uint tail_jump_edge_cnt = TypeFunc::Parms+2; |
690 | RegMask *tail_jump_rms = init_input_masks( tail_jump_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); |
691 | |
692 | // TailCalls have 2 returned values (target & moop), whose masks come |
693 | // from the usual MachNode/MachOper mechanism. Find a sample |
694 | // TailCall to extract these masks and put the correct masks into |
695 | // the tail_call_rms array. |
696 | for( i=1; i < root->req(); i++ ) { |
697 | MachReturnNode *m = root->in(i)->as_MachReturn(); |
698 | if( m->ideal_Opcode() == Op_TailCall ) { |
699 | tail_call_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0); |
700 | tail_call_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1); |
701 | break; |
702 | } |
703 | } |
704 | |
705 | // TailJumps have 2 returned values (target & ex_oop), whose masks come |
706 | // from the usual MachNode/MachOper mechanism. Find a sample |
707 | // TailJump to extract these masks and put the correct masks into |
708 | // the tail_jump_rms array. |
709 | for( i=1; i < root->req(); i++ ) { |
710 | MachReturnNode *m = root->in(i)->as_MachReturn(); |
711 | if( m->ideal_Opcode() == Op_TailJump ) { |
712 | tail_jump_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0); |
713 | tail_jump_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1); |
714 | break; |
715 | } |
716 | } |
717 | |
718 | // Input RegMask array shared by all Halts |
719 | uint halt_edge_cnt = TypeFunc::Parms; |
720 | RegMask *halt_rms = init_input_masks( halt_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); |
721 | |
722 | // Capture the return input masks into each exit flavor |
723 | for( i=1; i < root->req(); i++ ) { |
724 | MachReturnNode *exit = root->in(i)->as_MachReturn(); |
725 | switch( exit->ideal_Opcode() ) { |
726 | case Op_Return : exit->_in_rms = ret_rms; break; |
727 | case Op_Rethrow : exit->_in_rms = reth_rms; break; |
728 | case Op_TailCall : exit->_in_rms = tail_call_rms; break; |
729 | case Op_TailJump : exit->_in_rms = tail_jump_rms; break; |
730 | case Op_Halt : exit->_in_rms = halt_rms; break; |
731 | default : ShouldNotReachHere(); |
732 | } |
733 | } |
734 | |
735 | // Next unused projection number from Start. |
736 | int proj_cnt = C->tf()->domain()->cnt(); |
737 | |
738 | // Do all the save-on-entry registers. Make projections from Start for |
739 | // them, and give them a use at the exit points. To the allocator, they |
740 | // look like incoming register arguments. |
741 | for( i = 0; i < _last_Mach_Reg; i++ ) { |
742 | if( is_save_on_entry(i) ) { |
743 | |
744 | // Add the save-on-entry to the mask array |
745 | ret_rms [ ret_edge_cnt] = mreg2regmask[i]; |
746 | reth_rms [ reth_edge_cnt] = mreg2regmask[i]; |
747 | tail_call_rms[tail_call_edge_cnt] = mreg2regmask[i]; |
748 | tail_jump_rms[tail_jump_edge_cnt] = mreg2regmask[i]; |
749 | // Halts need the SOE registers, but only in the stack as debug info. |
750 | // A just-prior uncommon-trap or deoptimization will use the SOE regs. |
751 | halt_rms [ halt_edge_cnt] = *idealreg2spillmask[_register_save_type[i]]; |
752 | |
753 | Node *mproj; |
754 | |
755 | // Is this a RegF low half of a RegD? Double up 2 adjacent RegF's |
756 | // into a single RegD. |
757 | if( (i&1) == 0 && |
758 | _register_save_type[i ] == Op_RegF && |
759 | _register_save_type[i+1] == Op_RegF && |
760 | is_save_on_entry(i+1) ) { |
761 | // Add other bit for double |
762 | ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1)); |
763 | reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1)); |
764 | tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1)); |
765 | tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1)); |
766 | halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1)); |
767 | mproj = new MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegD ); |
768 | proj_cnt += 2; // Skip 2 for doubles |
769 | } |
770 | else if( (i&1) == 1 && // Else check for high half of double |
771 | _register_save_type[i-1] == Op_RegF && |
772 | _register_save_type[i ] == Op_RegF && |
773 | is_save_on_entry(i-1) ) { |
774 | ret_rms [ ret_edge_cnt] = RegMask::Empty; |
775 | reth_rms [ reth_edge_cnt] = RegMask::Empty; |
776 | tail_call_rms[tail_call_edge_cnt] = RegMask::Empty; |
777 | tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty; |
778 | halt_rms [ halt_edge_cnt] = RegMask::Empty; |
779 | mproj = C->top(); |
780 | } |
781 | // Is this a RegI low half of a RegL? Double up 2 adjacent RegI's |
782 | // into a single RegL. |
783 | else if( (i&1) == 0 && |
784 | _register_save_type[i ] == Op_RegI && |
785 | _register_save_type[i+1] == Op_RegI && |
786 | is_save_on_entry(i+1) ) { |
787 | // Add other bit for long |
788 | ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1)); |
789 | reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1)); |
790 | tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1)); |
791 | tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1)); |
792 | halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1)); |
793 | mproj = new MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegL ); |
794 | proj_cnt += 2; // Skip 2 for longs |
795 | } |
796 | else if( (i&1) == 1 && // Else check for high half of long |
797 | _register_save_type[i-1] == Op_RegI && |
798 | _register_save_type[i ] == Op_RegI && |
799 | is_save_on_entry(i-1) ) { |
800 | ret_rms [ ret_edge_cnt] = RegMask::Empty; |
801 | reth_rms [ reth_edge_cnt] = RegMask::Empty; |
802 | tail_call_rms[tail_call_edge_cnt] = RegMask::Empty; |
803 | tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty; |
804 | halt_rms [ halt_edge_cnt] = RegMask::Empty; |
805 | mproj = C->top(); |
806 | } else { |
807 | // Make a projection for it off the Start |
808 | mproj = new MachProjNode( start, proj_cnt++, ret_rms[ret_edge_cnt], _register_save_type[i] ); |
809 | } |
810 | |
811 | ret_edge_cnt ++; |
812 | reth_edge_cnt ++; |
813 | tail_call_edge_cnt ++; |
814 | tail_jump_edge_cnt ++; |
815 | halt_edge_cnt ++; |
816 | |
817 | // Add a use of the SOE register to all exit paths |
818 | for( uint j=1; j < root->req(); j++ ) |
819 | root->in(j)->add_req(mproj); |
820 | } // End of if a save-on-entry register |
821 | } // End of for all machine registers |
822 | } |
823 | |
824 | //------------------------------init_spill_mask-------------------------------- |
825 | void Matcher::init_spill_mask( Node *ret ) { |
826 | if( idealreg2regmask[Op_RegI] ) return; // One time only init |
827 | |
828 | OptoReg::c_frame_pointer = c_frame_pointer(); |
829 | c_frame_ptr_mask = c_frame_pointer(); |
830 | #ifdef _LP64 |
831 | // pointers are twice as big |
832 | c_frame_ptr_mask.Insert(OptoReg::add(c_frame_pointer(),1)); |
833 | #endif |
834 | |
835 | // Start at OptoReg::stack0() |
836 | STACK_ONLY_mask.Clear(); |
837 | OptoReg::Name init = OptoReg::stack2reg(0); |
838 | // STACK_ONLY_mask is all stack bits |
839 | OptoReg::Name i; |
840 | for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) |
841 | STACK_ONLY_mask.Insert(i); |
842 | // Also set the "infinite stack" bit. |
843 | STACK_ONLY_mask.set_AllStack(); |
844 | |
845 | // Copy the register names over into the shared world |
846 | for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) { |
847 | // SharedInfo::regName[i] = regName[i]; |
848 | // Handy RegMasks per machine register |
849 | mreg2regmask[i].Insert(i); |
850 | } |
851 | |
852 | // Grab the Frame Pointer |
853 | Node *fp = ret->in(TypeFunc::FramePtr); |
854 | Node *mem = ret->in(TypeFunc::Memory); |
855 | const TypePtr* atp = TypePtr::BOTTOM; |
856 | // Share frame pointer while making spill ops |
857 | set_shared(fp); |
858 | |
859 | // Compute generic short-offset Loads |
860 | #ifdef _LP64 |
861 | MachNode *spillCP = match_tree(new LoadNNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM,MemNode::unordered)); |
862 | #endif |
863 | MachNode *spillI = match_tree(new LoadINode(NULL,mem,fp,atp,TypeInt::INT,MemNode::unordered)); |
864 | MachNode *spillL = match_tree(new LoadLNode(NULL,mem,fp,atp,TypeLong::LONG,MemNode::unordered, LoadNode::DependsOnlyOnTest, false)); |
865 | MachNode *spillF = match_tree(new LoadFNode(NULL,mem,fp,atp,Type::FLOAT,MemNode::unordered)); |
866 | MachNode *spillD = match_tree(new LoadDNode(NULL,mem,fp,atp,Type::DOUBLE,MemNode::unordered)); |
867 | MachNode *spillP = match_tree(new LoadPNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM,MemNode::unordered)); |
868 | assert(spillI != NULL && spillL != NULL && spillF != NULL && |
869 | spillD != NULL && spillP != NULL, "" ); |
870 | // Get the ADLC notion of the right regmask, for each basic type. |
871 | #ifdef _LP64 |
872 | idealreg2regmask[Op_RegN] = &spillCP->out_RegMask(); |
873 | #endif |
874 | idealreg2regmask[Op_RegI] = &spillI->out_RegMask(); |
875 | idealreg2regmask[Op_RegL] = &spillL->out_RegMask(); |
876 | idealreg2regmask[Op_RegF] = &spillF->out_RegMask(); |
877 | idealreg2regmask[Op_RegD] = &spillD->out_RegMask(); |
878 | idealreg2regmask[Op_RegP] = &spillP->out_RegMask(); |
879 | |
880 | // Vector regmasks. |
881 | if (Matcher::vector_size_supported(T_BYTE,4)) { |
882 | TypeVect::VECTS = TypeVect::make(T_BYTE, 4); |
883 | MachNode *spillVectS = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTS)); |
884 | idealreg2regmask[Op_VecS] = &spillVectS->out_RegMask(); |
885 | } |
886 | if (Matcher::vector_size_supported(T_FLOAT,2)) { |
887 | MachNode *spillVectD = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTD)); |
888 | idealreg2regmask[Op_VecD] = &spillVectD->out_RegMask(); |
889 | } |
890 | if (Matcher::vector_size_supported(T_FLOAT,4)) { |
891 | MachNode *spillVectX = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTX)); |
892 | idealreg2regmask[Op_VecX] = &spillVectX->out_RegMask(); |
893 | } |
894 | if (Matcher::vector_size_supported(T_FLOAT,8)) { |
895 | MachNode *spillVectY = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTY)); |
896 | idealreg2regmask[Op_VecY] = &spillVectY->out_RegMask(); |
897 | } |
898 | if (Matcher::vector_size_supported(T_FLOAT,16)) { |
899 | MachNode *spillVectZ = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTZ)); |
900 | idealreg2regmask[Op_VecZ] = &spillVectZ->out_RegMask(); |
901 | } |
902 | } |
903 | |
904 | #ifdef ASSERT |
905 | static void match_alias_type(Compile* C, Node* n, Node* m) { |
906 | if (!VerifyAliases) return; // do not go looking for trouble by default |
907 | const TypePtr* nat = n->adr_type(); |
908 | const TypePtr* mat = m->adr_type(); |
909 | int nidx = C->get_alias_index(nat); |
910 | int midx = C->get_alias_index(mat); |
911 | // Detune the assert for cases like (AndI 0xFF (LoadB p)). |
