Mercurial > hg > graal-compiler
annotate src/share/vm/opto/matcher.cpp @ 681:fbde8ec322d0
6761600: Use sse 4.2 in intrinsics
Summary: Use SSE 4.2 in intrinsics for String.{compareTo/equals/indexOf} and Arrays.equals.
Reviewed-by: kvn, never, jrose
author | cfang |
---|---|
date | Tue, 31 Mar 2009 14:07:08 -0700 |
parents | bd441136a5ce |
children | 04fa5affa478 |
rev | line source |
---|---|
0 | 1 /* |
579 | 2 * Copyright 1997-2009 Sun Microsystems, Inc. All Rights Reserved. |
0 | 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
20 * CA 95054 USA or visit www.sun.com if you need additional information or | |
21 * have any questions. | |
22 * | |
23 */ | |
24 | |
25 #include "incls/_precompiled.incl" | |
26 #include "incls/_matcher.cpp.incl" | |
27 | |
28 OptoReg::Name OptoReg::c_frame_pointer; | |
29 | |
30 | |
31 | |
32 const int Matcher::base2reg[Type::lastype] = { | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
33 Node::NotAMachineReg,0,0, Op_RegI, Op_RegL, 0, Op_RegN, |
0 | 34 Node::NotAMachineReg, Node::NotAMachineReg, /* tuple, array */ |
35 Op_RegP, Op_RegP, Op_RegP, Op_RegP, Op_RegP, Op_RegP, /* the pointers */ | |
36 0, 0/*abio*/, | |
37 Op_RegP /* Return address */, 0, /* the memories */ | |
38 Op_RegF, Op_RegF, Op_RegF, Op_RegD, Op_RegD, Op_RegD, | |
39 0 /*bottom*/ | |
40 }; | |
41 | |
42 const RegMask *Matcher::idealreg2regmask[_last_machine_leaf]; | |
43 RegMask Matcher::mreg2regmask[_last_Mach_Reg]; | |
44 RegMask Matcher::STACK_ONLY_mask; | |
45 RegMask Matcher::c_frame_ptr_mask; | |
46 const uint Matcher::_begin_rematerialize = _BEGIN_REMATERIALIZE; | |
47 const uint Matcher::_end_rematerialize = _END_REMATERIALIZE; | |
48 | |
49 //---------------------------Matcher------------------------------------------- | |
50 Matcher::Matcher( Node_List &proj_list ) : | |
51 PhaseTransform( Phase::Ins_Select ), | |
52 #ifdef ASSERT | |
53 _old2new_map(C->comp_arena()), | |
222 | 54 _new2old_map(C->comp_arena()), |
0 | 55 #endif |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
56 _shared_nodes(C->comp_arena()), |
0 | 57 _reduceOp(reduceOp), _leftOp(leftOp), _rightOp(rightOp), |
58 _swallowed(swallowed), | |
59 _begin_inst_chain_rule(_BEGIN_INST_CHAIN_RULE), | |
60 _end_inst_chain_rule(_END_INST_CHAIN_RULE), | |
61 _must_clone(must_clone), _proj_list(proj_list), | |
62 _register_save_policy(register_save_policy), | |
63 _c_reg_save_policy(c_reg_save_policy), | |
64 _register_save_type(register_save_type), | |
65 _ruleName(ruleName), | |
66 _allocation_started(false), | |
67 _states_arena(Chunk::medium_size), | |
68 _visited(&_states_arena), | |
69 _shared(&_states_arena), | |
70 _dontcare(&_states_arena) { | |
71 C->set_matcher(this); | |
72 | |
73 idealreg2spillmask[Op_RegI] = NULL; | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
74 idealreg2spillmask[Op_RegN] = NULL; |
0 | 75 idealreg2spillmask[Op_RegL] = NULL; |
76 idealreg2spillmask[Op_RegF] = NULL; | |
77 idealreg2spillmask[Op_RegD] = NULL; | |
78 idealreg2spillmask[Op_RegP] = NULL; | |
79 | |
80 idealreg2debugmask[Op_RegI] = NULL; | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
81 idealreg2debugmask[Op_RegN] = NULL; |
0 | 82 idealreg2debugmask[Op_RegL] = NULL; |
83 idealreg2debugmask[Op_RegF] = NULL; | |
84 idealreg2debugmask[Op_RegD] = NULL; | |
85 idealreg2debugmask[Op_RegP] = NULL; | |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
86 debug_only(_mem_node = NULL;) // Ideal memory node consumed by mach node |
0 | 87 } |
88 | |
89 //------------------------------warp_incoming_stk_arg------------------------ | |
90 // This warps a VMReg into an OptoReg::Name | |
91 OptoReg::Name Matcher::warp_incoming_stk_arg( VMReg reg ) { | |
92 OptoReg::Name warped; | |
93 if( reg->is_stack() ) { // Stack slot argument? | |
94 warped = OptoReg::add(_old_SP, reg->reg2stack() ); | |
95 warped = OptoReg::add(warped, C->out_preserve_stack_slots()); | |
96 if( warped >= _in_arg_limit ) | |
97 _in_arg_limit = OptoReg::add(warped, 1); // Bump max stack slot seen | |
98 if (!RegMask::can_represent(warped)) { | |
99 // the compiler cannot represent this method's calling sequence | |
100 C->record_method_not_compilable_all_tiers("unsupported incoming calling sequence"); | |
101 return OptoReg::Bad; | |
102 } | |
103 return warped; | |
104 } | |
105 return OptoReg::as_OptoReg(reg); | |
106 } | |
107 | |
108 //---------------------------compute_old_SP------------------------------------ | |
109 OptoReg::Name Compile::compute_old_SP() { | |
110 int fixed = fixed_slots(); | |
111 int preserve = in_preserve_stack_slots(); | |
112 return OptoReg::stack2reg(round_to(fixed + preserve, Matcher::stack_alignment_in_slots())); | |
113 } | |
114 | |
115 | |
116 | |
117 #ifdef ASSERT | |
118 void Matcher::verify_new_nodes_only(Node* xroot) { | |
119 // Make sure that the new graph only references new nodes | |
120 ResourceMark rm; | |
121 Unique_Node_List worklist; | |
122 VectorSet visited(Thread::current()->resource_area()); | |
123 worklist.push(xroot); | |
124 while (worklist.size() > 0) { | |
125 Node* n = worklist.pop(); | |
126 visited <<= n->_idx; | |
127 assert(C->node_arena()->contains(n), "dead node"); | |
128 for (uint j = 0; j < n->req(); j++) { | |
129 Node* in = n->in(j); | |
130 if (in != NULL) { | |
131 assert(C->node_arena()->contains(in), "dead node"); | |
132 if (!visited.test(in->_idx)) { | |
133 worklist.push(in); | |
134 } | |
135 } | |
136 } | |
137 } | |
138 } | |
139 #endif | |
140 | |
141 | |
142 //---------------------------match--------------------------------------------- | |
143 void Matcher::match( ) { | |
144 // One-time initialization of some register masks. | |
145 init_spill_mask( C->root()->in(1) ); | |
146 _return_addr_mask = return_addr(); | |
147 #ifdef _LP64 | |
148 // Pointers take 2 slots in 64-bit land | |
149 _return_addr_mask.Insert(OptoReg::add(return_addr(),1)); | |
150 #endif | |
151 | |
152 // Map a Java-signature return type into return register-value | |
153 // machine registers for 0, 1 and 2 returned values. | |
154 const TypeTuple *range = C->tf()->range(); | |
155 if( range->cnt() > TypeFunc::Parms ) { // If not a void function | |
156 // Get ideal-register return type | |
157 int ireg = base2reg[range->field_at(TypeFunc::Parms)->base()]; | |
158 // Get machine return register | |
159 uint sop = C->start()->Opcode(); | |
160 OptoRegPair regs = return_value(ireg, false); | |
161 | |
162 // And mask for same | |
163 _return_value_mask = RegMask(regs.first()); | |
164 if( OptoReg::is_valid(regs.second()) ) | |
165 _return_value_mask.Insert(regs.second()); | |
166 } | |
167 | |
168 // --------------- | |
169 // Frame Layout | |
170 | |
171 // Need the method signature to determine the incoming argument types, | |
172 // because the types determine which registers the incoming arguments are | |
173 // in, and this affects the matched code. | |
174 const TypeTuple *domain = C->tf()->domain(); | |
175 uint argcnt = domain->cnt() - TypeFunc::Parms; | |
176 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt ); | |
177 VMRegPair *vm_parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt ); | |
178 _parm_regs = NEW_RESOURCE_ARRAY( OptoRegPair, argcnt ); | |
179 _calling_convention_mask = NEW_RESOURCE_ARRAY( RegMask, argcnt ); | |
180 uint i; | |
181 for( i = 0; i<argcnt; i++ ) { | |
182 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type(); | |
183 } | |
184 | |
185 // Pass array of ideal registers and length to USER code (from the AD file) | |
186 // that will convert this to an array of register numbers. | |
187 const StartNode *start = C->start(); | |
188 start->calling_convention( sig_bt, vm_parm_regs, argcnt ); | |
189 #ifdef ASSERT | |
190 // Sanity check users' calling convention. Real handy while trying to | |
191 // get the initial port correct. | |
192 { for (uint i = 0; i<argcnt; i++) { | |
193 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) { | |
194 assert(domain->field_at(i+TypeFunc::Parms)==Type::HALF, "only allowed on halve" ); | |
195 _parm_regs[i].set_bad(); | |
196 continue; | |
197 } | |
198 VMReg parm_reg = vm_parm_regs[i].first(); | |
199 assert(parm_reg->is_valid(), "invalid arg?"); | |
200 if (parm_reg->is_reg()) { | |
201 OptoReg::Name opto_parm_reg = OptoReg::as_OptoReg(parm_reg); | |
202 assert(can_be_java_arg(opto_parm_reg) || | |
203 C->stub_function() == CAST_FROM_FN_PTR(address, OptoRuntime::rethrow_C) || | |
204 opto_parm_reg == inline_cache_reg(), | |
205 "parameters in register must be preserved by runtime stubs"); | |
206 } | |
207 for (uint j = 0; j < i; j++) { | |
208 assert(parm_reg != vm_parm_regs[j].first(), | |
209 "calling conv. must produce distinct regs"); | |
210 } | |
211 } | |
212 } | |
213 #endif | |
214 | |
215 // Do some initial frame layout. | |
216 | |
217 // Compute the old incoming SP (may be called FP) as | |
218 // OptoReg::stack0() + locks + in_preserve_stack_slots + pad2. | |
219 _old_SP = C->compute_old_SP(); | |
220 assert( is_even(_old_SP), "must be even" ); | |
221 | |
222 // Compute highest incoming stack argument as | |
223 // _old_SP + out_preserve_stack_slots + incoming argument size. | |
224 _in_arg_limit = OptoReg::add(_old_SP, C->out_preserve_stack_slots()); | |
225 assert( is_even(_in_arg_limit), "out_preserve must be even" ); | |
226 for( i = 0; i < argcnt; i++ ) { | |
227 // Permit args to have no register | |
228 _calling_convention_mask[i].Clear(); | |
229 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) { | |
230 continue; | |
231 } | |
232 // calling_convention returns stack arguments as a count of | |
233 // slots beyond OptoReg::stack0()/VMRegImpl::stack0. We need to convert this to | |
234 // the allocators point of view, taking into account all the | |
235 // preserve area, locks & pad2. | |
236 | |
237 OptoReg::Name reg1 = warp_incoming_stk_arg(vm_parm_regs[i].first()); | |
238 if( OptoReg::is_valid(reg1)) | |
239 _calling_convention_mask[i].Insert(reg1); | |
240 | |
241 OptoReg::Name reg2 = warp_incoming_stk_arg(vm_parm_regs[i].second()); | |
242 if( OptoReg::is_valid(reg2)) | |
243 _calling_convention_mask[i].Insert(reg2); | |
244 | |
245 // Saved biased stack-slot register number | |
246 _parm_regs[i].set_pair(reg2, reg1); | |
247 } | |
248 | |
249 // Finally, make sure the incoming arguments take up an even number of | |
250 // words, in case the arguments or locals need to contain doubleword stack | |
251 // slots. The rest of the system assumes that stack slot pairs (in | |
252 // particular, in the spill area) which look aligned will in fact be | |
253 // aligned relative to the stack pointer in the target machine. Double | |
254 // stack slots will always be allocated aligned. | |
255 _new_SP = OptoReg::Name(round_to(_in_arg_limit, RegMask::SlotsPerLong)); | |
256 | |
257 // Compute highest outgoing stack argument as | |
258 // _new_SP + out_preserve_stack_slots + max(outgoing argument size). | |
259 _out_arg_limit = OptoReg::add(_new_SP, C->out_preserve_stack_slots()); | |
260 assert( is_even(_out_arg_limit), "out_preserve must be even" ); | |
261 | |
262 if (!RegMask::can_represent(OptoReg::add(_out_arg_limit,-1))) { | |
263 // the compiler cannot represent this method's calling sequence | |
264 C->record_method_not_compilable("must be able to represent all call arguments in reg mask"); | |
265 } | |
266 | |
267 if (C->failing()) return; // bailed out on incoming arg failure | |
268 | |
269 // --------------- | |
270 // Collect roots of matcher trees. Every node for which | |
271 // _shared[_idx] is cleared is guaranteed to not be shared, and thus | |
272 // can be a valid interior of some tree. | |
273 find_shared( C->root() ); | |
274 find_shared( C->top() ); | |
275 | |
367
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
309
diff
changeset
|
276 C->print_method("Before Matching"); |
0 | 277 |
278 // Swap out to old-space; emptying new-space | |
279 Arena *old = C->node_arena()->move_contents(C->old_arena()); | |
280 | |
281 // Save debug and profile information for nodes in old space: | |
282 _old_node_note_array = C->node_note_array(); | |
283 if (_old_node_note_array != NULL) { | |
284 C->set_node_note_array(new(C->comp_arena()) GrowableArray<Node_Notes*> | |
285 (C->comp_arena(), _old_node_note_array->length(), | |
286 0, NULL)); | |
287 } | |
288 | |
289 // Pre-size the new_node table to avoid the need for range checks. | |
290 grow_new_node_array(C->unique()); | |
291 | |
292 // Reset node counter so MachNodes start with _idx at 0 | |
293 int nodes = C->unique(); // save value | |
294 C->set_unique(0); | |
295 | |
296 // Recursively match trees from old space into new space. | |
297 // Correct leaves of new-space Nodes; they point to old-space. | |
298 _visited.Clear(); // Clear visit bits for xform call | |
299 C->set_cached_top_node(xform( C->top(), nodes )); | |
300 if (!C->failing()) { | |
301 Node* xroot = xform( C->root(), 1 ); | |
302 if (xroot == NULL) { | |
303 Matcher::soft_match_failure(); // recursive matching process failed | |
304 C->record_method_not_compilable("instruction match failed"); | |
305 } else { | |
306 // During matching shared constants were attached to C->root() | |
307 // because xroot wasn't available yet, so transfer the uses to | |
308 // the xroot. | |
309 for( DUIterator_Fast jmax, j = C->root()->fast_outs(jmax); j < jmax; j++ ) { | |
310 Node* n = C->root()->fast_out(j); | |
311 if (C->node_arena()->contains(n)) { | |
312 assert(n->in(0) == C->root(), "should be control user"); | |
313 n->set_req(0, xroot); | |
314 --j; | |
315 --jmax; | |
316 } | |
317 } | |
318 | |
319 C->set_root(xroot->is_Root() ? xroot->as_Root() : NULL); | |
320 #ifdef ASSERT | |
321 verify_new_nodes_only(xroot); | |
322 #endif | |
323 } | |
324 } | |
325 if (C->top() == NULL || C->root() == NULL) { | |
326 C->record_method_not_compilable("graph lost"); // %%% cannot happen? | |
327 } | |
328 if (C->failing()) { | |
329 // delete old; | |
330 old->destruct_contents(); | |
331 return; | |
332 } | |
333 assert( C->top(), "" ); | |
334 assert( C->root(), "" ); | |
335 validate_null_checks(); | |
336 | |
337 // Now smoke old-space | |
338 NOT_DEBUG( old->destruct_contents() ); | |
339 | |
340 // ------------------------ | |
341 // Set up save-on-entry registers | |
342 Fixup_Save_On_Entry( ); | |
343 } | |
344 | |
345 | |
346 //------------------------------Fixup_Save_On_Entry---------------------------- | |
347 // The stated purpose of this routine is to take care of save-on-entry | |
348 // registers. However, the overall goal of the Match phase is to convert into | |
349 // machine-specific instructions which have RegMasks to guide allocation. | |
350 // So what this procedure really does is put a valid RegMask on each input | |
351 // to the machine-specific variations of all Return, TailCall and Halt | |
352 // instructions. It also adds edgs to define the save-on-entry values (and of | |
353 // course gives them a mask). | |
354 | |
355 static RegMask *init_input_masks( uint size, RegMask &ret_adr, RegMask &fp ) { | |
356 RegMask *rms = NEW_RESOURCE_ARRAY( RegMask, size ); | |
357 // Do all the pre-defined register masks | |
358 rms[TypeFunc::Control ] = RegMask::Empty; | |
359 rms[TypeFunc::I_O ] = RegMask::Empty; | |
360 rms[TypeFunc::Memory ] = RegMask::Empty; | |
361 rms[TypeFunc::ReturnAdr] = ret_adr; | |
362 rms[TypeFunc::FramePtr ] = fp; | |
363 return rms; | |
364 } | |
365 | |
366 //---------------------------init_first_stack_mask----------------------------- | |
367 // Create the initial stack mask used by values spilling to the stack. | |
368 // Disallow any debug info in outgoing argument areas by setting the | |
369 // initial mask accordingly. | |
370 void Matcher::init_first_stack_mask() { | |
371 | |
372 // Allocate storage for spill masks as masks for the appropriate load type. | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
373 RegMask *rms = (RegMask*)C->comp_arena()->Amalloc_D(sizeof(RegMask)*12); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
374 idealreg2spillmask[Op_RegN] = &rms[0]; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
375 idealreg2spillmask[Op_RegI] = &rms[1]; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
376 idealreg2spillmask[Op_RegL] = &rms[2]; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
377 idealreg2spillmask[Op_RegF] = &rms[3]; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
378 idealreg2spillmask[Op_RegD] = &rms[4]; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
379 idealreg2spillmask[Op_RegP] = &rms[5]; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
380 idealreg2debugmask[Op_RegN] = &rms[6]; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
381 idealreg2debugmask[Op_RegI] = &rms[7]; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
382 idealreg2debugmask[Op_RegL] = &rms[8]; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
383 idealreg2debugmask[Op_RegF] = &rms[9]; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
384 idealreg2debugmask[Op_RegD] = &rms[10]; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
385 idealreg2debugmask[Op_RegP] = &rms[11]; |
0 | 386 |
387 OptoReg::Name i; | |
388 | |
389 // At first, start with the empty mask | |
390 C->FIRST_STACK_mask().Clear(); | |
391 | |
392 // Add in the incoming argument area | |
393 OptoReg::Name init = OptoReg::add(_old_SP, C->out_preserve_stack_slots()); | |
394 for (i = init; i < _in_arg_limit; i = OptoReg::add(i,1)) | |
395 C->FIRST_STACK_mask().Insert(i); | |
396 | |
397 // Add in all bits past the outgoing argument area | |
398 guarantee(RegMask::can_represent(OptoReg::add(_out_arg_limit,-1)), | |
399 "must be able to represent all call arguments in reg mask"); | |
400 init = _out_arg_limit; | |
401 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) | |
402 C->FIRST_STACK_mask().Insert(i); | |
403 | |
404 // Finally, set the "infinite stack" bit. | |
405 C->FIRST_STACK_mask().set_AllStack(); | |
406 | |
407 // Make spill masks. Registers for their class, plus FIRST_STACK_mask. | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
408 #ifdef _LP64 |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
409 *idealreg2spillmask[Op_RegN] = *idealreg2regmask[Op_RegN]; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
410 idealreg2spillmask[Op_RegN]->OR(C->FIRST_STACK_mask()); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
411 #endif |
0 | 412 *idealreg2spillmask[Op_RegI] = *idealreg2regmask[Op_RegI]; |
413 idealreg2spillmask[Op_RegI]->OR(C->FIRST_STACK_mask()); | |
414 *idealreg2spillmask[Op_RegL] = *idealreg2regmask[Op_RegL]; | |
415 idealreg2spillmask[Op_RegL]->OR(C->FIRST_STACK_mask()); | |
416 *idealreg2spillmask[Op_RegF] = *idealreg2regmask[Op_RegF]; | |
417 idealreg2spillmask[Op_RegF]->OR(C->FIRST_STACK_mask()); | |
418 *idealreg2spillmask[Op_RegD] = *idealreg2regmask[Op_RegD]; | |
419 idealreg2spillmask[Op_RegD]->OR(C->FIRST_STACK_mask()); | |
420 *idealreg2spillmask[Op_RegP] = *idealreg2regmask[Op_RegP]; | |
421 idealreg2spillmask[Op_RegP]->OR(C->FIRST_STACK_mask()); | |
422 | |
423 // Make up debug masks. Any spill slot plus callee-save registers. | |
424 // Caller-save registers are assumed to be trashable by the various | |
425 // inline-cache fixup routines. | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
426 *idealreg2debugmask[Op_RegN]= *idealreg2spillmask[Op_RegN]; |
0 | 427 *idealreg2debugmask[Op_RegI]= *idealreg2spillmask[Op_RegI]; |
428 *idealreg2debugmask[Op_RegL]= *idealreg2spillmask[Op_RegL]; | |
429 *idealreg2debugmask[Op_RegF]= *idealreg2spillmask[Op_RegF]; | |