912 | if (nidx == Compile::AliasIdxTop && midx >= Compile::AliasIdxRaw) { |
913 | for (uint i = 1; i < n->req(); i++) { |
914 | Node* n1 = n->in(i); |
915 | const TypePtr* n1at = n1->adr_type(); |
916 | if (n1at != NULL) { |
917 | nat = n1at; |
918 | nidx = C->get_alias_index(n1at); |
919 | } |
920 | } |
921 | } |
922 | // %%% Kludgery. Instead, fix ideal adr_type methods for all these cases: |
923 | if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxRaw) { |
924 | switch (n->Opcode()) { |
925 | case Op_PrefetchAllocation: |
926 | nidx = Compile::AliasIdxRaw; |
927 | nat = TypeRawPtr::BOTTOM; |
928 | break; |
929 | } |
930 | } |
931 | if (nidx == Compile::AliasIdxRaw && midx == Compile::AliasIdxTop) { |
932 | switch (n->Opcode()) { |
933 | case Op_ClearArray: |
934 | midx = Compile::AliasIdxRaw; |
935 | mat = TypeRawPtr::BOTTOM; |
936 | break; |
937 | } |
938 | } |
939 | if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxBot) { |
940 | switch (n->Opcode()) { |
941 | case Op_Return: |
942 | case Op_Rethrow: |
943 | case Op_Halt: |
944 | case Op_TailCall: |
945 | case Op_TailJump: |
946 | nidx = Compile::AliasIdxBot; |
947 | nat = TypePtr::BOTTOM; |
948 | break; |
949 | } |
950 | } |
951 | if (nidx == Compile::AliasIdxBot && midx == Compile::AliasIdxTop) { |
952 | switch (n->Opcode()) { |
953 | case Op_StrComp: |
954 | case Op_StrEquals: |
955 | case Op_StrIndexOf: |
956 | case Op_StrIndexOfChar: |
957 | case Op_AryEq: |
958 | case Op_HasNegatives: |
959 | case Op_MemBarVolatile: |
960 | case Op_MemBarCPUOrder: // %%% these ideals should have narrower adr_type? |
961 | case Op_StrInflatedCopy: |
962 | case Op_StrCompressedCopy: |
963 | case Op_OnSpinWait: |
964 | case Op_EncodeISOArray: |
965 | nidx = Compile::AliasIdxTop; |
966 | nat = NULL; |
967 | break; |
968 | } |
969 | } |
970 | if (nidx != midx) { |
971 | if (PrintOpto || (PrintMiscellaneous && (WizardMode || Verbose))) { |
972 | tty->print_cr("==== Matcher alias shift %d => %d" , nidx, midx); |
973 | n->dump(); |
974 | m->dump(); |
975 | } |
976 | assert(C->subsume_loads() && C->must_alias(nat, midx), |
977 | "must not lose alias info when matching" ); |
978 | } |
979 | } |
980 | #endif |
981 | |
982 | //------------------------------xform------------------------------------------ |
983 | // Given a Node in old-space, Match him (Label/Reduce) to produce a machine |
984 | // Node in new-space. Given a new-space Node, recursively walk his children. |
985 | Node *Matcher::transform( Node *n ) { ShouldNotCallThis(); return n; } |
986 | Node *Matcher::xform( Node *n, int max_stack ) { |
987 | // Use one stack to keep both: child's node/state and parent's node/index |
988 | MStack mstack(max_stack * 2 * 2); // usually: C->live_nodes() * 2 * 2 |
989 | mstack.push(n, Visit, NULL, -1); // set NULL as parent to indicate root |
990 | while (mstack.is_nonempty()) { |
991 | C->check_node_count(NodeLimitFudgeFactor, "too many nodes matching instructions" ); |
992 | if (C->failing()) return NULL; |
993 | n = mstack.node(); // Leave node on stack |
994 | Node_State nstate = mstack.state(); |
995 | if (nstate == Visit) { |
996 | mstack.set_state(Post_Visit); |
997 | Node *oldn = n; |
998 | // Old-space or new-space check |
999 | if (!C->node_arena()->contains(n)) { |
1000 | // Old space! |
1001 | Node* m; |
1002 | if (has_new_node(n)) { // Not yet Label/Reduced |
1003 | m = new_node(n); |
1004 | } else { |
1005 | if (!is_dontcare(n)) { // Matcher can match this guy |
1006 | // Calls match special. They match alone with no children. |
1007 | // Their children, the incoming arguments, match normally. |
1008 | m = n->is_SafePoint() ? match_sfpt(n->as_SafePoint()):match_tree(n); |
1009 | if (C->failing()) return NULL; |
1010 | if (m == NULL) { Matcher::soft_match_failure(); return NULL; } |
1011 | if (n->is_MemBar()) { |
1012 | m->as_MachMemBar()->set_adr_type(n->adr_type()); |
1013 | } |
1014 | } else { // Nothing the matcher cares about |
1015 | if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Multi()) { // Projections? |
1016 | // Convert to machine-dependent projection |
1017 | m = n->in(0)->as_Multi()->match( n->as_Proj(), this ); |
1018 | #ifdef ASSERT |
1019 | _new2old_map.map(m->_idx, n); |
1020 | #endif |
1021 | if (m->in(0) != NULL) // m might be top |
1022 | collect_null_checks(m, n); |
1023 | } else { // Else just a regular 'ol guy |
1024 | m = n->clone(); // So just clone into new-space |
1025 | #ifdef ASSERT |
1026 | _new2old_map.map(m->_idx, n); |
1027 | #endif |
1028 | // Def-Use edges will be added incrementally as Uses |
1029 | // of this node are matched. |
1030 | assert(m->outcnt() == 0, "no Uses of this clone yet" ); |
1031 | } |
1032 | } |
1033 | |
1034 | set_new_node(n, m); // Map old to new |
1035 | if (_old_node_note_array != NULL) { |
1036 | Node_Notes* nn = C->locate_node_notes(_old_node_note_array, |
1037 | n->_idx); |
1038 | C->set_node_notes_at(m->_idx, nn); |
1039 | } |
1040 | debug_only(match_alias_type(C, n, m)); |
1041 | } |
1042 | n = m; // n is now a new-space node |
1043 | mstack.set_node(n); |
1044 | } |
1045 | |
1046 | // New space! |
1047 | if (_visited.test_set(n->_idx)) continue; // while(mstack.is_nonempty()) |
1048 | |
1049 | int i; |
1050 | // Put precedence edges on stack first (match them last). |
1051 | for (i = oldn->req(); (uint)i < oldn->len(); i++) { |
1052 | Node *m = oldn->in(i); |
1053 | if (m == NULL) break; |
1054 | // set -1 to call add_prec() instead of set_req() during Step1 |
1055 | mstack.push(m, Visit, n, -1); |
1056 | } |
1057 | |
1058 | // Handle precedence edges for interior nodes |
1059 | for (i = n->len()-1; (uint)i >= n->req(); i--) { |
1060 | Node *m = n->in(i); |
1061 | if (m == NULL || C->node_arena()->contains(m)) continue; |
1062 | n->rm_prec(i); |
1063 | // set -1 to call add_prec() instead of set_req() during Step1 |
1064 | mstack.push(m, Visit, n, -1); |
1065 | } |
1066 | |
1067 | // For constant debug info, I'd rather have unmatched constants. |
1068 | int cnt = n->req(); |
1069 | JVMState* jvms = n->jvms(); |
1070 | int debug_cnt = jvms ? jvms->debug_start() : cnt; |
1071 | |
1072 | // Now do only debug info. Clone constants rather than matching. |
1073 | // Constants are represented directly in the debug info without |
1074 | // the need for executable machine instructions. |
1075 | // Monitor boxes are also represented directly. |
1076 | for (i = cnt - 1; i >= debug_cnt; --i) { // For all debug inputs do |
1077 | Node *m = n->in(i); // Get input |
1078 | int op = m->Opcode(); |
1079 | assert((op == Op_BoxLock) == jvms->is_monitor_use(i), "boxes only at monitor sites" ); |
1080 | if( op == Op_ConI || op == Op_ConP || op == Op_ConN || op == Op_ConNKlass || |
1081 | op == Op_ConF || op == Op_ConD || op == Op_ConL |
1082 | // || op == Op_BoxLock // %%%% enable this and remove (+++) in chaitin.cpp |
1083 | ) { |
1084 | m = m->clone(); |
1085 | #ifdef ASSERT |
1086 | _new2old_map.map(m->_idx, n); |
1087 | #endif |
1088 | mstack.push(m, Post_Visit, n, i); // Don't need to visit |
1089 | mstack.push(m->in(0), Visit, m, 0); |
1090 | } else { |
1091 | mstack.push(m, Visit, n, i); |
1092 | } |
1093 | } |
1094 | |
1095 | // And now walk his children, and convert his inputs to new-space. |
1096 | for( ; i >= 0; --i ) { // For all normal inputs do |
1097 | Node *m = n->in(i); // Get input |
1098 | if(m != NULL) |
1099 | mstack.push(m, Visit, n, i); |
1100 | } |
1101 | |
1102 | } |
1103 | else if (nstate == Post_Visit) { |
1104 | // Set xformed input |
1105 | Node *p = mstack.parent(); |
1106 | if (p != NULL) { // root doesn't have parent |
1107 | int i = (int)mstack.index(); |
1108 | if (i >= 0) |
1109 | p->set_req(i, n); // required input |
1110 | else if (i == -1) |
1111 | p->add_prec(n); // precedence input |
1112 | else |
1113 | ShouldNotReachHere(); |
1114 | } |
1115 | mstack.pop(); // remove processed node from stack |
1116 | } |
1117 | else { |
1118 | ShouldNotReachHere(); |
1119 | } |
1120 | } // while (mstack.is_nonempty()) |
1121 | return n; // Return new-space Node |
1122 | } |
1123 | |
1124 | //------------------------------warp_outgoing_stk_arg------------------------ |
1125 | OptoReg::Name Matcher::warp_outgoing_stk_arg( VMReg reg, OptoReg::Name begin_out_arg_area, OptoReg::Name &out_arg_limit_per_call ) { |
1126 | // Convert outgoing argument location to a pre-biased stack offset |
1127 | if (reg->is_stack()) { |
1128 | OptoReg::Name warped = reg->reg2stack(); |
1129 | // Adjust the stack slot offset to be the register number used |
1130 | // by the allocator. |
1131 | warped = OptoReg::add(begin_out_arg_area, warped); |
1132 | // Keep track of the largest numbered stack slot used for an arg. |
1133 | // Largest used slot per call-site indicates the amount of stack |
1134 | // that is killed by the call. |
1135 | if( warped >= out_arg_limit_per_call ) |
1136 | out_arg_limit_per_call = OptoReg::add(warped,1); |
1137 | if (!RegMask::can_represent_arg(warped)) { |
1138 | C->record_method_not_compilable("unsupported calling sequence" ); |
1139 | return OptoReg::Bad; |
1140 | } |
1141 | return warped; |
1142 | } |
1143 | return OptoReg::as_OptoReg(reg); |
1144 | } |
1145 | |
1146 | |
1147 | //------------------------------match_sfpt------------------------------------- |
1148 | // Helper function to match call instructions. Calls match special. |
1149 | // They match alone with no children. Their children, the incoming |
1150 | // arguments, match normally. |
1151 | MachNode *Matcher::match_sfpt( SafePointNode *sfpt ) { |
1152 | MachSafePointNode *msfpt = NULL; |
1153 | MachCallNode *mcall = NULL; |
1154 | uint cnt; |
1155 | // Split out case for SafePoint vs Call |
1156 | CallNode *call; |
1157 | const TypeTuple *domain; |
1158 | ciMethod* method = NULL; |
1159 | bool is_method_handle_invoke = false; // for special kill effects |
1160 | if( sfpt->is_Call() ) { |
1161 | call = sfpt->as_Call(); |
1162 | domain = call->tf()->domain(); |
1163 | cnt = domain->cnt(); |
1164 | |
1165 | // Match just the call, nothing else |
1166 | MachNode *m = match_tree(call); |
1167 | if (C->failing()) return NULL; |
1168 | if( m == NULL ) { Matcher::soft_match_failure(); return NULL; } |
1169 | |
1170 | // Copy data from the Ideal SafePoint to the machine version |
1171 | mcall = m->as_MachCall(); |
1172 | |
1173 | mcall->set_tf( call->tf()); |
1174 | mcall->set_entry_point(call->entry_point()); |
1175 | mcall->set_cnt( call->cnt()); |
1176 | |
1177 | if( mcall->is_MachCallJava() ) { |
1178 | MachCallJavaNode *mcall_java = mcall->as_MachCallJava(); |
1179 | const CallJavaNode *call_java = call->as_CallJava(); |
1180 | assert(call_java->validate_symbolic_info(), "inconsistent info" ); |