430 *idealreg2debugmask[Op_RegD]= *idealreg2spillmask[Op_RegD]; | |
431 *idealreg2debugmask[Op_RegP]= *idealreg2spillmask[Op_RegP]; | |
432 | |
433 // Prevent stub compilations from attempting to reference | |
434 // callee-saved registers from debug info | |
435 bool exclude_soe = !Compile::current()->is_method_compilation(); | |
436 | |
437 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) { | |
438 // registers the caller has to save do not work | |
439 if( _register_save_policy[i] == 'C' || | |
440 _register_save_policy[i] == 'A' || | |
441 (_register_save_policy[i] == 'E' && exclude_soe) ) { | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
442 idealreg2debugmask[Op_RegN]->Remove(i); |
0 | 443 idealreg2debugmask[Op_RegI]->Remove(i); // Exclude save-on-call |
444 idealreg2debugmask[Op_RegL]->Remove(i); // registers from debug | |
445 idealreg2debugmask[Op_RegF]->Remove(i); // masks | |
446 idealreg2debugmask[Op_RegD]->Remove(i); | |
447 idealreg2debugmask[Op_RegP]->Remove(i); | |
448 } | |
449 } | |
450 } | |
451 | |
452 //---------------------------is_save_on_entry---------------------------------- | |
453 bool Matcher::is_save_on_entry( int reg ) { | |
454 return | |
455 _register_save_policy[reg] == 'E' || | |
456 _register_save_policy[reg] == 'A' || // Save-on-entry register? | |
457 // Also save argument registers in the trampolining stubs | |
458 (C->save_argument_registers() && is_spillable_arg(reg)); | |
459 } | |
460 | |
461 //---------------------------Fixup_Save_On_Entry------------------------------- | |
462 void Matcher::Fixup_Save_On_Entry( ) { | |
463 init_first_stack_mask(); | |
464 | |
465 Node *root = C->root(); // Short name for root | |
466 // Count number of save-on-entry registers. | |
467 uint soe_cnt = number_of_saved_registers(); | |
468 uint i; | |
469 | |
470 // Find the procedure Start Node | |
471 StartNode *start = C->start(); | |
472 assert( start, "Expect a start node" ); | |
473 | |
474 // Save argument registers in the trampolining stubs | |
475 if( C->save_argument_registers() ) | |
476 for( i = 0; i < _last_Mach_Reg; i++ ) | |
477 if( is_spillable_arg(i) ) | |
478 soe_cnt++; | |
479 | |
480 // Input RegMask array shared by all Returns. | |
481 // The type for doubles and longs has a count of 2, but | |
482 // there is only 1 returned value | |
483 uint ret_edge_cnt = TypeFunc::Parms + ((C->tf()->range()->cnt() == TypeFunc::Parms) ? 0 : 1); | |
484 RegMask *ret_rms = init_input_masks( ret_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); | |
485 // Returns have 0 or 1 returned values depending on call signature. | |
486 // Return register is specified by return_value in the AD file. | |
487 if (ret_edge_cnt > TypeFunc::Parms) | |
488 ret_rms[TypeFunc::Parms+0] = _return_value_mask; | |
489 | |
490 // Input RegMask array shared by all Rethrows. | |
491 uint reth_edge_cnt = TypeFunc::Parms+1; | |
492 RegMask *reth_rms = init_input_masks( reth_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); | |
493 // Rethrow takes exception oop only, but in the argument 0 slot. | |
494 reth_rms[TypeFunc::Parms] = mreg2regmask[find_receiver(false)]; | |
495 #ifdef _LP64 | |
496 // Need two slots for ptrs in 64-bit land | |
497 reth_rms[TypeFunc::Parms].Insert(OptoReg::add(OptoReg::Name(find_receiver(false)),1)); | |
498 #endif | |
499 | |
500 // Input RegMask array shared by all TailCalls | |
501 uint tail_call_edge_cnt = TypeFunc::Parms+2; | |
502 RegMask *tail_call_rms = init_input_masks( tail_call_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); | |
503 | |
504 // Input RegMask array shared by all TailJumps | |
505 uint tail_jump_edge_cnt = TypeFunc::Parms+2; | |
506 RegMask *tail_jump_rms = init_input_masks( tail_jump_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); | |
507 | |
508 // TailCalls have 2 returned values (target & moop), whose masks come | |
509 // from the usual MachNode/MachOper mechanism. Find a sample | |
510 // TailCall to extract these masks and put the correct masks into | |
511 // the tail_call_rms array. | |
512 for( i=1; i < root->req(); i++ ) { | |
513 MachReturnNode *m = root->in(i)->as_MachReturn(); | |
514 if( m->ideal_Opcode() == Op_TailCall ) { | |
515 tail_call_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0); | |
516 tail_call_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1); | |
517 break; | |
518 } | |
519 } | |
520 | |
521 // TailJumps have 2 returned values (target & ex_oop), whose masks come | |
522 // from the usual MachNode/MachOper mechanism. Find a sample | |
523 // TailJump to extract these masks and put the correct masks into | |
524 // the tail_jump_rms array. | |
525 for( i=1; i < root->req(); i++ ) { | |
526 MachReturnNode *m = root->in(i)->as_MachReturn(); | |
527 if( m->ideal_Opcode() == Op_TailJump ) { | |
528 tail_jump_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0); | |
529 tail_jump_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1); | |
530 break; | |
531 } | |
532 } | |
533 | |
534 // Input RegMask array shared by all Halts | |
535 uint halt_edge_cnt = TypeFunc::Parms; | |
536 RegMask *halt_rms = init_input_masks( halt_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); | |
537 | |
538 // Capture the return input masks into each exit flavor | |
539 for( i=1; i < root->req(); i++ ) { | |
540 MachReturnNode *exit = root->in(i)->as_MachReturn(); | |
541 switch( exit->ideal_Opcode() ) { | |
542 case Op_Return : exit->_in_rms = ret_rms; break; | |
543 case Op_Rethrow : exit->_in_rms = reth_rms; break; | |
544 case Op_TailCall : exit->_in_rms = tail_call_rms; break; | |
545 case Op_TailJump : exit->_in_rms = tail_jump_rms; break; | |
546 case Op_Halt : exit->_in_rms = halt_rms; break; | |
547 default : ShouldNotReachHere(); | |
548 } | |
549 } | |
550 | |
551 // Next unused projection number from Start. | |
552 int proj_cnt = C->tf()->domain()->cnt(); | |
553 | |
554 // Do all the save-on-entry registers. Make projections from Start for | |
555 // them, and give them a use at the exit points. To the allocator, they | |
556 // look like incoming register arguments. | |
557 for( i = 0; i < _last_Mach_Reg; i++ ) { | |
558 if( is_save_on_entry(i) ) { | |
559 | |
560 // Add the save-on-entry to the mask array | |
561 ret_rms [ ret_edge_cnt] = mreg2regmask[i]; | |
562 reth_rms [ reth_edge_cnt] = mreg2regmask[i]; | |
563 tail_call_rms[tail_call_edge_cnt] = mreg2regmask[i]; | |
564 tail_jump_rms[tail_jump_edge_cnt] = mreg2regmask[i]; | |
565 // Halts need the SOE registers, but only in the stack as debug info. | |
566 // A just-prior uncommon-trap or deoptimization will use the SOE regs. | |
567 halt_rms [ halt_edge_cnt] = *idealreg2spillmask[_register_save_type[i]]; | |
568 | |
569 Node *mproj; | |
570 | |
571 // Is this a RegF low half of a RegD? Double up 2 adjacent RegF's | |
572 // into a single RegD. | |
573 if( (i&1) == 0 && | |
574 _register_save_type[i ] == Op_RegF && | |
575 _register_save_type[i+1] == Op_RegF && | |
576 is_save_on_entry(i+1) ) { | |
577 // Add other bit for double | |
578 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1)); | |
579 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1)); | |
580 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1)); | |
581 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1)); | |
582 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1)); | |
583 mproj = new (C, 1) MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegD ); | |
584 proj_cnt += 2; // Skip 2 for doubles | |
585 } | |
586 else if( (i&1) == 1 && // Else check for high half of double | |
587 _register_save_type[i-1] == Op_RegF && | |
588 _register_save_type[i ] == Op_RegF && | |
589 is_save_on_entry(i-1) ) { | |
590 ret_rms [ ret_edge_cnt] = RegMask::Empty; | |
591 reth_rms [ reth_edge_cnt] = RegMask::Empty; | |
592 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty; | |
593 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty; | |
594 halt_rms [ halt_edge_cnt] = RegMask::Empty; | |
595 mproj = C->top(); | |
596 } | |
597 // Is this a RegI low half of a RegL? Double up 2 adjacent RegI's | |
598 // into a single RegL. | |
599 else if( (i&1) == 0 && | |
600 _register_save_type[i ] == Op_RegI && | |
601 _register_save_type[i+1] == Op_RegI && | |
602 is_save_on_entry(i+1) ) { | |
603 // Add other bit for long | |
604 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1)); | |
605 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1)); | |
606 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1)); | |
607 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1)); | |
608 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1)); | |
609 mproj = new (C, 1) MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegL ); | |
610 proj_cnt += 2; // Skip 2 for longs | |
611 } | |
612 else if( (i&1) == 1 && // Else check for high half of long | |
613 _register_save_type[i-1] == Op_RegI && | |
614 _register_save_type[i ] == Op_RegI && | |
615 is_save_on_entry(i-1) ) { | |
616 ret_rms [ ret_edge_cnt] = RegMask::Empty; | |
617 reth_rms [ reth_edge_cnt] = RegMask::Empty; | |
618 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty; | |
619 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty; | |
620 halt_rms [ halt_edge_cnt] = RegMask::Empty; | |
621 mproj = C->top(); | |
622 } else { | |
623 // Make a projection for it off the Start | |
624 mproj = new (C, 1) MachProjNode( start, proj_cnt++, ret_rms[ret_edge_cnt], _register_save_type[i] ); | |
625 } | |
626 | |
627 ret_edge_cnt ++; | |
628 reth_edge_cnt ++; | |
629 tail_call_edge_cnt ++; | |
630 tail_jump_edge_cnt ++; | |
631 halt_edge_cnt ++; | |
632 | |
633 // Add a use of the SOE register to all exit paths | |
634 for( uint j=1; j < root->req(); j++ ) | |
635 root->in(j)->add_req(mproj); | |
636 } // End of if a save-on-entry register | |
637 } // End of for all machine registers | |
638 } | |
639 | |
640 //------------------------------init_spill_mask-------------------------------- | |
641 void Matcher::init_spill_mask( Node *ret ) { | |
642 if( idealreg2regmask[Op_RegI] ) return; // One time only init | |
643 | |
644 OptoReg::c_frame_pointer = c_frame_pointer(); | |
645 c_frame_ptr_mask = c_frame_pointer(); | |
646 #ifdef _LP64 | |
647 // pointers are twice as big | |
648 c_frame_ptr_mask.Insert(OptoReg::add(c_frame_pointer(),1)); | |
649 #endif | |
650 | |
651 // Start at OptoReg::stack0() | |
652 STACK_ONLY_mask.Clear(); | |
653 OptoReg::Name init = OptoReg::stack2reg(0); | |
654 // STACK_ONLY_mask is all stack bits | |
655 OptoReg::Name i; | |
656 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) | |
657 STACK_ONLY_mask.Insert(i); | |
658 // Also set the "infinite stack" bit. | |
659 STACK_ONLY_mask.set_AllStack(); | |
660 | |
661 // Copy the register names over into the shared world | |
662 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) { | |
663 // SharedInfo::regName[i] = regName[i]; | |
664 // Handy RegMasks per machine register | |
665 mreg2regmask[i].Insert(i); | |
666 } | |
667 | |
668 // Grab the Frame Pointer | |
669 Node *fp = ret->in(TypeFunc::FramePtr); | |
670 Node *mem = ret->in(TypeFunc::Memory); | |
671 const TypePtr* atp = TypePtr::BOTTOM; | |
672 // Share frame pointer while making spill ops | |
673 set_shared(fp); | |
674 | |
675 // Compute generic short-offset Loads | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
676 #ifdef _LP64 |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
677 MachNode *spillCP = match_tree(new (C, 3) LoadNNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM)); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
678 #endif |
0 | 679 MachNode *spillI = match_tree(new (C, 3) LoadINode(NULL,mem,fp,atp)); |
680 MachNode *spillL = match_tree(new (C, 3) LoadLNode(NULL,mem,fp,atp)); | |
681 MachNode *spillF = match_tree(new (C, 3) LoadFNode(NULL,mem,fp,atp)); | |
682 MachNode *spillD = match_tree(new (C, 3) LoadDNode(NULL,mem,fp,atp)); | |
683 MachNode *spillP = match_tree(new (C, 3) LoadPNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM)); | |
684 assert(spillI != NULL && spillL != NULL && spillF != NULL && | |
685 spillD != NULL && spillP != NULL, ""); | |
686 | |
687 // Get the ADLC notion of the right regmask, for each basic type. | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
688 #ifdef _LP64 |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
689 idealreg2regmask[Op_RegN] = &spillCP->out_RegMask(); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
690 #endif |
0 | 691 idealreg2regmask[Op_RegI] = &spillI->out_RegMask(); |
692 idealreg2regmask[Op_RegL] = &spillL->out_RegMask(); | |
693 idealreg2regmask[Op_RegF] = &spillF->out_RegMask(); | |
694 idealreg2regmask[Op_RegD] = &spillD->out_RegMask(); | |
695 idealreg2regmask[Op_RegP] = &spillP->out_RegMask(); | |
696 } | |
697 | |
698 #ifdef ASSERT | |
699 static void match_alias_type(Compile* C, Node* n, Node* m) { | |
700 if (!VerifyAliases) return; // do not go looking for trouble by default | |
701 const TypePtr* nat = n->adr_type(); | |
702 const TypePtr* mat = m->adr_type(); | |
703 int nidx = C->get_alias_index(nat); | |
704 int midx = C->get_alias_index(mat); | |
705 // Detune the assert for cases like (AndI 0xFF (LoadB p)). | |
706 if (nidx == Compile::AliasIdxTop && midx >= Compile::AliasIdxRaw) { | |
707 for (uint i = 1; i < n->req(); i++) { | |
708 Node* n1 = n->in(i); | |
709 const TypePtr* n1at = n1->adr_type(); | |
710 if (n1at != NULL) { | |
711 nat = n1at; | |
712 nidx = C->get_alias_index(n1at); | |
713 } | |
714 } | |
715 } | |
716 // %%% Kludgery. Instead, fix ideal adr_type methods for all these cases: | |
717 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxRaw) { | |
718 switch (n->Opcode()) { | |
719 case Op_PrefetchRead: | |
720 case Op_PrefetchWrite: | |
721 nidx = Compile::AliasIdxRaw; | |
722 nat = TypeRawPtr::BOTTOM; | |
723 break; | |
724 } | |
725 } | |
726 if (nidx == Compile::AliasIdxRaw && midx == Compile::AliasIdxTop) { | |
727 switch (n->Opcode()) { | |
728 case Op_ClearArray: | |
729 midx = Compile::AliasIdxRaw; | |
730 mat = TypeRawPtr::BOTTOM; | |
731 break; | |
732 } | |
733 } | |
734 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxBot) { | |
735 switch (n->Opcode()) { | |
736 case Op_Return: | |
737 case Op_Rethrow: | |
738 case Op_Halt: | |
739 case Op_TailCall: | |
740 case Op_TailJump: | |
741 nidx = Compile::AliasIdxBot; | |
742 nat = TypePtr::BOTTOM; | |
743 break; | |
744 } | |
745 } | |
746 if (nidx == Compile::AliasIdxBot && midx == Compile::AliasIdxTop) { | |
747 switch (n->Opcode()) { | |
748 case Op_StrComp: | |
681 | 749 case Op_StrEquals: |
750 case Op_StrIndexOf: | |
169
9148c65abefc
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
168
diff
changeset
|
751 case Op_AryEq: |
0 | 752 case Op_MemBarVolatile: |
753 case Op_MemBarCPUOrder: // %%% these ideals should have narrower adr_type? | |
754 nidx = Compile::AliasIdxTop; | |
755 nat = NULL; | |
756 break; | |
757 } | |
758 } | |
759 if (nidx != midx) { | |
760 if (PrintOpto || (PrintMiscellaneous && (WizardMode || Verbose))) { | |
761 tty->print_cr("==== Matcher alias shift %d => %d", nidx, midx); | |
762 n->dump(); | |
763 m->dump(); | |
764 } | |
765 assert(C->subsume_loads() && C->must_alias(nat, midx), | |
766 "must not lose alias info when matching"); | |
767 } | |
768 } | |
769 #endif | |
770 | |
771 | |
772 //------------------------------MStack----------------------------------------- | |
773 // State and MStack class used in xform() and find_shared() iterative methods. | |
774 enum Node_State { Pre_Visit, // node has to be pre-visited | |
775 Visit, // visit node | |
776 Post_Visit, // post-visit node | |
777 Alt_Post_Visit // alternative post-visit path | |
778 }; | |
779 | |
780 class MStack: public Node_Stack { | |
781 public: | |
782 MStack(int size) : Node_Stack(size) { } | |
783 | |
784 void push(Node *n, Node_State ns) { | |
785 Node_Stack::push(n, (uint)ns); | |
786 } | |
787 void push(Node *n, Node_State ns, Node *parent, int indx) { | |
788 ++_inode_top; | |
789 if ((_inode_top + 1) >= _inode_max) grow(); | |
790 _inode_top->node = parent; | |
791 _inode_top->indx = (uint)indx; | |
792 ++_inode_top; | |
793 _inode_top->node = n; | |
794 _inode_top->indx = (uint)ns; | |
795 } | |
796 Node *parent() { | |
797 pop(); | |
798 return node(); | |
799 } | |
800 Node_State state() const { | |
801 return (Node_State)index(); | |
802 } | |
803 void set_state(Node_State ns) { | |
804 set_index((uint)ns); | |
805 } | |
806 }; | |
807 | |
808 | |
809 //------------------------------xform------------------------------------------ | |
810 // Given a Node in old-space, Match him (Label/Reduce) to produce a machine | |
811 // Node in new-space. Given a new-space Node, recursively walk his children. | |
812 Node *Matcher::transform( Node *n ) { ShouldNotCallThis(); return n; } | |
813 Node *Matcher::xform( Node *n, int max_stack ) { | |
814 // Use one stack to keep both: child's node/state and parent's node/index | |
815 MStack mstack(max_stack * 2 * 2); // C->unique() * 2 * 2 | |
816 mstack.push(n, Visit, NULL, -1); // set NULL as parent to indicate root | |
817 | |
818 while (mstack.is_nonempty()) { | |
819 n = mstack.node(); // Leave node on stack | |
820 Node_State nstate = mstack.state(); | |
821 if (nstate == Visit) { | |
822 mstack.set_state(Post_Visit); | |
823 Node *oldn = n; | |
824 // Old-space or new-space check | |
825 if (!C->node_arena()->contains(n)) { | |
826 // Old space! | |
827 Node* m; | |
828 if (has_new_node(n)) { // Not yet Label/Reduced | |
829 m = new_node(n); | |
830 } else { | |
831 if (!is_dontcare(n)) { // Matcher can match this guy | |
832 // Calls match special. They match alone with no children. | |
833 // Their children, the incoming arguments, match normally. | |
834 m = n->is_SafePoint() ? match_sfpt(n->as_SafePoint()):match_tree(n); | |
835 if (C->failing()) return NULL; | |
836 if (m == NULL) { Matcher::soft_match_failure(); return NULL; } | |
837 } else { // Nothing the matcher cares about | |
838 if( n->is_Proj() && n->in(0)->is_Multi()) { // Projections? | |
839 // Convert to machine-dependent projection | |
840 m = n->in(0)->as_Multi()->match( n->as_Proj(), this ); | |
222 | 841 #ifdef ASSERT |
842 _new2old_map.map(m->_idx, n); | |
843 #endif | |
0 | 844 if (m->in(0) != NULL) // m might be top |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
845 collect_null_checks(m, n); |
0 | 846 } else { // Else just a regular 'ol guy |
847 m = n->clone(); // So just clone into new-space | |
222 | 848 #ifdef ASSERT |
849 _new2old_map.map(m->_idx, n); | |
850 #endif | |
0 | 851 // Def-Use edges will be added incrementally as Uses |
852 // of this node are matched. | |
853 assert(m->outcnt() == 0, "no Uses of this clone yet"); | |
854 } | |
855 } | |
856 | |
857 set_new_node(n, m); // Map old to new | |
858 if (_old_node_note_array != NULL) { | |
859 Node_Notes* nn = C->locate_node_notes(_old_node_note_array, | |
860 n->_idx); | |
861 C->set_node_notes_at(m->_idx, nn); | |
862 } | |
863 debug_only(match_alias_type(C, n, m)); | |
864 } | |
865 n = m; // n is now a new-space node | |
866 mstack.set_node(n); | |
867 } | |
868 | |
869 // New space! | |
870 if (_visited.test_set(n->_idx)) continue; // while(mstack.is_nonempty()) | |
871 | |
872 int i; | |
873 // Put precedence edges on stack first (match them last). | |
874 for (i = oldn->req(); (uint)i < oldn->len(); i++) { | |
875 Node *m = oldn->in(i); | |
876 if (m == NULL) break; | |
877 // set -1 to call add_prec() instead of set_req() during Step1 | |
878 mstack.push(m, Visit, n, -1); | |