1181 | method = call_java->method(); |
1182 | mcall_java->_method = method; |
1183 | mcall_java->_bci = call_java->_bci; |
1184 | mcall_java->_optimized_virtual = call_java->is_optimized_virtual(); |
1185 | is_method_handle_invoke = call_java->is_method_handle_invoke(); |
1186 | mcall_java->_method_handle_invoke = is_method_handle_invoke; |
1187 | mcall_java->_override_symbolic_info = call_java->override_symbolic_info(); |
1188 | if (is_method_handle_invoke) { |
1189 | C->set_has_method_handle_invokes(true); |
1190 | } |
1191 | if( mcall_java->is_MachCallStaticJava() ) |
1192 | mcall_java->as_MachCallStaticJava()->_name = |
1193 | call_java->as_CallStaticJava()->_name; |
1194 | if( mcall_java->is_MachCallDynamicJava() ) |
1195 | mcall_java->as_MachCallDynamicJava()->_vtable_index = |
1196 | call_java->as_CallDynamicJava()->_vtable_index; |
1197 | } |
1198 | else if( mcall->is_MachCallRuntime() ) { |
1199 | mcall->as_MachCallRuntime()->_name = call->as_CallRuntime()->_name; |
1200 | } |
1201 | msfpt = mcall; |
1202 | } |
1203 | // This is a non-call safepoint |
1204 | else { |
1205 | call = NULL; |
1206 | domain = NULL; |
1207 | MachNode *mn = match_tree(sfpt); |
1208 | if (C->failing()) return NULL; |
1209 | msfpt = mn->as_MachSafePoint(); |
1210 | cnt = TypeFunc::Parms; |
1211 | } |
1212 | |
1213 | // Advertise the correct memory effects (for anti-dependence computation). |
1214 | msfpt->set_adr_type(sfpt->adr_type()); |
1215 | |
1216 | // Allocate a private array of RegMasks. These RegMasks are not shared. |
1217 | msfpt->_in_rms = NEW_RESOURCE_ARRAY( RegMask, cnt ); |
1218 | // Empty them all. |
1219 | for (uint i = 0; i < cnt; i++) ::new (&(msfpt->_in_rms[i])) RegMask(); |
1220 | |
1221 | // Do all the pre-defined non-Empty register masks |
1222 | msfpt->_in_rms[TypeFunc::ReturnAdr] = _return_addr_mask; |
1223 | msfpt->_in_rms[TypeFunc::FramePtr ] = c_frame_ptr_mask; |
1224 | |
1225 | // Place first outgoing argument can possibly be put. |
1226 | OptoReg::Name begin_out_arg_area = OptoReg::add(_new_SP, C->out_preserve_stack_slots()); |
1227 | assert( is_even(begin_out_arg_area), "" ); |
1228 | // Compute max outgoing register number per call site. |
1229 | OptoReg::Name out_arg_limit_per_call = begin_out_arg_area; |
1230 | // Calls to C may hammer extra stack slots above and beyond any arguments. |
1231 | // These are usually backing store for register arguments for varargs. |
1232 | if( call != NULL && call->is_CallRuntime() ) |
1233 | out_arg_limit_per_call = OptoReg::add(out_arg_limit_per_call,C->varargs_C_out_slots_killed()); |
1234 | |
1235 | |
1236 | // Do the normal argument list (parameters) register masks |
1237 | int argcnt = cnt - TypeFunc::Parms; |
1238 | if( argcnt > 0 ) { // Skip it all if we have no args |
1239 | BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt ); |
1240 | VMRegPair *parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt ); |
1241 | int i; |
1242 | for( i = 0; i < argcnt; i++ ) { |
1243 | sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type(); |
1244 | } |
1245 | // V-call to pick proper calling convention |
1246 | call->calling_convention( sig_bt, parm_regs, argcnt ); |
1247 | |
1248 | #ifdef ASSERT |
1249 | // Sanity check users' calling convention. Really handy during |
1250 | // the initial porting effort. Fairly expensive otherwise. |
1251 | { for (int i = 0; i<argcnt; i++) { |
1252 | if( !parm_regs[i].first()->is_valid() && |
1253 | !parm_regs[i].second()->is_valid() ) continue; |
1254 | VMReg reg1 = parm_regs[i].first(); |
1255 | VMReg reg2 = parm_regs[i].second(); |
1256 | for (int j = 0; j < i; j++) { |
1257 | if( !parm_regs[j].first()->is_valid() && |
1258 | !parm_regs[j].second()->is_valid() ) continue; |
1259 | VMReg reg3 = parm_regs[j].first(); |
1260 | VMReg reg4 = parm_regs[j].second(); |
1261 | if( !reg1->is_valid() ) { |
1262 | assert( !reg2->is_valid(), "valid halvsies" ); |
1263 | } else if( !reg3->is_valid() ) { |
1264 | assert( !reg4->is_valid(), "valid halvsies" ); |
1265 | } else { |
1266 | assert( reg1 != reg2, "calling conv. must produce distinct regs" ); |
1267 | assert( reg1 != reg3, "calling conv. must produce distinct regs" ); |
1268 | assert( reg1 != reg4, "calling conv. must produce distinct regs" ); |
1269 | assert( reg2 != reg3, "calling conv. must produce distinct regs" ); |
1270 | assert( reg2 != reg4 || !reg2->is_valid(), "calling conv. must produce distinct regs" ); |
1271 | assert( reg3 != reg4, "calling conv. must produce distinct regs" ); |
1272 | } |
1273 | } |
1274 | } |
1275 | } |
1276 | #endif |
1277 | |
1278 | // Visit each argument. Compute its outgoing register mask. |
1279 | // Return results now can have 2 bits returned. |
1280 | // Compute max over all outgoing arguments both per call-site |
1281 | // and over the entire method. |
1282 | for( i = 0; i < argcnt; i++ ) { |
1283 | // Address of incoming argument mask to fill in |
1284 | RegMask *rm = &mcall->_in_rms[i+TypeFunc::Parms]; |
1285 | if( !parm_regs[i].first()->is_valid() && |
1286 | !parm_regs[i].second()->is_valid() ) { |
1287 | continue; // Avoid Halves |
1288 | } |
1289 | // Grab first register, adjust stack slots and insert in mask. |
1290 | OptoReg::Name reg1 = warp_outgoing_stk_arg(parm_regs[i].first(), begin_out_arg_area, out_arg_limit_per_call ); |
1291 | if (OptoReg::is_valid(reg1)) |
1292 | rm->Insert( reg1 ); |
1293 | // Grab second register (if any), adjust stack slots and insert in mask. |
1294 | OptoReg::Name reg2 = warp_outgoing_stk_arg(parm_regs[i].second(), begin_out_arg_area, out_arg_limit_per_call ); |
1295 | if (OptoReg::is_valid(reg2)) |
1296 | rm->Insert( reg2 ); |
1297 | } // End of for all arguments |
1298 | |
1299 | // Compute number of stack slots needed to restore stack in case of |
1300 | // Pascal-style argument popping. |
1301 | mcall->_argsize = out_arg_limit_per_call - begin_out_arg_area; |
1302 | } |
1303 | |
1304 | // Compute the max stack slot killed by any call. These will not be |
1305 | // available for debug info, and will be used to adjust FIRST_STACK_mask |
1306 | // after all call sites have been visited. |
1307 | if( _out_arg_limit < out_arg_limit_per_call) |
1308 | _out_arg_limit = out_arg_limit_per_call; |
1309 | |
1310 | if (mcall) { |
1311 | // Kill the outgoing argument area, including any non-argument holes and |
1312 | // any legacy C-killed slots. Use Fat-Projections to do the killing. |
1313 | // Since the max-per-method covers the max-per-call-site and debug info |
1314 | // is excluded on the max-per-method basis, debug info cannot land in |
1315 | // this killed area. |
1316 | uint r_cnt = mcall->tf()->range()->cnt(); |
1317 | MachProjNode *proj = new MachProjNode( mcall, r_cnt+10000, RegMask::Empty, MachProjNode::fat_proj ); |
1318 | if (!RegMask::can_represent_arg(OptoReg::Name(out_arg_limit_per_call-1))) { |
1319 | C->record_method_not_compilable("unsupported outgoing calling sequence" ); |
1320 | } else { |
1321 | for (int i = begin_out_arg_area; i < out_arg_limit_per_call; i++) |
1322 | proj->_rout.Insert(OptoReg::Name(i)); |
1323 | } |
1324 | if (proj->_rout.is_NotEmpty()) { |
1325 | push_projection(proj); |
1326 | } |
1327 | } |
1328 | // Transfer the safepoint information from the call to the mcall |
1329 | // Move the JVMState list |
1330 | msfpt->set_jvms(sfpt->jvms()); |
1331 | for (JVMState* jvms = msfpt->jvms(); jvms; jvms = jvms->caller()) { |
1332 | jvms->set_map(sfpt); |
1333 | } |
1334 | |
1335 | // Debug inputs begin just after the last incoming parameter |
1336 | assert((mcall == NULL) || (mcall->jvms() == NULL) || |
1337 | (mcall->jvms()->debug_start() + mcall->_jvmadj == mcall->tf()->domain()->cnt()), "" ); |
1338 | |
1339 | // Move the OopMap |
1340 | msfpt->_oop_map = sfpt->_oop_map; |
1341 | |
1342 | // Add additional edges. |
1343 | if (msfpt->mach_constant_base_node_input() != (uint)-1 && !msfpt->is_MachCallLeaf()) { |
1344 | // For these calls we can not add MachConstantBase in expand(), as the |
1345 | // ins are not complete then. |
1346 | msfpt->ins_req(msfpt->mach_constant_base_node_input(), C->mach_constant_base_node()); |
1347 | if (msfpt->jvms() && |
1348 | msfpt->mach_constant_base_node_input() <= msfpt->jvms()->debug_start() + msfpt->_jvmadj) { |
1349 | // We added an edge before jvms, so we must adapt the position of the ins. |
1350 | msfpt->jvms()->adapt_position(+1); |
1351 | } |
1352 | } |
1353 | |
1354 | // Registers killed by the call are set in the local scheduling pass |
1355 | // of Global Code Motion. |
1356 | return msfpt; |
1357 | } |
1358 | |
1359 | //---------------------------match_tree---------------------------------------- |
1360 | // Match a Ideal Node DAG - turn it into a tree; Label & Reduce. Used as part |
1361 | // of the whole-sale conversion from Ideal to Mach Nodes. Also used for |
1362 | // making GotoNodes while building the CFG and in init_spill_mask() to identify |
1363 | // a Load's result RegMask for memoization in idealreg2regmask[] |
1364 | MachNode *Matcher::match_tree( const Node *n ) { |
1365 | assert( n->Opcode() != Op_Phi, "cannot match" ); |
1366 | assert( !n->is_block_start(), "cannot match" ); |
1367 | // Set the mark for all locally allocated State objects. |
1368 | // When this call returns, the _states_arena arena will be reset |
1369 | // freeing all State objects. |
1370 | ResourceMark rm( &_states_arena ); |
1371 | |
1372 | LabelRootDepth = 0; |
1373 | |
1374 | // StoreNodes require their Memory input to match any LoadNodes |
1375 | Node *mem = n->is_Store() ? n->in(MemNode::Memory) : (Node*)1 ; |
1376 | #ifdef ASSERT |
1377 | Node* save_mem_node = _mem_node; |
1378 | _mem_node = n->is_Store() ? (Node*)n : NULL; |
1379 | #endif |
1380 | // State object for root node of match tree |
1381 | // Allocate it on _states_arena - stack allocation can cause stack overflow. |
1382 | State *s = new (&_states_arena) State; |
1383 | s->_kids[0] = NULL; |
1384 | s->_kids[1] = NULL; |
1385 | s->_leaf = (Node*)n; |
1386 | // Label the input tree, allocating labels from top-level arena |
1387 | Label_Root( n, s, n->in(0), mem ); |
1388 | if (C->failing()) return NULL; |
1389 | |
1390 | // The minimum cost match for the whole tree is found at the root State |
1391 | uint mincost = max_juint; |
1392 | uint cost = max_juint; |
1393 | uint i; |
1394 | for( i = 0; i < NUM_OPERANDS; i++ ) { |
1395 | if( s->valid(i) && // valid entry and |
1396 | s->_cost[i] < cost && // low cost and |
1397 | s->_rule[i] >= NUM_OPERANDS ) // not an operand |
1398 | cost = s->_cost[mincost=i]; |
1399 | } |
1400 | if (mincost == max_juint) { |
1401 | #ifndef PRODUCT |