879 } | |
880 | |
881 // For constant debug info, I'd rather have unmatched constants. | |
882 int cnt = n->req(); | |
883 JVMState* jvms = n->jvms(); | |
884 int debug_cnt = jvms ? jvms->debug_start() : cnt; | |
885 | |
886 // Now do only debug info. Clone constants rather than matching. | |
887 // Constants are represented directly in the debug info without | |
888 // the need for executable machine instructions. | |
889 // Monitor boxes are also represented directly. | |
890 for (i = cnt - 1; i >= debug_cnt; --i) { // For all debug inputs do | |
891 Node *m = n->in(i); // Get input | |
892 int op = m->Opcode(); | |
893 assert((op == Op_BoxLock) == jvms->is_monitor_use(i), "boxes only at monitor sites"); | |
163 | 894 if( op == Op_ConI || op == Op_ConP || op == Op_ConN || |
0 | 895 op == Op_ConF || op == Op_ConD || op == Op_ConL |
896 // || op == Op_BoxLock // %%%% enable this and remove (+++) in chaitin.cpp | |
897 ) { | |
898 m = m->clone(); | |
222 | 899 #ifdef ASSERT |
900 _new2old_map.map(m->_idx, n); | |
901 #endif | |
605 | 902 mstack.push(m, Post_Visit, n, i); // Don't need to visit |
0 | 903 mstack.push(m->in(0), Visit, m, 0); |
904 } else { | |
905 mstack.push(m, Visit, n, i); | |
906 } | |
907 } | |
908 | |
909 // And now walk his children, and convert his inputs to new-space. | |
910 for( ; i >= 0; --i ) { // For all normal inputs do | |
911 Node *m = n->in(i); // Get input | |
912 if(m != NULL) | |
913 mstack.push(m, Visit, n, i); | |
914 } | |
915 | |
916 } | |
917 else if (nstate == Post_Visit) { | |
918 // Set xformed input | |
919 Node *p = mstack.parent(); | |
920 if (p != NULL) { // root doesn't have parent | |
921 int i = (int)mstack.index(); | |
922 if (i >= 0) | |
923 p->set_req(i, n); // required input | |
924 else if (i == -1) | |
925 p->add_prec(n); // precedence input | |
926 else | |
927 ShouldNotReachHere(); | |
928 } | |
929 mstack.pop(); // remove processed node from stack | |
930 } | |
931 else { | |
932 ShouldNotReachHere(); | |
933 } | |
934 } // while (mstack.is_nonempty()) | |
935 return n; // Return new-space Node | |
936 } | |
937 | |
938 //------------------------------warp_outgoing_stk_arg------------------------ | |
939 OptoReg::Name Matcher::warp_outgoing_stk_arg( VMReg reg, OptoReg::Name begin_out_arg_area, OptoReg::Name &out_arg_limit_per_call ) { | |
940 // Convert outgoing argument location to a pre-biased stack offset | |
941 if (reg->is_stack()) { | |
942 OptoReg::Name warped = reg->reg2stack(); | |
943 // Adjust the stack slot offset to be the register number used | |
944 // by the allocator. | |
945 warped = OptoReg::add(begin_out_arg_area, warped); | |
946 // Keep track of the largest numbered stack slot used for an arg. | |
947 // Largest used slot per call-site indicates the amount of stack | |
948 // that is killed by the call. | |
949 if( warped >= out_arg_limit_per_call ) | |
950 out_arg_limit_per_call = OptoReg::add(warped,1); | |
951 if (!RegMask::can_represent(warped)) { | |
952 C->record_method_not_compilable_all_tiers("unsupported calling sequence"); | |
953 return OptoReg::Bad; | |
954 } | |
955 return warped; | |
956 } | |
957 return OptoReg::as_OptoReg(reg); | |
958 } | |
959 | |
960 | |
961 //------------------------------match_sfpt------------------------------------- | |
962 // Helper function to match call instructions. Calls match special. | |
963 // They match alone with no children. Their children, the incoming | |
964 // arguments, match normally. | |
965 MachNode *Matcher::match_sfpt( SafePointNode *sfpt ) { | |
966 MachSafePointNode *msfpt = NULL; | |
967 MachCallNode *mcall = NULL; | |
968 uint cnt; | |
969 // Split out case for SafePoint vs Call | |
970 CallNode *call; | |
971 const TypeTuple *domain; | |
972 ciMethod* method = NULL; | |
973 if( sfpt->is_Call() ) { | |
974 call = sfpt->as_Call(); | |
975 domain = call->tf()->domain(); | |
976 cnt = domain->cnt(); | |
977 | |
978 // Match just the call, nothing else | |
979 MachNode *m = match_tree(call); | |
980 if (C->failing()) return NULL; | |
981 if( m == NULL ) { Matcher::soft_match_failure(); return NULL; } | |
982 | |
983 // Copy data from the Ideal SafePoint to the machine version | |
984 mcall = m->as_MachCall(); | |
985 | |
986 mcall->set_tf( call->tf()); | |
987 mcall->set_entry_point(call->entry_point()); | |
988 mcall->set_cnt( call->cnt()); | |
989 | |
990 if( mcall->is_MachCallJava() ) { | |
991 MachCallJavaNode *mcall_java = mcall->as_MachCallJava(); | |
992 const CallJavaNode *call_java = call->as_CallJava(); | |
993 method = call_java->method(); | |
994 mcall_java->_method = method; | |
995 mcall_java->_bci = call_java->_bci; | |
996 mcall_java->_optimized_virtual = call_java->is_optimized_virtual(); | |
997 if( mcall_java->is_MachCallStaticJava() ) | |
998 mcall_java->as_MachCallStaticJava()->_name = | |
999 call_java->as_CallStaticJava()->_name; | |
1000 if( mcall_java->is_MachCallDynamicJava() ) | |
1001 mcall_java->as_MachCallDynamicJava()->_vtable_index = | |
1002 call_java->as_CallDynamicJava()->_vtable_index; | |
1003 } | |
1004 else if( mcall->is_MachCallRuntime() ) { | |
1005 mcall->as_MachCallRuntime()->_name = call->as_CallRuntime()->_name; | |
1006 } | |
1007 msfpt = mcall; | |
1008 } | |
1009 // This is a non-call safepoint | |
1010 else { | |
1011 call = NULL; | |
1012 domain = NULL; | |
1013 MachNode *mn = match_tree(sfpt); | |
1014 if (C->failing()) return NULL; | |
1015 msfpt = mn->as_MachSafePoint(); | |
1016 cnt = TypeFunc::Parms; | |
1017 } | |
1018 | |
1019 // Advertise the correct memory effects (for anti-dependence computation). | |
1020 msfpt->set_adr_type(sfpt->adr_type()); | |
1021 | |
1022 // Allocate a private array of RegMasks. These RegMasks are not shared. | |
1023 msfpt->_in_rms = NEW_RESOURCE_ARRAY( RegMask, cnt ); | |
1024 // Empty them all. | |
1025 memset( msfpt->_in_rms, 0, sizeof(RegMask)*cnt ); | |
1026 | |
1027 // Do all the pre-defined non-Empty register masks | |
1028 msfpt->_in_rms[TypeFunc::ReturnAdr] = _return_addr_mask; | |
1029 msfpt->_in_rms[TypeFunc::FramePtr ] = c_frame_ptr_mask; | |
1030 | |
1031 // Place first outgoing argument can possibly be put. | |
1032 OptoReg::Name begin_out_arg_area = OptoReg::add(_new_SP, C->out_preserve_stack_slots()); | |
1033 assert( is_even(begin_out_arg_area), "" ); | |
1034 // Compute max outgoing register number per call site. | |
1035 OptoReg::Name out_arg_limit_per_call = begin_out_arg_area; | |
1036 // Calls to C may hammer extra stack slots above and beyond any arguments. | |
1037 // These are usually backing store for register arguments for varargs. | |
1038 if( call != NULL && call->is_CallRuntime() ) | |
1039 out_arg_limit_per_call = OptoReg::add(out_arg_limit_per_call,C->varargs_C_out_slots_killed()); | |
1040 | |
1041 | |
1042 // Do the normal argument list (parameters) register masks | |
1043 int argcnt = cnt - TypeFunc::Parms; | |
1044 if( argcnt > 0 ) { // Skip it all if we have no args | |
1045 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt ); | |
1046 VMRegPair *parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt ); | |
1047 int i; | |
1048 for( i = 0; i < argcnt; i++ ) { | |
1049 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type(); | |
1050 } | |
1051 // V-call to pick proper calling convention | |
1052 call->calling_convention( sig_bt, parm_regs, argcnt ); | |
1053 | |
1054 #ifdef ASSERT | |
1055 // Sanity check users' calling convention. Really handy during | |
1056 // the initial porting effort. Fairly expensive otherwise. | |
1057 { for (int i = 0; i<argcnt; i++) { | |
1058 if( !parm_regs[i].first()->is_valid() && | |
1059 !parm_regs[i].second()->is_valid() ) continue; | |
1060 VMReg reg1 = parm_regs[i].first(); | |
1061 VMReg reg2 = parm_regs[i].second(); | |
1062 for (int j = 0; j < i; j++) { | |
1063 if( !parm_regs[j].first()->is_valid() && | |
1064 !parm_regs[j].second()->is_valid() ) continue; | |
1065 VMReg reg3 = parm_regs[j].first(); | |
1066 VMReg reg4 = parm_regs[j].second(); | |
1067 if( !reg1->is_valid() ) { | |
1068 assert( !reg2->is_valid(), "valid halvsies" ); | |
1069 } else if( !reg3->is_valid() ) { | |
1070 assert( !reg4->is_valid(), "valid halvsies" ); | |
1071 } else { | |
1072 assert( reg1 != reg2, "calling conv. must produce distinct regs"); | |
1073 assert( reg1 != reg3, "calling conv. must produce distinct regs"); | |
1074 assert( reg1 != reg4, "calling conv. must produce distinct regs"); | |
1075 assert( reg2 != reg3, "calling conv. must produce distinct regs"); | |
1076 assert( reg2 != reg4 || !reg2->is_valid(), "calling conv. must produce distinct regs"); | |
1077 assert( reg3 != reg4, "calling conv. must produce distinct regs"); | |
1078 } | |
1079 } | |
1080 } | |
1081 } | |
1082 #endif | |
1083 | |
1084 // Visit each argument. Compute its outgoing register mask. | |
1085 // Return results now can have 2 bits returned. | |
1086 // Compute max over all outgoing arguments both per call-site | |
1087 // and over the entire method. | |
1088 for( i = 0; i < argcnt; i++ ) { | |
1089 // Address of incoming argument mask to fill in | |
1090 RegMask *rm = &mcall->_in_rms[i+TypeFunc::Parms]; | |
1091 if( !parm_regs[i].first()->is_valid() && | |
1092 !parm_regs[i].second()->is_valid() ) { | |
1093 continue; // Avoid Halves | |
1094 } | |
1095 // Grab first register, adjust stack slots and insert in mask. | |
1096 OptoReg::Name reg1 = warp_outgoing_stk_arg(parm_regs[i].first(), begin_out_arg_area, out_arg_limit_per_call ); | |
1097 if (OptoReg::is_valid(reg1)) | |
1098 rm->Insert( reg1 ); | |
1099 // Grab second register (if any), adjust stack slots and insert in mask. | |
1100 OptoReg::Name reg2 = warp_outgoing_stk_arg(parm_regs[i].second(), begin_out_arg_area, out_arg_limit_per_call ); | |
1101 if (OptoReg::is_valid(reg2)) | |