1402 | tty->print("No matching rule for:" ); |
1403 | s->dump(); |
1404 | #endif |
1405 | Matcher::soft_match_failure(); |
1406 | return NULL; |
1407 | } |
1408 | // Reduce input tree based upon the state labels to machine Nodes |
1409 | MachNode *m = ReduceInst( s, s->_rule[mincost], mem ); |
1410 | #ifdef ASSERT |
1411 | _old2new_map.map(n->_idx, m); |
1412 | _new2old_map.map(m->_idx, (Node*)n); |
1413 | #endif |
1414 | |
1415 | // Add any Matcher-ignored edges |
1416 | uint cnt = n->req(); |
1417 | uint start = 1; |
1418 | if( mem != (Node*)1 ) start = MemNode::Memory+1; |
1419 | if( n->is_AddP() ) { |
1420 | assert( mem == (Node*)1, "" ); |
1421 | start = AddPNode::Base+1; |
1422 | } |
1423 | for( i = start; i < cnt; i++ ) { |
1424 | if( !n->match_edge(i) ) { |
1425 | if( i < m->req() ) |
1426 | m->ins_req( i, n->in(i) ); |
1427 | else |
1428 | m->add_req( n->in(i) ); |
1429 | } |
1430 | } |
1431 | |
1432 | debug_only( _mem_node = save_mem_node; ) |
1433 | return m; |
1434 | } |
1435 | |
1436 | |
1437 | //------------------------------match_into_reg--------------------------------- |
1438 | // Choose to either match this Node in a register or part of the current |
1439 | // match tree. Return true for requiring a register and false for matching |
1440 | // as part of the current match tree. |
1441 | static bool match_into_reg( const Node *n, Node *m, Node *control, int i, bool shared ) { |
1442 | |
1443 | const Type *t = m->bottom_type(); |
1444 | |
1445 | if (t->singleton()) { |
1446 | // Never force constants into registers. Allow them to match as |
1447 | // constants or registers. Copies of the same value will share |
1448 | // the same register. See find_shared_node. |
1449 | return false; |
1450 | } else { // Not a constant |
1451 | // Stop recursion if they have different Controls. |
1452 | Node* m_control = m->in(0); |
1453 | // Control of load's memory can post-dominates load's control. |
1454 | // So use it since load can't float above its memory. |
1455 | Node* mem_control = (m->is_Load()) ? m->in(MemNode::Memory)->in(0) : NULL; |
1456 | if (control && m_control && control != m_control && control != mem_control) { |
1457 | |
1458 | // Actually, we can live with the most conservative control we |
1459 | // find, if it post-dominates the others. This allows us to |
1460 | // pick up load/op/store trees where the load can float a little |
1461 | // above the store. |
1462 | Node *x = control; |
1463 | const uint max_scan = 6; // Arbitrary scan cutoff |
1464 | uint j; |
1465 | for (j=0; j<max_scan; j++) { |
1466 | if (x->is_Region()) // Bail out at merge points |
1467 | return true; |
1468 | x = x->in(0); |
1469 | if (x == m_control) // Does 'control' post-dominate |
1470 | break; // m->in(0)? If so, we can use it |
1471 | if (x == mem_control) // Does 'control' post-dominate |
1472 | break; // mem_control? If so, we can use it |
1473 | } |
1474 | if (j == max_scan) // No post-domination before scan end? |
1475 | return true; // Then break the match tree up |
1476 | } |
1477 | if ((m->is_DecodeN() && Matcher::narrow_oop_use_complex_address()) || |
1478 | (m->is_DecodeNKlass() && Matcher::narrow_klass_use_complex_address())) { |
1479 | // These are commonly used in address expressions and can |
1480 | // efficiently fold into them on X64 in some cases. |
1481 | return false; |
1482 | } |
1483 | } |
1484 | |
1485 | // Not forceable cloning. If shared, put it into a register. |
1486 | return shared; |
1487 | } |
1488 | |
1489 | |
1490 | //------------------------------Instruction Selection-------------------------- |
1491 | // Label method walks a "tree" of nodes, using the ADLC generated DFA to match |
1492 | // ideal nodes to machine instructions. Trees are delimited by shared Nodes, |
1493 | // things the Matcher does not match (e.g., Memory), and things with different |
1494 | // Controls (hence forced into different blocks). We pass in the Control |
1495 | // selected for this entire State tree. |
1496 | |
1497 | // The Matcher works on Trees, but an Intel add-to-memory requires a DAG: the |
1498 | // Store and the Load must have identical Memories (as well as identical |
1499 | // pointers). Since the Matcher does not have anything for Memory (and |
1500 | // does not handle DAGs), I have to match the Memory input myself. If the |
1501 | // Tree root is a Store, I require all Loads to have the identical memory. |
1502 | Node *Matcher::Label_Root( const Node *n, State *svec, Node *control, const Node *mem){ |
1503 | // Since Label_Root is a recursive function, its possible that we might run |
1504 | // out of stack space. See bugs 6272980 & 6227033 for more info. |
1505 | LabelRootDepth++; |
1506 | if (LabelRootDepth > MaxLabelRootDepth) { |
1507 | C->record_method_not_compilable("Out of stack space, increase MaxLabelRootDepth" ); |
1508 | return NULL; |
1509 | } |
1510 | uint care = 0; // Edges matcher cares about |
1511 | uint cnt = n->req(); |
1512 | uint i = 0; |
1513 | |
1514 | // Examine children for memory state |
1515 | // Can only subsume a child into your match-tree if that child's memory state |
1516 | // is not modified along the path to another input. |
1517 | // It is unsafe even if the other inputs are separate roots. |
1518 | Node *input_mem = NULL; |
1519 | for( i = 1; i < cnt; i++ ) { |
1520 | if( !n->match_edge(i) ) continue; |
1521 | Node *m = n->in(i); // Get ith input |
1522 | assert( m, "expect non-null children" ); |
1523 | if( m->is_Load() ) { |
1524 | if( input_mem == NULL ) { |
1525 | input_mem = m->in(MemNode::Memory); |
1526 | } else if( input_mem != m->in(MemNode::Memory) ) { |
1527 | input_mem = NodeSentinel; |
1528 | } |
1529 | } |
1530 | } |
1531 | |
1532 | for( i = 1; i < cnt; i++ ){// For my children |
1533 | if( !n->match_edge(i) ) continue; |
1534 | Node *m = n->in(i); // Get ith input |
1535 | // Allocate states out of a private arena |
1536 | State *s = new (&_states_arena) State; |
1537 | svec->_kids[care++] = s; |
1538 | assert( care <= 2, "binary only for now" ); |
1539 | |
1540 | // Recursively label the State tree. |
1541 | s->_kids[0] = NULL; |
1542 | s->_kids[1] = NULL; |
1543 | s->_leaf = m; |
1544 | |
1545 | // Check for leaves of the State Tree; things that cannot be a part of |
1546 | // the current tree. If it finds any, that value is matched as a |
1547 | // register operand. If not, then the normal matching is used. |
1548 | if( match_into_reg(n, m, control, i, is_shared(m)) || |
1549 | // |
1550 | // Stop recursion if this is LoadNode and the root of this tree is a |
1551 | // StoreNode and the load & store have different memories. |
1552 | ((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) || |
1553 | // Can NOT include the match of a subtree when its memory state |
1554 | // is used by any of the other subtrees |
1555 | (input_mem == NodeSentinel) ) { |
1556 | // Print when we exclude matching due to different memory states at input-loads |
1557 | if (PrintOpto && (Verbose && WizardMode) && (input_mem == NodeSentinel) |
1558 | && !((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem)) { |
1559 | tty->print_cr("invalid input_mem" ); |
1560 | } |
1561 | // Switch to a register-only opcode; this value must be in a register |
1562 | // and cannot be subsumed as part of a larger instruction. |
1563 | s->DFA( m->ideal_reg(), m ); |
1564 | |
1565 | } else { |
1566 | // If match tree has no control and we do, adopt it for entire tree |
1567 | if( control == NULL && m->in(0) != NULL && m->req() > 1 ) |
1568 | control = m->in(0); // Pick up control |
1569 | // Else match as a normal part of the match tree. |
1570 | control = Label_Root(m,s,control,mem); |
1571 | if (C->failing()) return NULL; |
1572 | } |
1573 | } |
1574 | |
1575 | |
1576 | // Call DFA to match this node, and return |
1577 | svec->DFA( n->Opcode(), n ); |
1578 | |
1579 | #ifdef ASSERT |
1580 | uint x; |
1581 | for( x = 0; x < _LAST_MACH_OPER; x++ ) |
1582 | if( svec->valid(x) ) |
1583 | break; |
1584 | |
1585 | if (x >= _LAST_MACH_OPER) { |
1586 | n->dump(); |
1587 | svec->dump(); |
1588 | assert( false, "bad AD file" ); |
1589 | } |
1590 | #endif |
1591 | return control; |
1592 | } |
1593 | |
1594 | |
1595 | // Con nodes reduced using the same rule can share their MachNode |
1596 | // which reduces the number of copies of a constant in the final |
1597 | // program. The register allocator is free to split uses later to |
1598 | // split live ranges. |
1599 | MachNode* Matcher::find_shared_node(Node* leaf, uint rule) { |
1600 | if (!leaf->is_Con() && !leaf->is_DecodeNarrowPtr()) return NULL; |
1601 | |
1602 | // See if this Con has already been reduced using this rule. |
1603 | if (_shared_nodes.Size() <= leaf->_idx) return NULL; |
1604 | MachNode* last = (MachNode*)_shared_nodes.at(leaf->_idx); |
1605 | if (last != NULL && rule == last->rule()) { |
1606 | // Don't expect control change for DecodeN |
1607 | if (leaf->is_DecodeNarrowPtr()) |
1608 | return last; |
1609 | // Get the new space root. |
1610 | Node* xroot = new_node(C->root()); |
1611 | if (xroot == NULL) { |
1612 | // This shouldn't happen give the order of matching. |
1613 | return NULL; |
1614 | } |
1615 | |
1616 | // Shared constants need to have their control be root so they |
1617 | // can be scheduled properly. |
1618 | Node* control = last->in(0); |
1619 | if (control != xroot) { |
1620 | if (control == NULL || control == C->root()) { |
1621 | last->set_req(0, xroot); |
1622 | } else { |
1623 | assert(false, "unexpected control" ); |
1624 | return NULL; |
1625 | } |
1626 | } |
1627 | return last; |
1628 | } |
1629 | return NULL; |
1630 | } |
1631 | |
1632 | |
1633 | //------------------------------ReduceInst------------------------------------- |
1634 | // Reduce a State tree (with given Control) into a tree of MachNodes. |
1635 | // This routine (and it's cohort ReduceOper) convert Ideal Nodes into |
1636 | // complicated machine Nodes. Each MachNode covers some tree of Ideal Nodes. |
1637 | // Each MachNode has a number of complicated MachOper operands; each |
1638 | // MachOper also covers a further tree of Ideal Nodes. |
1639 | |
1640 | // The root of the Ideal match tree is always an instruction, so we enter |
1641 | // the recursion here. After building the MachNode, we need to recurse |
1642 | // the tree checking for these cases: |
1643 | // (1) Child is an instruction - |
1644 | // Build the instruction (recursively), add it as an edge. |
1645 | // Build a simple operand (register) to hold the result of the instruction. |
1646 | // (2) Child is an interior part of an instruction - |
1647 | // Skip over it (do nothing) |
1648 | // (3) Child is the start of a operand - |