1102 rm->Insert( reg2 ); | |
1103 } // End of for all arguments | |
1104 | |
1105 // Compute number of stack slots needed to restore stack in case of | |
1106 // Pascal-style argument popping. | |
1107 mcall->_argsize = out_arg_limit_per_call - begin_out_arg_area; | |
1108 } | |
1109 | |
1110 // Compute the max stack slot killed by any call. These will not be | |
1111 // available for debug info, and will be used to adjust FIRST_STACK_mask | |
1112 // after all call sites have been visited. | |
1113 if( _out_arg_limit < out_arg_limit_per_call) | |
1114 _out_arg_limit = out_arg_limit_per_call; | |
1115 | |
1116 if (mcall) { | |
1117 // Kill the outgoing argument area, including any non-argument holes and | |
1118 // any legacy C-killed slots. Use Fat-Projections to do the killing. | |
1119 // Since the max-per-method covers the max-per-call-site and debug info | |
1120 // is excluded on the max-per-method basis, debug info cannot land in | |
1121 // this killed area. | |
1122 uint r_cnt = mcall->tf()->range()->cnt(); | |
1123 MachProjNode *proj = new (C, 1) MachProjNode( mcall, r_cnt+10000, RegMask::Empty, MachProjNode::fat_proj ); | |
1124 if (!RegMask::can_represent(OptoReg::Name(out_arg_limit_per_call-1))) { | |
1125 C->record_method_not_compilable_all_tiers("unsupported outgoing calling sequence"); | |
1126 } else { | |
1127 for (int i = begin_out_arg_area; i < out_arg_limit_per_call; i++) | |
1128 proj->_rout.Insert(OptoReg::Name(i)); | |
1129 } | |
1130 if( proj->_rout.is_NotEmpty() ) | |
1131 _proj_list.push(proj); | |
1132 } | |
1133 // Transfer the safepoint information from the call to the mcall | |
1134 // Move the JVMState list | |
1135 msfpt->set_jvms(sfpt->jvms()); | |
1136 for (JVMState* jvms = msfpt->jvms(); jvms; jvms = jvms->caller()) { | |
1137 jvms->set_map(sfpt); | |
1138 } | |
1139 | |
1140 // Debug inputs begin just after the last incoming parameter | |
1141 assert( (mcall == NULL) || (mcall->jvms() == NULL) || | |
1142 (mcall->jvms()->debug_start() + mcall->_jvmadj == mcall->tf()->domain()->cnt()), "" ); | |
1143 | |
1144 // Move the OopMap | |
1145 msfpt->_oop_map = sfpt->_oop_map; | |
1146 | |
1147 // Registers killed by the call are set in the local scheduling pass | |
1148 // of Global Code Motion. | |
1149 return msfpt; | |
1150 } | |
1151 | |
1152 //---------------------------match_tree---------------------------------------- | |
1153 // Match a Ideal Node DAG - turn it into a tree; Label & Reduce. Used as part | |
1154 // of the whole-sale conversion from Ideal to Mach Nodes. Also used for | |
1155 // making GotoNodes while building the CFG and in init_spill_mask() to identify | |
1156 // a Load's result RegMask for memoization in idealreg2regmask[] | |
1157 MachNode *Matcher::match_tree( const Node *n ) { | |
1158 assert( n->Opcode() != Op_Phi, "cannot match" ); | |
1159 assert( !n->is_block_start(), "cannot match" ); | |
1160 // Set the mark for all locally allocated State objects. | |
1161 // When this call returns, the _states_arena arena will be reset | |
1162 // freeing all State objects. | |
1163 ResourceMark rm( &_states_arena ); | |
1164 | |
1165 LabelRootDepth = 0; | |
1166 | |
1167 // StoreNodes require their Memory input to match any LoadNodes | |
1168 Node *mem = n->is_Store() ? n->in(MemNode::Memory) : (Node*)1 ; | |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1169 #ifdef ASSERT |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1170 Node* save_mem_node = _mem_node; |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1171 _mem_node = n->is_Store() ? (Node*)n : NULL; |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1172 #endif |
0 | 1173 // State object for root node of match tree |
1174 // Allocate it on _states_arena - stack allocation can cause stack overflow. | |
1175 State *s = new (&_states_arena) State; | |
1176 s->_kids[0] = NULL; | |
1177 s->_kids[1] = NULL; | |
1178 s->_leaf = (Node*)n; | |
1179 // Label the input tree, allocating labels from top-level arena | |
1180 Label_Root( n, s, n->in(0), mem ); | |
1181 if (C->failing()) return NULL; | |
1182 | |
1183 // The minimum cost match for the whole tree is found at the root State | |
1184 uint mincost = max_juint; | |
1185 uint cost = max_juint; | |
1186 uint i; | |
1187 for( i = 0; i < NUM_OPERANDS; i++ ) { | |
1188 if( s->valid(i) && // valid entry and | |
1189 s->_cost[i] < cost && // low cost and | |
1190 s->_rule[i] >= NUM_OPERANDS ) // not an operand | |
1191 cost = s->_cost[mincost=i]; | |
1192 } | |
1193 if (mincost == max_juint) { | |
1194 #ifndef PRODUCT | |
1195 tty->print("No matching rule for:"); | |
1196 s->dump(); | |
1197 #endif | |
1198 Matcher::soft_match_failure(); | |
1199 return NULL; | |
1200 } | |
1201 // Reduce input tree based upon the state labels to machine Nodes | |
1202 MachNode *m = ReduceInst( s, s->_rule[mincost], mem ); | |
1203 #ifdef ASSERT | |
1204 _old2new_map.map(n->_idx, m); | |
222 | 1205 _new2old_map.map(m->_idx, (Node*)n); |
0 | 1206 #endif |
1207 | |
1208 // Add any Matcher-ignored edges | |
1209 uint cnt = n->req(); | |
1210 uint start = 1; | |
1211 if( mem != (Node*)1 ) start = MemNode::Memory+1; | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1212 if( n->is_AddP() ) { |
0 | 1213 assert( mem == (Node*)1, "" ); |
1214 start = AddPNode::Base+1; | |
1215 } | |
1216 for( i = start; i < cnt; i++ ) { | |
1217 if( !n->match_edge(i) ) { | |
1218 if( i < m->req() ) | |
1219 m->ins_req( i, n->in(i) ); | |
1220 else | |
1221 m->add_req( n->in(i) ); | |
1222 } | |
1223 } | |
1224 | |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1225 debug_only( _mem_node = save_mem_node; ) |
0 | 1226 return m; |
1227 } | |
1228 | |
1229 | |
1230 //------------------------------match_into_reg--------------------------------- | |
1231 // Choose to either match this Node in a register or part of the current | |
1232 // match tree. Return true for requiring a register and false for matching | |
1233 // as part of the current match tree. | |
1234 static bool match_into_reg( const Node *n, Node *m, Node *control, int i, bool shared ) { | |
1235 | |
1236 const Type *t = m->bottom_type(); | |
1237 | |
1238 if( t->singleton() ) { | |
1239 // Never force constants into registers. Allow them to match as | |
1240 // constants or registers. Copies of the same value will share | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1241 // the same register. See find_shared_node. |
0 | 1242 return false; |
1243 } else { // Not a constant | |
1244 // Stop recursion if they have different Controls. | |
1245 // Slot 0 of constants is not really a Control. | |
1246 if( control && m->in(0) && control != m->in(0) ) { | |
1247 | |
1248 // Actually, we can live with the most conservative control we | |
1249 // find, if it post-dominates the others. This allows us to | |
1250 // pick up load/op/store trees where the load can float a little | |
1251 // above the store. | |
1252 Node *x = control; | |
1253 const uint max_scan = 6; // Arbitrary scan cutoff | |
1254 uint j; | |
1255 for( j=0; j<max_scan; j++ ) { | |
1256 if( x->is_Region() ) // Bail out at merge points | |
1257 return true; | |
1258 x = x->in(0); | |
1259 if( x == m->in(0) ) // Does 'control' post-dominate | |
1260 break; // m->in(0)? If so, we can use it | |
1261 } | |
1262 if( j == max_scan ) // No post-domination before scan end? | |
1263 return true; // Then break the match tree up | |
1264 } | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1265 if (m->is_DecodeN() && Matcher::clone_shift_expressions) { |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
1266 // These are commonly used in address expressions and can |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1267 // efficiently fold into them on X64 in some cases. |
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1268 return false; |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
1269 } |
0 | 1270 } |
1271 | |
605 | 1272 // Not forceable cloning. If shared, put it into a register. |
0 | 1273 return shared; |
1274 } | |
1275 | |
1276 | |
1277 //------------------------------Instruction Selection-------------------------- | |
1278 // Label method walks a "tree" of nodes, using the ADLC generated DFA to match | |
1279 // ideal nodes to machine instructions. Trees are delimited by shared Nodes, | |
1280 // things the Matcher does not match (e.g., Memory), and things with different | |
1281 // Controls (hence forced into different blocks). We pass in the Control | |
1282 // selected for this entire State tree. | |
1283 | |
1284 // The Matcher works on Trees, but an Intel add-to-memory requires a DAG: the | |
1285 // Store and the Load must have identical Memories (as well as identical | |
1286 // pointers). Since the Matcher does not have anything for Memory (and | |
1287 // does not handle DAGs), I have to match the Memory input myself. If the | |
1288 // Tree root is a Store, I require all Loads to have the identical memory. | |
1289 Node *Matcher::Label_Root( const Node *n, State *svec, Node *control, const Node *mem){ | |
1290 // Since Label_Root is a recursive function, its possible that we might run | |
1291 // out of stack space. See bugs 6272980 & 6227033 for more info. | |
1292 LabelRootDepth++; | |
1293 if (LabelRootDepth > MaxLabelRootDepth) { | |
1294 C->record_method_not_compilable_all_tiers("Out of stack space, increase MaxLabelRootDepth"); | |
1295 return NULL; | |
1296 } | |
1297 uint care = 0; // Edges matcher cares about | |
1298 uint cnt = n->req(); | |
1299 uint i = 0; | |
1300 | |
1301 // Examine children for memory state | |
1302 // Can only subsume a child into your match-tree if that child's memory state | |
1303 // is not modified along the path to another input. | |
1304 // It is unsafe even if the other inputs are separate roots. | |