1649 | // Build the operand, place it inside the instruction |
1650 | // Call ReduceOper. |
1651 | MachNode *Matcher::ReduceInst( State *s, int rule, Node *&mem ) { |
1652 | assert( rule >= NUM_OPERANDS, "called with operand rule" ); |
1653 | |
1654 | MachNode* shared_node = find_shared_node(s->_leaf, rule); |
1655 | if (shared_node != NULL) { |
1656 | return shared_node; |
1657 | } |
1658 | |
1659 | // Build the object to represent this state & prepare for recursive calls |
1660 | MachNode *mach = s->MachNodeGenerator(rule); |
1661 | guarantee(mach != NULL, "Missing MachNode" ); |
1662 | mach->_opnds[0] = s->MachOperGenerator(_reduceOp[rule]); |
1663 | assert( mach->_opnds[0] != NULL, "Missing result operand" ); |
1664 | Node *leaf = s->_leaf; |
1665 | // Check for instruction or instruction chain rule |
1666 | if( rule >= _END_INST_CHAIN_RULE || rule < _BEGIN_INST_CHAIN_RULE ) { |
1667 | assert(C->node_arena()->contains(s->_leaf) || !has_new_node(s->_leaf), |
1668 | "duplicating node that's already been matched" ); |
1669 | // Instruction |
1670 | mach->add_req( leaf->in(0) ); // Set initial control |
1671 | // Reduce interior of complex instruction |
1672 | ReduceInst_Interior( s, rule, mem, mach, 1 ); |
1673 | } else { |
1674 | // Instruction chain rules are data-dependent on their inputs |
1675 | mach->add_req(0); // Set initial control to none |
1676 | ReduceInst_Chain_Rule( s, rule, mem, mach ); |
1677 | } |
1678 | |
1679 | // If a Memory was used, insert a Memory edge |
1680 | if( mem != (Node*)1 ) { |
1681 | mach->ins_req(MemNode::Memory,mem); |
1682 | #ifdef ASSERT |
1683 | // Verify adr type after matching memory operation |
1684 | const MachOper* oper = mach->memory_operand(); |
1685 | if (oper != NULL && oper != (MachOper*)-1) { |
1686 | // It has a unique memory operand. Find corresponding ideal mem node. |
1687 | Node* m = NULL; |
1688 | if (leaf->is_Mem()) { |
1689 | m = leaf; |
1690 | } else { |
1691 | m = _mem_node; |
1692 | assert(m != NULL && m->is_Mem(), "expecting memory node" ); |
1693 | } |
1694 | const Type* mach_at = mach->adr_type(); |
1695 | // DecodeN node consumed by an address may have different type |
1696 | // than its input. Don't compare types for such case. |
1697 | if (m->adr_type() != mach_at && |
1698 | (m->in(MemNode::Address)->is_DecodeNarrowPtr() || |
1699 | (m->in(MemNode::Address)->is_AddP() && |
1700 | m->in(MemNode::Address)->in(AddPNode::Address)->is_DecodeNarrowPtr()) || |
1701 | (m->in(MemNode::Address)->is_AddP() && |
1702 | m->in(MemNode::Address)->in(AddPNode::Address)->is_AddP() && |
1703 | m->in(MemNode::Address)->in(AddPNode::Address)->in(AddPNode::Address)->is_DecodeNarrowPtr()))) { |
1704 | mach_at = m->adr_type(); |
1705 | } |
1706 | if (m->adr_type() != mach_at) { |
1707 | m->dump(); |
1708 | tty->print_cr("mach:" ); |
1709 | mach->dump(1); |
1710 | } |
1711 | assert(m->adr_type() == mach_at, "matcher should not change adr type" ); |
1712 | } |
1713 | #endif |
1714 | } |
1715 | |
1716 | // If the _leaf is an AddP, insert the base edge |
1717 | if (leaf->is_AddP()) { |
1718 | mach->ins_req(AddPNode::Base,leaf->in(AddPNode::Base)); |
1719 | } |
1720 | |
1721 | uint number_of_projections_prior = number_of_projections(); |
1722 | |
1723 | // Perform any 1-to-many expansions required |
1724 | MachNode *ex = mach->Expand(s, _projection_list, mem); |
1725 | if (ex != mach) { |
1726 | assert(ex->ideal_reg() == mach->ideal_reg(), "ideal types should match" ); |
1727 | if( ex->in(1)->is_Con() ) |
1728 | ex->in(1)->set_req(0, C->root()); |
1729 | // Remove old node from the graph |
1730 | for( uint i=0; i<mach->req(); i++ ) { |
1731 | mach->set_req(i,NULL); |
1732 | } |
1733 | #ifdef ASSERT |
1734 | _new2old_map.map(ex->_idx, s->_leaf); |
1735 | #endif |
1736 | } |
1737 | |
1738 | // PhaseChaitin::fixup_spills will sometimes generate spill code |
1739 | // via the matcher. By the time, nodes have been wired into the CFG, |
1740 | // and any further nodes generated by expand rules will be left hanging |
1741 | // in space, and will not get emitted as output code. Catch this. |
1742 | // Also, catch any new register allocation constraints ("projections") |
1743 | // generated belatedly during spill code generation. |
1744 | if (_allocation_started) { |
1745 | guarantee(ex == mach, "no expand rules during spill generation" ); |
1746 | guarantee(number_of_projections_prior == number_of_projections(), "no allocation during spill generation" ); |
1747 | } |
1748 | |
1749 | if (leaf->is_Con() || leaf->is_DecodeNarrowPtr()) { |
1750 | // Record the con for sharing |
1751 | _shared_nodes.map(leaf->_idx, ex); |
1752 | } |
1753 | |
1754 | return ex; |
1755 | } |
1756 | |
1757 | void Matcher::handle_precedence_edges(Node* n, MachNode *mach) { |
1758 | for (uint i = n->req(); i < n->len(); i++) { |
1759 | if (n->in(i) != NULL) { |
1760 | mach->add_prec(n->in(i)); |
1761 | } |
1762 | } |
1763 | } |
1764 | |
1765 | void Matcher::ReduceInst_Chain_Rule( State *s, int rule, Node *&mem, MachNode *mach ) { |
1766 | // 'op' is what I am expecting to receive |
1767 | int op = _leftOp[rule]; |
1768 | // Operand type to catch childs result |
1769 | // This is what my child will give me. |
1770 | int opnd_class_instance = s->_rule[op]; |
1771 | // Choose between operand class or not. |
1772 | // This is what I will receive. |
1773 | int catch_op = (FIRST_OPERAND_CLASS <= op && op < NUM_OPERANDS) ? opnd_class_instance : op; |
1774 | // New rule for child. Chase operand classes to get the actual rule. |
1775 | int newrule = s->_rule[catch_op]; |
1776 | |
1777 | if( newrule < NUM_OPERANDS ) { |
1778 | // Chain from operand or operand class, may be output of shared node |
1779 | assert( 0 <= opnd_class_instance && opnd_class_instance < NUM_OPERANDS, |
1780 | "Bad AD file: Instruction chain rule must chain from operand" ); |
1781 | // Insert operand into array of operands for this instruction |
1782 | mach->_opnds[1] = s->MachOperGenerator(opnd_class_instance); |
1783 | |
1784 | ReduceOper( s, newrule, mem, mach ); |
1785 | } else { |
1786 | // Chain from the result of an instruction |
1787 | assert( newrule >= _LAST_MACH_OPER, "Do NOT chain from internal operand" ); |
1788 | mach->_opnds[1] = s->MachOperGenerator(_reduceOp[catch_op]); |
1789 | Node *mem1 = (Node*)1; |
1790 | debug_only(Node *save_mem_node = _mem_node;) |
1791 | mach->add_req( ReduceInst(s, newrule, mem1) ); |
1792 | debug_only(_mem_node = save_mem_node;) |
1793 | } |
1794 | return; |
1795 | } |
1796 | |
1797 | |
1798 | uint Matcher::ReduceInst_Interior( State *s, int rule, Node *&mem, MachNode *mach, uint num_opnds ) { |
1799 | handle_precedence_edges(s->_leaf, mach); |
1800 | |
1801 | if( s->_leaf->is_Load() ) { |
1802 | Node *mem2 = s->_leaf->in(MemNode::Memory); |
1803 | assert( mem == (Node*)1 || mem == mem2, "multiple Memories being matched at once?" ); |
1804 | debug_only( if( mem == (Node*)1 ) _mem_node = s->_leaf;) |
1805 | mem = mem2; |
1806 | } |
1807 | if( s->_leaf->in(0) != NULL && s->_leaf->req() > 1) { |
1808 | if( mach->in(0) == NULL ) |
1809 | mach->set_req(0, s->_leaf->in(0)); |
1810 | } |
1811 | |
1812 | // Now recursively walk the state tree & add operand list. |
1813 | for( uint i=0; i<2; i++ ) { // binary tree |
1814 | State *newstate = s->_kids[i]; |
1815 | if( newstate == NULL ) break; // Might only have 1 child |
1816 | // 'op' is what I am expecting to receive |
1817 | int op; |
1818 | if( i == 0 ) { |
1819 | op = _leftOp[rule]; |
1820 | } else { |
1821 | op = _rightOp[rule]; |
1822 | } |
1823 | // Operand type to catch childs result |
1824 | // This is what my child will give me. |
1825 | int opnd_class_instance = newstate->_rule[op]; |
1826 | // Choose between operand class or not. |
1827 | // This is what I will receive. |
1828 | int catch_op = (op >= FIRST_OPERAND_CLASS && op < NUM_OPERANDS) ? opnd_class_instance : op; |
1829 | // New rule for child. Chase operand classes to get the actual rule. |
1830 | int newrule = newstate->_rule[catch_op]; |
1831 | |
1832 | if( newrule < NUM_OPERANDS ) { // Operand/operandClass or internalOp/instruction? |
1833 | // Operand/operandClass |
1834 | // Insert operand into array of operands for this instruction |
1835 | mach->_opnds[num_opnds++] = newstate->MachOperGenerator(opnd_class_instance); |
1836 | ReduceOper( newstate, newrule, mem, mach ); |
1837 | |
1838 | } else { // Child is internal operand or new instruction |
1839 | if( newrule < _LAST_MACH_OPER ) { // internal operand or instruction? |
1840 | // internal operand --> call ReduceInst_Interior |
1841 | // Interior of complex instruction. Do nothing but recurse. |
1842 | num_opnds = ReduceInst_Interior( newstate, newrule, mem, mach, num_opnds ); |
1843 | } else { |
1844 | // instruction --> call build operand( ) to catch result |
1845 | // --> ReduceInst( newrule ) |
1846 | mach->_opnds[num_opnds++] = s->MachOperGenerator(_reduceOp[catch_op]); |
1847 | Node *mem1 = (Node*)1; |
1848 | debug_only(Node *save_mem_node = _mem_node;) |
1849 | mach->add_req( ReduceInst( newstate, newrule, mem1 ) ); |
1850 | debug_only(_mem_node = save_mem_node;) |
1851 | } |
1852 | } |
1853 | assert( mach->_opnds[num_opnds-1], "" ); |
1854 | } |
1855 | return num_opnds; |
1856 | } |
1857 | |
1858 | // This routine walks the interior of possible complex operands. |
1859 | // At each point we check our children in the match tree: |
1860 | // (1) No children - |
1861 | // We are a leaf; add _leaf field as an input to the MachNode |
1862 | // (2) Child is an internal operand - |
1863 | // Skip over it ( do nothing ) |
1864 | // (3) Child is an instruction - |
1865 | // Call ReduceInst recursively and |
1866 | // and instruction as an input to the MachNode |
1867 | void Matcher::ReduceOper( State *s, int rule, Node *&mem, MachNode *mach ) { |
1868 | assert( rule < _LAST_MACH_OPER, "called with operand rule" ); |
1869 | State *kid = s->_kids[0]; |
1870 | assert( kid == NULL || s->_leaf->in(0) == NULL, "internal operands have no control" ); |
1871 | |
1872 | // Leaf? And not subsumed? |
1873 | if( kid == NULL && !_swallowed[rule] ) { |
1874 | mach->add_req( s->_leaf ); // Add leaf pointer |
1875 | return; // Bail out |
1876 | } |
1877 | |
1878 | if( s->_leaf->is_Load() ) { |
1879 | assert( mem == (Node*)1, "multiple Memories being matched at once?" ); |
1880 | mem = s->_leaf->in(MemNode::Memory); |
1881 | debug_only(_mem_node = s->_leaf;) |
1882 | } |
1883 | |
1884 | handle_precedence_edges(s->_leaf, mach); |
1885 | |
1886 | if( s->_leaf->in(0) && s->_leaf->req() > 1) { |
1887 | if( !mach->in(0) ) |
1888 | mach->set_req(0,s->_leaf->in(0)); |
1889 | else { |
1890 | assert( s->_leaf->in(0) == mach->in(0), "same instruction, differing controls?" ); |
1891 | } |
1892 | } |
1893 | |
1894 | for( uint i=0; kid != NULL && i<2; kid = s->_kids[1], i++ ) { // binary tree |
1895 | int newrule; |
1896 | if( i == 0) |
1897 | newrule = kid->_rule[_leftOp[rule]]; |
1898 | else |
1899 | newrule = kid->_rule[_rightOp[rule]]; |
1900 | |
1901 | if( newrule < _LAST_MACH_OPER ) { // Operand or instruction? |
1902 | // Internal operand; recurse but do nothing else |
1903 | ReduceOper( kid, newrule, mem, mach ); |
1904 | |
1905 | } else { // Child is a new instruction |
1906 | // Reduce the instruction, and add a direct pointer from this |
1907 | // machine instruction to the newly reduced one. |
1908 | Node *mem1 = (Node*)1; |
1909 | debug_only(Node *save_mem_node = _mem_node;) |
1910 | mach->add_req( ReduceInst( kid, newrule, mem1 ) ); |
1911 | debug_only(_mem_node = save_mem_node;) |
1912 | } |
1913 | } |
1914 | } |
1915 | |
1916 | |
1917 | // ------------------------------------------------------------------------- |
1918 | // Java-Java calling convention |
1919 | // (what you use when Java calls Java) |
1920 | |
1921 | //------------------------------find_receiver---------------------------------- |
1922 | // For a given signature, return the OptoReg for parameter 0. |
1923 | OptoReg::Name Matcher::find_receiver( bool is_outgoing ) { |
1924 | VMRegPair regs; |
1925 | BasicType sig_bt = T_OBJECT; |
1926 | calling_convention(&sig_bt, ®s, 1, is_outgoing); |
1927 | // Return argument 0 register. In the LP64 build pointers |
1928 | // take 2 registers, but the VM wants only the 'main' name. |
1929 | return OptoReg::as_OptoReg(regs.first()); |
1930 | } |
1931 | |
1932 | // This function identifies sub-graphs in which a 'load' node is |
1933 | // input to two different nodes, and such that it can be matched |
1934 | // with BMI instructions like blsi, blsr, etc. |
1935 | // Example : for b = -a[i] & a[i] can be matched to blsi r32, m32. |
1936 | // The graph is (AndL (SubL Con0 LoadL*) LoadL*), where LoadL* |
1937 | // refers to the same node. |
1938 | #ifdef X86 |
1939 | // Match the generic fused operations pattern (op1 (op2 Con{ConType} mop) mop) |
1940 | // This is a temporary solution until we make DAGs expressible in ADL. |
1941 | template<typename ConType> |
1942 | class FusedPatternMatcher { |
1943 | Node* _op1_node; |
1944 | Node* _mop_node; |
1945 | int _con_op; |
1946 | |
1947 | static int match_next(Node* n, int next_op, int next_op_idx) { |
1948 | if (n->in(1) == NULL || n->in(2) == NULL) { |
1949 | return -1; |
1950 | } |
1951 | |
1952 | if (next_op_idx == -1) { // n is commutative, try rotations |
1953 | if (n->in(1)->Opcode() == next_op) { |
1954 | return 1; |
1955 | } else if (n->in(2)->Opcode() == next_op) { |
1956 | return 2; |
1957 | } |
1958 | } else { |
1959 | assert(next_op_idx > 0 && next_op_idx <= 2, "Bad argument index" ); |
1960 | if (n->in(next_op_idx)->Opcode() == next_op) { |
1961 | return next_op_idx; |
1962 | } |
1963 | } |
1964 | return -1; |
1965 | } |
1966 | public: |
1967 | FusedPatternMatcher(Node* op1_node, Node *mop_node, int con_op) : |
1968 | _op1_node(op1_node), _mop_node(mop_node), _con_op(con_op) { } |
1969 | |
1970 | bool match(int op1, int op1_op2_idx, // op1 and the index of the op1->op2 edge, -1 if op1 is commutative |
1971 | int op2, int op2_con_idx, // op2 and the index of the op2->con edge, -1 if op2 is commutative |
1972 | typename ConType::NativeType con_value) { |
1973 | if (_op1_node->Opcode() != op1) { |
1974 | return false; |
1975 | } |
1976 | if (_mop_node->outcnt() > 2) { |
1977 | return false; |
1978 | } |
1979 | op1_op2_idx = match_next(_op1_node, op2, op1_op2_idx); |
1980 | if (op1_op2_idx == -1) { |
1981 | return false; |
1982 | } |
1983 | // Memory operation must be the other edge |
1984 | int op1_mop_idx = (op1_op2_idx & 1) + 1; |
1985 | |
1986 | // Check that the mop node is really what we want |
1987 | if (_op1_node->in(op1_mop_idx) == _mop_node) { |
1988 | Node *op2_node = _op1_node->in(op1_op2_idx); |
1989 | if (op2_node->outcnt() > 1) { |
1990 | return false; |
1991 | } |
1992 | assert(op2_node->Opcode() == op2, "Should be" ); |
1993 | op2_con_idx = match_next(op2_node, _con_op, op2_con_idx); |
1994 | if (op2_con_idx == -1) { |
1995 | return false; |
1996 | } |
1997 | // Memory operation must be the other edge |
1998 | int op2_mop_idx = (op2_con_idx & 1) + 1; |
1999 | // Check that the memory operation is the same node |
2000 | if (op2_node->in(op2_mop_idx) == _mop_node) { |
2001 | // Now check the constant |
2002 | const Type* con_type = op2_node->in(op2_con_idx)->bottom_type(); |
2003 | if (con_type != Type::TOP && ConType::as_self(con_type)->get_con() == con_value) { |
2004 | return true; |
2005 | } |
2006 | } |
2007 | } |
2008 | return false; |
2009 | } |
2010 | }; |
2011 | |
2012 | |
2013 | bool Matcher::is_bmi_pattern(Node *n, Node *m) { |
2014 | if (n != NULL && m != NULL) { |
2015 | if (m->Opcode() == Op_LoadI) { |
2016 | FusedPatternMatcher<TypeInt> bmii(n, m, Op_ConI); |
2017 | return bmii.match(Op_AndI, -1, Op_SubI, 1, 0) || |
2018 | bmii.match(Op_AndI, -1, Op_AddI, -1, -1) || |
2019 | bmii.match(Op_XorI, -1, Op_AddI, -1, -1); |
2020 | } else if (m->Opcode() == Op_LoadL) { |
2021 | FusedPatternMatcher<TypeLong> bmil(n, m, Op_ConL); |
2022 | return bmil.match(Op_AndL, -1, Op_SubL, 1, 0) || |
2023 | bmil.match(Op_AndL, -1, Op_AddL, -1, -1) || |
2024 | bmil.match(Op_XorL, -1, Op_AddL, -1, -1); |
2025 | } |
2026 | } |
2027 | return false; |
2028 | } |
2029 | #endif // X86 |
2030 | |
2031 | bool Matcher::clone_base_plus_offset_address(AddPNode* m, Matcher::MStack& mstack, VectorSet& address_visited) { |
2032 | Node *off = m->in(AddPNode::Offset); |
2033 | if (off->is_Con()) { |
2034 | address_visited.test_set(m->_idx); // Flag as address_visited |
2035 | mstack.push(m->in(AddPNode::Address), Pre_Visit); |
2036 | // Clone X+offset as it also folds into most addressing expressions |
2037 | mstack.push(off, Visit); |
2038 | mstack.push(m->in(AddPNode::Base), Pre_Visit); |
2039 | return true; |
2040 | } |
2041 | return false; |
2042 | } |
2043 | |
2044 | // A method-klass-holder may be passed in the inline_cache_reg |
2045 | // and then expanded into the inline_cache_reg and a method_oop register |
2046 | // defined in ad_<arch>.cpp |
2047 | |
2048 | //------------------------------find_shared------------------------------------ |
2049 | // Set bits if Node is shared or otherwise a root |
2050 | void Matcher::find_shared( Node *n ) { |
2051 | // Allocate stack of size C->live_nodes() * 2 to avoid frequent realloc |
2052 | MStack mstack(C->live_nodes() * 2); |
2053 | // Mark nodes as address_visited if they are inputs to an address expression |
2054 | VectorSet address_visited(Thread::current()->resource_area()); |
2055 | mstack.push(n, Visit); // Don't need to pre-visit root node |
2056 | while (mstack.is_nonempty()) { |
2057 | n = mstack.node(); // Leave node on stack |
2058 | Node_State nstate = mstack.state(); |
2059 | uint nop = n->Opcode(); |
2060 | if (nstate == Pre_Visit) { |
2061 | if (address_visited.test(n->_idx)) { // Visited in address already? |
2062 | // Flag as visited and shared now. |
2063 | set_visited(n); |
2064 | } |
2065 | if (is_visited(n)) { // Visited already? |
2066 | // Node is shared and has no reason to clone. Flag it as shared. |
2067 | // This causes it to match into a register for the sharing. |
2068 | set_shared(n); // Flag as shared and |
2069 | if (n->is_DecodeNarrowPtr()) { |
2070 | // Oop field/array element loads must be shared but since |
2071 | // they are shared through a DecodeN they may appear to have |
2072 | // a single use so force sharing here. |
2073 | set_shared(n->in(1)); |
2074 | } |
2075 | mstack.pop(); // remove node from stack |
2076 | continue; |
2077 | } |
2078 | nstate = Visit; // Not already visited; so visit now |
2079 | } |
2080 | if (nstate == Visit) { |
2081 | mstack.set_state(Post_Visit); |
2082 | set_visited(n); // Flag as visited now |
2083 | bool mem_op = false; |
2084 | int mem_addr_idx = MemNode::Address; |
2085 | bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->matcher_find_shared_visit(this, mstack, n, nop, mem_op, mem_addr_idx); |
2086 | if (!gc_handled) { |
2087 | if (find_shared_visit(mstack, n, nop, mem_op, mem_addr_idx)) { |
2088 | continue; |
2089 | } |
2090 | } |
2091 | for(int i = n->req() - 1; i >= 0; --i) { // For my children |
2092 | Node *m = n->in(i); // Get ith input |
2093 | if (m == NULL) continue; // Ignore NULLs |
2094 | uint mop = m->Opcode(); |
2095 | |
2096 | // Must clone all producers of flags, or we will not match correctly. |
2097 | // Suppose a compare setting int-flags is shared (e.g., a switch-tree) |
2098 | // then it will match into an ideal Op_RegFlags. Alas, the fp-flags |
2099 | // are also there, so we may match a float-branch to int-flags and |
2100 | // expect the allocator to haul the flags from the int-side to the |
2101 | // fp-side. No can do. |
2102 | if( _must_clone[mop] ) { |
2103 | mstack.push(m, Visit); |
2104 | continue; // for(int i = ...) |
2105 | } |
2106 | |
2107 | // if 'n' and 'm' are part of a graph for BMI instruction, clone this node. |
2108 | #ifdef X86 |
2109 | if (UseBMI1Instructions && is_bmi_pattern(n, m)) { |
2110 | mstack.push(m, Visit); |
2111 | continue; |
2112 | } |
2113 | #endif |
2114 | |
2115 | // Clone addressing expressions as they are "free" in memory access instructions |
2116 | if (mem_op && i == mem_addr_idx && mop == Op_AddP && |
2117 | // When there are other uses besides address expressions |
2118 | // put it on stack and mark as shared. |
2119 | !is_visited(m)) { |
2120 | // Some inputs for address expression are not put on stack |
2121 | // to avoid marking them as shared and forcing them into register |
2122 | // if they are used only in address expressions. |
2123 | // But they should be marked as shared if there are other uses |
2124 | // besides address expressions. |
2125 | |
2126 | if (clone_address_expressions(m->as_AddP(), mstack, address_visited)) { |
2127 | continue; |
2128 | } |
2129 | } // if( mem_op && |
2130 | mstack.push(m, Pre_Visit); |
2131 | } // for(int i = ...) |
2132 | } |
2133 | else if (nstate == Alt_Post_Visit) { |
2134 | mstack.pop(); // Remove node from stack |
2135 | // We cannot remove the Cmp input from the Bool here, as the Bool may be |
2136 | // shared and all users of the Bool need to move the Cmp in parallel. |
2137 | // This leaves both the Bool and the If pointing at the Cmp. To |
2138 | // prevent the Matcher from trying to Match the Cmp along both paths |
2139 | // BoolNode::match_edge always returns a zero. |