1305 Node *input_mem = NULL; | |
1306 for( i = 1; i < cnt; i++ ) { | |
1307 if( !n->match_edge(i) ) continue; | |
1308 Node *m = n->in(i); // Get ith input | |
1309 assert( m, "expect non-null children" ); | |
1310 if( m->is_Load() ) { | |
1311 if( input_mem == NULL ) { | |
1312 input_mem = m->in(MemNode::Memory); | |
1313 } else if( input_mem != m->in(MemNode::Memory) ) { | |
1314 input_mem = NodeSentinel; | |
1315 } | |
1316 } | |
1317 } | |
1318 | |
1319 for( i = 1; i < cnt; i++ ){// For my children | |
1320 if( !n->match_edge(i) ) continue; | |
1321 Node *m = n->in(i); // Get ith input | |
1322 // Allocate states out of a private arena | |
1323 State *s = new (&_states_arena) State; | |
1324 svec->_kids[care++] = s; | |
1325 assert( care <= 2, "binary only for now" ); | |
1326 | |
1327 // Recursively label the State tree. | |
1328 s->_kids[0] = NULL; | |
1329 s->_kids[1] = NULL; | |
1330 s->_leaf = m; | |
1331 | |
1332 // Check for leaves of the State Tree; things that cannot be a part of | |
1333 // the current tree. If it finds any, that value is matched as a | |
1334 // register operand. If not, then the normal matching is used. | |
1335 if( match_into_reg(n, m, control, i, is_shared(m)) || | |
1336 // | |
1337 // Stop recursion if this is LoadNode and the root of this tree is a | |
1338 // StoreNode and the load & store have different memories. | |
1339 ((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) || | |
1340 // Can NOT include the match of a subtree when its memory state | |
1341 // is used by any of the other subtrees | |
1342 (input_mem == NodeSentinel) ) { | |
1343 #ifndef PRODUCT | |
1344 // Print when we exclude matching due to different memory states at input-loads | |
1345 if( PrintOpto && (Verbose && WizardMode) && (input_mem == NodeSentinel) | |
1346 && !((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) ) { | |
1347 tty->print_cr("invalid input_mem"); | |
1348 } | |
1349 #endif | |
1350 // Switch to a register-only opcode; this value must be in a register | |
1351 // and cannot be subsumed as part of a larger instruction. | |
1352 s->DFA( m->ideal_reg(), m ); | |
1353 | |
1354 } else { | |
1355 // If match tree has no control and we do, adopt it for entire tree | |
1356 if( control == NULL && m->in(0) != NULL && m->req() > 1 ) | |
1357 control = m->in(0); // Pick up control | |
1358 // Else match as a normal part of the match tree. | |
1359 control = Label_Root(m,s,control,mem); | |
1360 if (C->failing()) return NULL; | |
1361 } | |
1362 } | |
1363 | |
1364 | |
1365 // Call DFA to match this node, and return | |
1366 svec->DFA( n->Opcode(), n ); | |
1367 | |
1368 #ifdef ASSERT | |
1369 uint x; | |
1370 for( x = 0; x < _LAST_MACH_OPER; x++ ) | |
1371 if( svec->valid(x) ) | |
1372 break; | |
1373 | |
1374 if (x >= _LAST_MACH_OPER) { | |
1375 n->dump(); | |
1376 svec->dump(); | |
1377 assert( false, "bad AD file" ); | |
1378 } | |
1379 #endif | |
1380 return control; | |
1381 } | |
1382 | |
1383 | |
1384 // Con nodes reduced using the same rule can share their MachNode | |
1385 // which reduces the number of copies of a constant in the final | |
1386 // program. The register allocator is free to split uses later to | |
1387 // split live ranges. | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1388 MachNode* Matcher::find_shared_node(Node* leaf, uint rule) { |
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1389 if (!leaf->is_Con() && !leaf->is_DecodeN()) return NULL; |
0 | 1390 |
1391 // See if this Con has already been reduced using this rule. | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1392 if (_shared_nodes.Size() <= leaf->_idx) return NULL; |
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1393 MachNode* last = (MachNode*)_shared_nodes.at(leaf->_idx); |
0 | 1394 if (last != NULL && rule == last->rule()) { |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1395 // Don't expect control change for DecodeN |
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1396 if (leaf->is_DecodeN()) |
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1397 return last; |
0 | 1398 // Get the new space root. |
1399 Node* xroot = new_node(C->root()); | |
1400 if (xroot == NULL) { | |
1401 // This shouldn't happen give the order of matching. | |
1402 return NULL; | |
1403 } | |
1404 | |
1405 // Shared constants need to have their control be root so they | |
1406 // can be scheduled properly. | |
1407 Node* control = last->in(0); | |
1408 if (control != xroot) { | |
1409 if (control == NULL || control == C->root()) { | |
1410 last->set_req(0, xroot); | |
1411 } else { | |
1412 assert(false, "unexpected control"); | |
1413 return NULL; | |
1414 } | |
1415 } | |
1416 return last; | |
1417 } | |
1418 return NULL; | |
1419 } | |
1420 | |
1421 | |
1422 //------------------------------ReduceInst------------------------------------- | |
1423 // Reduce a State tree (with given Control) into a tree of MachNodes. | |
1424 // This routine (and it's cohort ReduceOper) convert Ideal Nodes into | |
1425 // complicated machine Nodes. Each MachNode covers some tree of Ideal Nodes. | |
1426 // Each MachNode has a number of complicated MachOper operands; each | |
1427 // MachOper also covers a further tree of Ideal Nodes. | |
1428 | |
1429 // The root of the Ideal match tree is always an instruction, so we enter | |
1430 // the recursion here. After building the MachNode, we need to recurse | |
1431 // the tree checking for these cases: | |
1432 // (1) Child is an instruction - | |
1433 // Build the instruction (recursively), add it as an edge. | |
1434 // Build a simple operand (register) to hold the result of the instruction. | |
1435 // (2) Child is an interior part of an instruction - | |
1436 // Skip over it (do nothing) | |
1437 // (3) Child is the start of a operand - | |
1438 // Build the operand, place it inside the instruction | |
1439 // Call ReduceOper. | |
1440 MachNode *Matcher::ReduceInst( State *s, int rule, Node *&mem ) { | |
1441 assert( rule >= NUM_OPERANDS, "called with operand rule" ); | |
1442 | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1443 MachNode* shared_node = find_shared_node(s->_leaf, rule); |
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1444 if (shared_node != NULL) { |
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1445 return shared_node; |
0 | 1446 } |
1447 | |
1448 // Build the object to represent this state & prepare for recursive calls | |
1449 MachNode *mach = s->MachNodeGenerator( rule, C ); | |
1450 mach->_opnds[0] = s->MachOperGenerator( _reduceOp[rule], C ); | |
1451 assert( mach->_opnds[0] != NULL, "Missing result operand" ); | |
1452 Node *leaf = s->_leaf; | |
1453 // Check for instruction or instruction chain rule | |
1454 if( rule >= _END_INST_CHAIN_RULE || rule < _BEGIN_INST_CHAIN_RULE ) { | |
309
eaf496ad4a14
6732698: crash with dead code from compressed oops in gcm
never
parents:
235
diff
changeset
|
1455 assert(C->node_arena()->contains(s->_leaf) || !has_new_node(s->_leaf), |
eaf496ad4a14
6732698: crash with dead code from compressed oops in gcm
never
parents:
235
diff
changeset
|
1456 "duplicating node that's already been matched"); |
0 | 1457 // Instruction |
1458 mach->add_req( leaf->in(0) ); // Set initial control | |
1459 // Reduce interior of complex instruction | |
1460 ReduceInst_Interior( s, rule, mem, mach, 1 ); | |
1461 } else { | |
1462 // Instruction chain rules are data-dependent on their inputs | |
1463 mach->add_req(0); // Set initial control to none | |
1464 ReduceInst_Chain_Rule( s, rule, mem, mach ); | |
1465 } | |
1466 | |
1467 // If a Memory was used, insert a Memory edge | |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1468 if( mem != (Node*)1 ) { |
0 | 1469 mach->ins_req(MemNode::Memory,mem); |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1470 #ifdef ASSERT |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1471 // Verify adr type after matching memory operation |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1472 const MachOper* oper = mach->memory_operand(); |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1473 if (oper != NULL && oper != (MachOper*)-1 && |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1474 mach->adr_type() != TypeRawPtr::BOTTOM) { // non-direct addressing mode |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1475 // It has a unique memory operand. Find corresponding ideal mem node. |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1476 Node* m = NULL; |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1477 if (leaf->is_Mem()) { |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1478 m = leaf; |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1479 } else { |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1480 m = _mem_node; |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1481 assert(m != NULL && m->is_Mem(), "expecting memory node"); |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1482 } |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
1483 const Type* mach_at = mach->adr_type(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
1484 // DecodeN node consumed by an address may have different type |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
1485 // then its input. Don't compare types for such case. |
642
660978a2a31a
6791178: Specialize for zero as the compressed oop vm heap base
kvn
parents:
605
diff
changeset
|
1486 if (m->adr_type() != mach_at && |
660978a2a31a
6791178: Specialize for zero as the compressed oop vm heap base
kvn
parents:
605
diff
changeset
|
1487 (m->in(MemNode::Address)->is_DecodeN() || |
660978a2a31a
6791178: Specialize for zero as the compressed oop vm heap base
kvn
parents:
605
diff
changeset
|
1488 m->in(MemNode::Address)->is_AddP() && |
660978a2a31a