2140 | |
2141 | // We reorder the Op_If in a pre-order manner, so we can visit without |
2142 | // accidentally sharing the Cmp (the Bool and the If make 2 users). |
2143 | n->add_req( n->in(1)->in(1) ); // Add the Cmp next to the Bool |
2144 | } |
2145 | else if (nstate == Post_Visit) { |
2146 | mstack.pop(); // Remove node from stack |
2147 | |
2148 | // Now hack a few special opcodes |
2149 | uint opcode = n->Opcode(); |
2150 | bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->matcher_find_shared_post_visit(this, n, opcode); |
2151 | if (!gc_handled) { |
2152 | find_shared_post_visit(n, opcode); |
2153 | } |
2154 | } |
2155 | else { |
2156 | ShouldNotReachHere(); |
2157 | } |
2158 | } // end of while (mstack.is_nonempty()) |
2159 | } |
2160 | |
2161 | bool Matcher::find_shared_visit(MStack& mstack, Node* n, uint opcode, bool& mem_op, int& mem_addr_idx) { |
2162 | switch(opcode) { // Handle some opcodes special |
2163 | case Op_Phi: // Treat Phis as shared roots |
2164 | case Op_Parm: |
2165 | case Op_Proj: // All handled specially during matching |
2166 | case Op_SafePointScalarObject: |
2167 | set_shared(n); |
2168 | set_dontcare(n); |
2169 | break; |
2170 | case Op_If: |
2171 | case Op_CountedLoopEnd: |
2172 | mstack.set_state(Alt_Post_Visit); // Alternative way |
2173 | // Convert (If (Bool (CmpX A B))) into (If (Bool) (CmpX A B)). Helps |
2174 | // with matching cmp/branch in 1 instruction. The Matcher needs the |
2175 | // Bool and CmpX side-by-side, because it can only get at constants |
2176 | // that are at the leaves of Match trees, and the Bool's condition acts |
2177 | // as a constant here. |
2178 | mstack.push(n->in(1), Visit); // Clone the Bool |
2179 | mstack.push(n->in(0), Pre_Visit); // Visit control input |
2180 | return true; // while (mstack.is_nonempty()) |
2181 | case Op_ConvI2D: // These forms efficiently match with a prior |
2182 | case Op_ConvI2F: // Load but not a following Store |
2183 | if( n->in(1)->is_Load() && // Prior load |
2184 | n->outcnt() == 1 && // Not already shared |
2185 | n->unique_out()->is_Store() ) // Following store |
2186 | set_shared(n); // Force it to be a root |
2187 | break; |
2188 | case Op_ReverseBytesI: |
2189 | case Op_ReverseBytesL: |
2190 | if( n->in(1)->is_Load() && // Prior load |
2191 | n->outcnt() == 1 ) // Not already shared |
2192 | set_shared(n); // Force it to be a root |
2193 | break; |
2194 | case Op_BoxLock: // Cant match until we get stack-regs in ADLC |
2195 | case Op_IfFalse: |
2196 | case Op_IfTrue: |
2197 | case Op_MachProj: |
2198 | case Op_MergeMem: |
2199 | case Op_Catch: |
2200 | case Op_CatchProj: |
2201 | case Op_CProj: |
2202 | case Op_JumpProj: |
2203 | case Op_JProj: |
2204 | case Op_NeverBranch: |
2205 | set_dontcare(n); |
2206 | break; |
2207 | case Op_Jump: |
2208 | mstack.push(n->in(1), Pre_Visit); // Switch Value (could be shared) |
2209 | mstack.push(n->in(0), Pre_Visit); // Visit Control input |
2210 | return true; // while (mstack.is_nonempty()) |
2211 | case Op_StrComp: |
2212 | case Op_StrEquals: |
2213 | case Op_StrIndexOf: |
2214 | case Op_StrIndexOfChar: |
2215 | case Op_AryEq: |
2216 | case Op_HasNegatives: |
2217 | case Op_StrInflatedCopy: |
2218 | case Op_StrCompressedCopy: |
2219 | case Op_EncodeISOArray: |
2220 | case Op_FmaD: |
2221 | case Op_FmaF: |
2222 | case Op_FmaVD: |
2223 | case Op_FmaVF: |
2224 | set_shared(n); // Force result into register (it will be anyways) |
2225 | break; |
2226 | case Op_ConP: { // Convert pointers above the centerline to NUL |
2227 | TypeNode *tn = n->as_Type(); // Constants derive from type nodes |
2228 | const TypePtr* tp = tn->type()->is_ptr(); |
2229 | if (tp->_ptr == TypePtr::AnyNull) { |
2230 | tn->set_type(TypePtr::NULL_PTR); |
2231 | } |
2232 | break; |
2233 | } |
2234 | case Op_ConN: { // Convert narrow pointers above the centerline to NUL |
2235 | TypeNode *tn = n->as_Type(); // Constants derive from type nodes |
2236 | const TypePtr* tp = tn->type()->make_ptr(); |
2237 | if (tp && tp->_ptr == TypePtr::AnyNull) { |
2238 | tn->set_type(TypeNarrowOop::NULL_PTR); |
2239 | } |
2240 | break; |
2241 | } |
2242 | case Op_Binary: // These are introduced in the Post_Visit state. |
2243 | ShouldNotReachHere(); |
2244 | break; |
2245 | case Op_ClearArray: |
2246 | case Op_SafePoint: |
2247 | mem_op = true; |
2248 | break; |
2249 | default: |
2250 | if( n->is_Store() ) { |
2251 | // Do match stores, despite no ideal reg |
2252 | mem_op = true; |
2253 | break; |
2254 | } |
2255 | if( n->is_Mem() ) { // Loads and LoadStores |
2256 | mem_op = true; |
2257 | // Loads must be root of match tree due to prior load conflict |
2258 | if( C->subsume_loads() == false ) |
2259 | set_shared(n); |
2260 | } |
2261 | // Fall into default case |
2262 | if( !n->ideal_reg() ) |
2263 | set_dontcare(n); // Unmatchable Nodes |
2264 | } // end_switch |
2265 | return false; |
2266 | } |
2267 | |
2268 | void Matcher::find_shared_post_visit(Node* n, uint opcode) { |
2269 | switch(opcode) { // Handle some opcodes special |
2270 | case Op_StorePConditional: |
2271 | case Op_StoreIConditional: |
2272 | case Op_StoreLConditional: |
2273 | case Op_CompareAndExchangeB: |
2274 | case Op_CompareAndExchangeS: |
2275 | case Op_CompareAndExchangeI: |
2276 | case Op_CompareAndExchangeL: |
2277 | case Op_CompareAndExchangeP: |
2278 | case Op_CompareAndExchangeN: |
2279 | case Op_WeakCompareAndSwapB: |
2280 | case Op_WeakCompareAndSwapS: |
2281 | case Op_WeakCompareAndSwapI: |
2282 | case Op_WeakCompareAndSwapL: |
2283 | case Op_WeakCompareAndSwapP: |
2284 | case Op_WeakCompareAndSwapN: |
2285 | case Op_CompareAndSwapB: |
2286 | case Op_CompareAndSwapS: |
2287 | case Op_CompareAndSwapI: |
2288 | case Op_CompareAndSwapL: |
2289 | case Op_CompareAndSwapP: |
2290 | case Op_CompareAndSwapN: { // Convert trinary to binary-tree |
2291 | Node* newval = n->in(MemNode::ValueIn); |
2292 | Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn); |
2293 | Node* pair = new BinaryNode(oldval, newval); |
2294 | n->set_req(MemNode::ValueIn, pair); |
2295 | n->del_req(LoadStoreConditionalNode::ExpectedIn); |
2296 | break; |
2297 | } |
2298 | case Op_CMoveD: // Convert trinary to binary-tree |
2299 | case Op_CMoveF: |
2300 | case Op_CMoveI: |
2301 | case Op_CMoveL: |
2302 | case Op_CMoveN: |
2303 | case Op_CMoveP: |
2304 | case Op_CMoveVF: |
2305 | case Op_CMoveVD: { |
2306 | // Restructure into a binary tree for Matching. It's possible that |
2307 | // we could move this code up next to the graph reshaping for IfNodes |
2308 | // or vice-versa, but I do not want to debug this for Ladybird. |
2309 | // 10/2/2000 CNC. |
2310 | Node* pair1 = new BinaryNode(n->in(1), n->in(1)->in(1)); |
2311 | n->set_req(1, pair1); |
2312 | Node* pair2 = new BinaryNode(n->in(2), n->in(3)); |
2313 | n->set_req(2, pair2); |
2314 | n->del_req(3); |
2315 | break; |
2316 | } |
2317 | case Op_LoopLimit: { |
2318 | Node* pair1 = new BinaryNode(n->in(1), n->in(2)); |
2319 | n->set_req(1, pair1); |
2320 | n->set_req(2, n->in(3)); |
2321 | n->del_req(3); |
2322 | break; |
2323 | } |
2324 | case Op_StrEquals: |
2325 | case Op_StrIndexOfChar: { |
2326 | Node* pair1 = new BinaryNode(n->in(2), n->in(3)); |
2327 | n->set_req(2, pair1); |
2328 | n->set_req(3, n->in(4)); |
2329 | n->del_req(4); |
2330 | break; |
2331 | } |
2332 | case Op_StrComp: |
2333 | case Op_StrIndexOf: { |
2334 | Node* pair1 = new BinaryNode(n->in(2), n->in(3)); |
2335 | n->set_req(2, pair1); |
2336 | Node* pair2 = new BinaryNode(n->in(4),n->in(5)); |
2337 | n->set_req(3, pair2); |
2338 | n->del_req(5); |
2339 | n->del_req(4); |
2340 | break; |
2341 | } |
2342 | case Op_StrCompressedCopy: |
2343 | case Op_StrInflatedCopy: |
2344 | case Op_EncodeISOArray: { |
2345 | // Restructure into a binary tree for Matching. |
2346 | Node* pair = new BinaryNode(n->in(3), n->in(4)); |
2347 | n->set_req(3, pair); |
2348 | n->del_req(4); |
2349 | break; |
2350 | } |
2351 | case Op_FmaD: |
2352 | case Op_FmaF: |
2353 | case Op_FmaVD: |
2354 | case Op_FmaVF: { |
2355 | // Restructure into a binary tree for Matching. |
2356 | Node* pair = new BinaryNode(n->in(1), n->in(2)); |
2357 | n->set_req(2, pair); |
2358 | n->set_req(1, n->in(3)); |
2359 | n->del_req(3); |
2360 | break; |
2361 | } |
2362 | case Op_MulAddS2I: { |
2363 | Node* pair1 = new BinaryNode(n->in(1), n->in(2)); |
2364 | Node* pair2 = new BinaryNode(n->in(3), n->in(4)); |
2365 | n->set_req(1, pair1); |
2366 | n->set_req(2, pair2); |
2367 | n->del_req(4); |
2368 | n->del_req(3); |
2369 | break; |
2370 | } |
2371 | default: |
2372 | break; |
2373 | } |
2374 | } |
2375 | |
2376 | #ifdef ASSERT |
2377 | // machine-independent root to machine-dependent root |
2378 | void Matcher::dump_old2new_map() { |
2379 | _old2new_map.dump(); |
2380 | } |
2381 | #endif |
2382 | |
2383 | //---------------------------collect_null_checks------------------------------- |
2384 | // Find null checks in the ideal graph; write a machine-specific node for |
2385 | // it. Used by later implicit-null-check handling. Actually collects |
2386 | // either an IfTrue or IfFalse for the common NOT-null path, AND the ideal |
2387 | // value being tested. |
2388 | void Matcher::collect_null_checks( Node *proj, Node *orig_proj ) { |
2389 | Node *iff = proj->in(0); |
2390 | if( iff->Opcode() == Op_If ) { |
2391 | // During matching If's have Bool & Cmp side-by-side |
2392 | BoolNode *b = iff->in(1)->as_Bool(); |
2393 | Node *cmp = iff->in(2); |
2394 | int opc = cmp->Opcode(); |
2395 | if (opc != Op_CmpP && opc != Op_CmpN) return; |
2396 | |
2397 | const Type* ct = cmp->in(2)->bottom_type(); |
2398 | if (ct == TypePtr::NULL_PTR || |
2399 | (opc == Op_CmpN && ct == TypeNarrowOop::NULL_PTR)) { |
2400 | |
2401 | bool push_it = false; |
2402 | if( proj->Opcode() == Op_IfTrue ) { |
2403 | #ifndef PRODUCT |
2404 | extern int all_null_checks_found; |
2405 | all_null_checks_found++; |
2406 | #endif |
2407 | if( b->_test._test == BoolTest::ne ) { |
2408 | push_it = true; |
2409 | } |
2410 | } else { |
2411 | assert( proj->Opcode() == Op_IfFalse, "" ); |
2412 | if( b->_test._test == BoolTest::eq ) { |
2413 | push_it = true; |
2414 | } |
2415 | } |
2416 | if( push_it ) { |
2417 | _null_check_tests.push(proj); |
2418 | Node* val = cmp->in(1); |
2419 | #ifdef _LP64 |
2420 | if (val->bottom_type()->isa_narrowoop() && |
2421 | !Matcher::narrow_oop_use_complex_address()) { |
2422 | // |
2423 | // Look for DecodeN node which should be pinned to orig_proj. |
2424 | // On platforms (Sparc) which can not handle 2 adds |
2425 | // in addressing mode we have to keep a DecodeN node and |
2426 | // use it to do implicit NULL check in address. |
2427 | // |
2428 | // DecodeN node was pinned to non-null path (orig_proj) during |
2429 | // CastPP transformation in final_graph_reshaping_impl(). |
2430 | // |
2431 | uint cnt = orig_proj->outcnt(); |
2432 | for (uint i = 0; i < orig_proj->outcnt(); i++) { |
2433 | Node* d = orig_proj->raw_out(i); |
2434 | if (d->is_DecodeN() && d->in(1) == val) { |
2435 | val = d; |
2436 | val->set_req(0, NULL); // Unpin now. |
2437 | // Mark this as special case to distinguish from |
2438 | // a regular case: CmpP(DecodeN, NULL). |
2439 | val = (Node*)(((intptr_t)val) | 1); |
2440 | break; |
2441 | } |
2442 | } |
2443 | } |
2444 | #endif |
2445 | _null_check_tests.push(val); |
2446 | } |
2447 | } |
2448 | } |
2449 | } |
2450 | |
2451 | //---------------------------validate_null_checks------------------------------ |
2452 | // Its possible that the value being NULL checked is not the root of a match |
2453 | // tree. If so, I cannot use the value in an implicit null check. |
2454 | void Matcher::validate_null_checks( ) { |
2455 | uint cnt = _null_check_tests.size(); |
2456 | for( uint i=0; i < cnt; i+=2 ) { |
2457 | Node *test = _null_check_tests[i]; |
2458 | Node *val = _null_check_tests[i+1]; |
2459 | bool is_decoden = ((intptr_t)val) & 1; |
2460 | val = (Node*)(((intptr_t)val) & ~1); |
2461 | if (has_new_node(val)) { |
2462 | Node* new_val = new_node(val); |
2463 | if (is_decoden) { |
2464 | assert(val->is_DecodeNarrowPtr() && val->in(0) == NULL, "sanity" ); |
2465 | // Note: new_val may have a control edge if |
2466 | // the original ideal node DecodeN was matched before |
2467 | // it was unpinned in Matcher::collect_null_checks(). |
2468 | // Unpin the mach node and mark it. |
2469 | new_val->set_req(0, NULL); |
2470 | new_val = (Node*)(((intptr_t)new_val) | 1); |
2471 | } |
2472 | // Is a match-tree root, so replace with the matched value |
2473 | _null_check_tests.map(i+1, new_val); |
2474 | } else { |
2475 | // Yank from candidate list |
2476 | _null_check_tests.map(i+1,_null_check_tests[--cnt]); |
2477 | _null_check_tests.map(i,_null_check_tests[--cnt]); |
2478 | _null_check_tests.pop(); |
2479 | _null_check_tests.pop(); |
2480 | i-=2; |
2481 | } |
2482 | } |
2483 | } |
2484 | |
2485 | bool Matcher::gen_narrow_oop_implicit_null_checks() { |
2486 | // Advice matcher to perform null checks on the narrow oop side. |
2487 | // Implicit checks are not possible on the uncompressed oop side anyway |
2488 | // (at least not for read accesses). |
2489 | // Performs significantly better (especially on Power 6). |
2490 | if (!os::zero_page_read_protected()) { |
2491 | return true; |
2492 | } |
2493 | return CompressedOops::use_implicit_null_checks() && |
2494 | (narrow_oop_use_complex_address() || |
2495 | CompressedOops::base() != NULL); |
2496 | } |
2497 | |
2498 | // Used by the DFA in dfa_xxx.cpp. Check for a following barrier or |
2499 | // atomic instruction acting as a store_load barrier without any |
2500 | // intervening volatile load, and thus we don't need a barrier here. |
2501 | // We retain the Node to act as a compiler ordering barrier. |
2502 | bool Matcher::post_store_load_barrier(const Node* vmb) { |
2503 | Compile* C = Compile::current(); |
2504 | assert(vmb->is_MemBar(), "" ); |
2505 | assert(vmb->Opcode() != Op_MemBarAcquire && vmb->Opcode() != Op_LoadFence, "" ); |
2506 | const MemBarNode* membar = vmb->as_MemBar(); |
2507 | |
2508 | // Get the Ideal Proj node, ctrl, that can be used to iterate forward |
2509 | Node* ctrl = NULL; |
2510 | for (DUIterator_Fast imax, i = membar->fast_outs(imax); i < imax; i++) { |
2511 | Node* p = membar->fast_out(i); |
2512 | assert(p->is_Proj(), "only projections here" ); |
2513 | if ((p->as_Proj()->_con == TypeFunc::Control) && |
2514 | !C->node_arena()->contains(p)) { // Unmatched old-space only |
2515 | ctrl = p; |
2516 | break; |
2517 | } |
2518 | } |
2519 | assert((ctrl != NULL), "missing control projection" ); |
2520 | |
2521 | for (DUIterator_Fast jmax, j = ctrl->fast_outs(jmax); j < jmax; j++) { |
2522 | Node *x = ctrl->fast_out(j); |
2523 | int xop = x->Opcode(); |
2524 | |
2525 | // We don't need current barrier if we see another or a lock |
2526 | // before seeing volatile load. |
2527 | // |
2528 | // Op_Fastunlock previously appeared in the Op_* list below. |
2529 | // With the advent of 1-0 lock operations we're no longer guaranteed |
2530 | // that a monitor exit operation contains a serializing instruction. |
2531 | |
2532 | if (xop == Op_MemBarVolatile || |
2533 | xop == Op_CompareAndExchangeB || |
2534 | xop == Op_CompareAndExchangeS || |
2535 | xop == Op_CompareAndExchangeI || |
2536 | xop == Op_CompareAndExchangeL || |
2537 | xop == Op_CompareAndExchangeP || |
2538 | xop == Op_CompareAndExchangeN || |
2539 | xop == Op_WeakCompareAndSwapB || |
2540 | xop == Op_WeakCompareAndSwapS || |
2541 | xop == Op_WeakCompareAndSwapL || |
2542 | xop == Op_WeakCompareAndSwapP || |
2543 | xop == Op_WeakCompareAndSwapN || |
2544 | xop == Op_WeakCompareAndSwapI || |
2545 | xop == Op_CompareAndSwapB || |
2546 | xop == Op_CompareAndSwapS || |
2547 | xop == Op_CompareAndSwapL || |
2548 | xop == Op_CompareAndSwapP || |
2549 | xop == Op_CompareAndSwapN || |
2550 | xop == Op_CompareAndSwapI || |
2551 | BarrierSet::barrier_set()->barrier_set_c2()->matcher_is_store_load_barrier(x, xop)) { |
2552 | return true; |
2553 | } |
2554 | |
2555 | // Op_FastLock previously appeared in the Op_* list above. |
2556 | // With biased locking we're no longer guaranteed that a monitor |
2557 | // enter operation contains a serializing instruction. |
2558 | if ((xop == Op_FastLock) && !UseBiasedLocking) { |
2559 | return true; |
2560 | } |
2561 | |
2562 | if (x->is_MemBar()) { |
2563 | // We must retain this membar if there is an upcoming volatile |
2564 | // load, which will be followed by acquire membar. |
2565 | if (xop == Op_MemBarAcquire || xop == Op_LoadFence) { |
2566 | return false; |
2567 | } else { |
2568 | // For other kinds of barriers, check by pretending we |
2569 | // are them, and seeing if we can be removed. |
2570 | return post_store_load_barrier(x->as_MemBar()); |
2571 | } |
2572 | } |
2573 | |
2574 | // probably not necessary to check for these |
2575 | if (x->is_Call() || x->is_SafePoint() || x->is_block_proj()) { |
2576 | return false; |
2577 | } |
2578 | } |
2579 | return false; |
2580 | } |
2581 | |
2582 | // Check whether node n is a branch to an uncommon trap that we could |
2583 | // optimize as test with very high branch costs in case of going to |
2584 | // the uncommon trap. The code must be able to be recompiled to use |
2585 | // a cheaper test. |
2586 | bool Matcher::branches_to_uncommon_trap(const Node *n) { |
2587 | // Don't do it for natives, adapters, or runtime stubs |
2588 | Compile *C = Compile::current(); |
2589 | if (!C->is_method_compilation()) return false; |
2590 | |
2591 | assert(n->is_If(), "You should only call this on if nodes." ); |
2592 | IfNode *ifn = n->as_If(); |
2593 | |
2594 | Node *ifFalse = NULL; |
2595 | for (DUIterator_Fast imax, i = ifn->fast_outs(imax); i < imax; i++) { |
2596 | if (ifn->fast_out(i)->is_IfFalse()) { |
2597 | ifFalse = ifn->fast_out(i); |
2598 | break; |
2599 | } |
2600 | } |
2601 | assert(ifFalse, "An If should have an ifFalse. Graph is broken." ); |
2602 | |
2603 | Node *reg = ifFalse; |
2604 | int cnt = 4; // We must protect against cycles. Limit to 4 iterations. |
2605 | // Alternatively use visited set? Seems too expensive. |
2606 | while (reg != NULL && cnt > 0) { |
2607 | CallNode *call = NULL; |
2608 | RegionNode *nxt_reg = NULL; |
2609 | for (DUIterator_Fast imax, i = reg->fast_outs(imax); i < imax; i++) { |
2610 | Node *o = reg->fast_out(i); |
2611 | if (o->is_Call()) { |
2612 | call = o->as_Call(); |
2613 | } |
2614 | if (o->is_Region()) { |
2615 | nxt_reg = o->as_Region(); |
2616 | } |
2617 | } |
2618 | |
2619 | if (call && |
2620 | call->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) { |
2621 | const Type* trtype = call->in(TypeFunc::Parms)->bottom_type(); |
2622 | if (trtype->isa_int() && trtype->is_int()->is_con()) { |
2623 | jint tr_con = trtype->is_int()->get_con(); |
2624 | Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con); |
2625 | Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con); |
2626 | assert((int)reason < (int)BitsPerInt, "recode bit map" ); |
2627 | |
2628 | if (is_set_nth_bit(C->allowed_deopt_reasons(), (int)reason) |
2629 | && action != Deoptimization::Action_none) { |
2630 | // This uncommon trap is sure to recompile, eventually. |
2631 | // When that happens, C->too_many_traps will prevent |
2632 | // this transformation from happening again. |
2633 | return true; |
2634 | } |
2635 | } |
2636 | } |
2637 | |
2638 | reg = nxt_reg; |
2639 | cnt--; |
2640 | } |
2641 | |
2642 | return false; |
2643 | } |
2644 | |
2645 | //============================================================================= |
2646 | //---------------------------State--------------------------------------------- |
2647 | State::State(void) { |
2648 | #ifdef ASSERT |
2649 | _id = 0; |
2650 | _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe); |
2651 | _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d); |
2652 | //memset(_cost, -1, sizeof(_cost)); |
2653 | //memset(_rule, -1, sizeof(_rule)); |
2654 | #endif |
2655 | memset(_valid, 0, sizeof(_valid)); |
2656 | } |
2657 | |
2658 | #ifdef ASSERT |
2659 | State::~State() { |
2660 | _id = 99; |
2661 | _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe); |
2662 | _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d); |
2663 | memset(_cost, -3, sizeof(_cost)); |
2664 | memset(_rule, -3, sizeof(_rule)); |
2665 | } |
2666 | #endif |
2667 | |
2668 | #ifndef PRODUCT |
2669 | //---------------------------dump---------------------------------------------- |
2670 | void State::dump() { |
2671 | tty->print("\n" ); |
2672 | dump(0); |
2673 | } |
2674 | |
2675 | void State::dump(int depth) { |
2676 | for( int j = 0; j < depth; j++ ) |
2677 | tty->print(" " ); |
2678 | tty->print("--N: " ); |
2679 | _leaf->dump(); |
2680 | uint i; |
2681 | for( i = 0; i < _LAST_MACH_OPER; i++ ) |
2682 | // Check for valid entry |
2683 | if( valid(i) ) { |
2684 | for( int j = 0; j < depth; j++ ) |
2685 | tty->print(" " ); |
2686 | assert(_cost[i] != max_juint, "cost must be a valid value" ); |
2687 | assert(_rule[i] < _last_Mach_Node, "rule[i] must be valid rule" ); |
2688 | tty->print_cr("%s %d %s" , |
2689 | ruleName[i], _cost[i], ruleName[_rule[i]] ); |
2690 | } |
2691 | tty->cr(); |
2692 | |
2693 | for( i=0; i<2; i++ ) |
2694 | if( _kids[i] ) |
2695 | _kids[i]->dump(depth+1); |
2696 | } |
2697 | #endif |
2698 | |