6791178: Specialize for zero as the compressed oop vm heap base
kvn
parents:
605
diff
changeset
|
1489 m->in(MemNode::Address)->in(AddPNode::Address)->is_DecodeN() || |
660978a2a31a
6791178: Specialize for zero as the compressed oop vm heap base
kvn
parents:
605
diff
changeset
|
1490 m->in(MemNode::Address)->is_AddP() && |
660978a2a31a
6791178: Specialize for zero as the compressed oop vm heap base
kvn
parents:
605
diff
changeset
|
1491 m->in(MemNode::Address)->in(AddPNode::Address)->is_AddP() && |
660978a2a31a
6791178: Specialize for zero as the compressed oop vm heap base
kvn
parents:
605
diff
changeset
|
1492 m->in(MemNode::Address)->in(AddPNode::Address)->in(AddPNode::Address)->is_DecodeN())) { |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
1493 mach_at = m->adr_type(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
1494 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
1495 if (m->adr_type() != mach_at) { |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1496 m->dump(); |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1497 tty->print_cr("mach:"); |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1498 mach->dump(1); |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1499 } |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
1500 assert(m->adr_type() == mach_at, "matcher should not change adr type"); |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1501 } |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1502 #endif |
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1503 } |
0 | 1504 |
1505 // If the _leaf is an AddP, insert the base edge | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1506 if( leaf->is_AddP() ) |
0 | 1507 mach->ins_req(AddPNode::Base,leaf->in(AddPNode::Base)); |
1508 | |
1509 uint num_proj = _proj_list.size(); | |
1510 | |
1511 // Perform any 1-to-many expansions required | |
1512 MachNode *ex = mach->Expand(s,_proj_list); | |
1513 if( ex != mach ) { | |
1514 assert(ex->ideal_reg() == mach->ideal_reg(), "ideal types should match"); | |
1515 if( ex->in(1)->is_Con() ) | |
1516 ex->in(1)->set_req(0, C->root()); | |
1517 // Remove old node from the graph | |
1518 for( uint i=0; i<mach->req(); i++ ) { | |
1519 mach->set_req(i,NULL); | |
1520 } | |
222 | 1521 #ifdef ASSERT |
1522 _new2old_map.map(ex->_idx, s->_leaf); | |
1523 #endif | |
0 | 1524 } |
1525 | |
1526 // PhaseChaitin::fixup_spills will sometimes generate spill code | |
1527 // via the matcher. By the time, nodes have been wired into the CFG, | |
1528 // and any further nodes generated by expand rules will be left hanging | |
1529 // in space, and will not get emitted as output code. Catch this. | |
1530 // Also, catch any new register allocation constraints ("projections") | |
1531 // generated belatedly during spill code generation. | |
1532 if (_allocation_started) { | |
1533 guarantee(ex == mach, "no expand rules during spill generation"); | |
1534 guarantee(_proj_list.size() == num_proj, "no allocation during spill generation"); | |
1535 } | |
1536 | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1537 if (leaf->is_Con() || leaf->is_DecodeN()) { |
0 | 1538 // Record the con for sharing |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1539 _shared_nodes.map(leaf->_idx, ex); |
0 | 1540 } |
1541 | |
1542 return ex; | |
1543 } | |
1544 | |
1545 void Matcher::ReduceInst_Chain_Rule( State *s, int rule, Node *&mem, MachNode *mach ) { | |
1546 // 'op' is what I am expecting to receive | |
1547 int op = _leftOp[rule]; | |
1548 // Operand type to catch childs result | |
1549 // This is what my child will give me. | |
1550 int opnd_class_instance = s->_rule[op]; | |
1551 // Choose between operand class or not. | |
605 | 1552 // This is what I will receive. |
0 | 1553 int catch_op = (FIRST_OPERAND_CLASS <= op && op < NUM_OPERANDS) ? opnd_class_instance : op; |
1554 // New rule for child. Chase operand classes to get the actual rule. | |
1555 int newrule = s->_rule[catch_op]; | |
1556 | |
1557 if( newrule < NUM_OPERANDS ) { | |
1558 // Chain from operand or operand class, may be output of shared node | |
1559 assert( 0 <= opnd_class_instance && opnd_class_instance < NUM_OPERANDS, | |
1560 "Bad AD file: Instruction chain rule must chain from operand"); | |
1561 // Insert operand into array of operands for this instruction | |
1562 mach->_opnds[1] = s->MachOperGenerator( opnd_class_instance, C ); | |
1563 | |
1564 ReduceOper( s, newrule, mem, mach ); | |
1565 } else { | |
1566 // Chain from the result of an instruction | |
1567 assert( newrule >= _LAST_MACH_OPER, "Do NOT chain from internal operand"); | |
1568 mach->_opnds[1] = s->MachOperGenerator( _reduceOp[catch_op], C ); | |
1569 Node *mem1 = (Node*)1; | |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1570 debug_only(Node *save_mem_node = _mem_node;) |
0 | 1571 mach->add_req( ReduceInst(s, newrule, mem1) ); |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1572 debug_only(_mem_node = save_mem_node;) |
0 | 1573 } |
1574 return; | |
1575 } | |
1576 | |
1577 | |
1578 uint Matcher::ReduceInst_Interior( State *s, int rule, Node *&mem, MachNode *mach, uint num_opnds ) { | |
1579 if( s->_leaf->is_Load() ) { | |
1580 Node *mem2 = s->_leaf->in(MemNode::Memory); | |
1581 assert( mem == (Node*)1 || mem == mem2, "multiple Memories being matched at once?" ); | |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1582 debug_only( if( mem == (Node*)1 ) _mem_node = s->_leaf;) |
0 | 1583 mem = mem2; |
1584 } | |
1585 if( s->_leaf->in(0) != NULL && s->_leaf->req() > 1) { | |
1586 if( mach->in(0) == NULL ) | |
1587 mach->set_req(0, s->_leaf->in(0)); | |
1588 } | |
1589 | |
1590 // Now recursively walk the state tree & add operand list. | |
1591 for( uint i=0; i<2; i++ ) { // binary tree | |
1592 State *newstate = s->_kids[i]; | |
1593 if( newstate == NULL ) break; // Might only have 1 child | |
1594 // 'op' is what I am expecting to receive | |
1595 int op; | |
1596 if( i == 0 ) { | |
1597 op = _leftOp[rule]; | |
1598 } else { | |
1599 op = _rightOp[rule]; | |
1600 } | |
1601 // Operand type to catch childs result | |
1602 // This is what my child will give me. | |
1603 int opnd_class_instance = newstate->_rule[op]; | |
1604 // Choose between operand class or not. | |
1605 // This is what I will receive. | |
1606 int catch_op = (op >= FIRST_OPERAND_CLASS && op < NUM_OPERANDS) ? opnd_class_instance : op; | |
1607 // New rule for child. Chase operand classes to get the actual rule. | |
1608 int newrule = newstate->_rule[catch_op]; | |
1609 | |
1610 if( newrule < NUM_OPERANDS ) { // Operand/operandClass or internalOp/instruction? | |
1611 // Operand/operandClass | |
1612 // Insert operand into array of operands for this instruction | |
1613 mach->_opnds[num_opnds++] = newstate->MachOperGenerator( opnd_class_instance, C ); | |
1614 ReduceOper( newstate, newrule, mem, mach ); | |
1615 | |
1616 } else { // Child is internal operand or new instruction | |
1617 if( newrule < _LAST_MACH_OPER ) { // internal operand or instruction? | |
1618 // internal operand --> call ReduceInst_Interior | |
1619 // Interior of complex instruction. Do nothing but recurse. | |
1620 num_opnds = ReduceInst_Interior( newstate, newrule, mem, mach, num_opnds ); | |
1621 } else { | |
1622 // instruction --> call build operand( ) to catch result | |
1623 // --> ReduceInst( newrule ) | |
1624 mach->_opnds[num_opnds++] = s->MachOperGenerator( _reduceOp[catch_op], C ); | |
1625 Node *mem1 = (Node*)1; | |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1626 debug_only(Node *save_mem_node = _mem_node;) |
0 | 1627 mach->add_req( ReduceInst( newstate, newrule, mem1 ) ); |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1628 debug_only(_mem_node = save_mem_node;) |
0 | 1629 } |
1630 } | |
1631 assert( mach->_opnds[num_opnds-1], "" ); | |
1632 } | |
1633 return num_opnds; | |
1634 } | |
1635 | |
1636 // This routine walks the interior of possible complex operands. | |
1637 // At each point we check our children in the match tree: | |
1638 // (1) No children - | |
1639 // We are a leaf; add _leaf field as an input to the MachNode | |
1640 // (2) Child is an internal operand - | |
1641 // Skip over it ( do nothing ) | |
1642 // (3) Child is an instruction - | |
1643 // Call ReduceInst recursively and | |
1644 // and instruction as an input to the MachNode | |
1645 void Matcher::ReduceOper( State *s, int rule, Node *&mem, MachNode *mach ) { | |
1646 assert( rule < _LAST_MACH_OPER, "called with operand rule" ); | |
1647 State *kid = s->_kids[0]; | |
1648 assert( kid == NULL || s->_leaf->in(0) == NULL, "internal operands have no control" ); | |
1649 | |
1650 // Leaf? And not subsumed? | |
1651 if( kid == NULL && !_swallowed[rule] ) { | |
1652 mach->add_req( s->_leaf ); // Add leaf pointer | |
1653 return; // Bail out | |
1654 } | |
1655 | |
1656 if( s->_leaf->is_Load() ) { | |
1657 assert( mem == (Node*)1, "multiple Memories being matched at once?" ); | |
1658 mem = s->_leaf->in(MemNode::Memory); | |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1659 debug_only(_mem_node = s->_leaf;) |
0 | 1660 } |
1661 if( s->_leaf->in(0) && s->_leaf->req() > 1) { | |
1662 if( !mach->in(0) ) | |
1663 mach->set_req(0,s->_leaf->in(0)); | |
1664 else { | |
1665 assert( s->_leaf->in(0) == mach->in(0), "same instruction, differing controls?" ); | |
1666 } | |
1667 } | |
1668 | |
1669 for( uint i=0; kid != NULL && i<2; kid = s->_kids[1], i++ ) { // binary tree | |
1670 int newrule; | |
1671 if( i == 0 ) | |
1672 newrule = kid->_rule[_leftOp[rule]]; | |
1673 else | |
1674 newrule = kid->_rule[_rightOp[rule]]; | |
1675 | |
1676 if( newrule < _LAST_MACH_OPER ) { // Operand or instruction? | |
1677 // Internal operand; recurse but do nothing else | |
1678 ReduceOper( kid, newrule, mem, mach ); | |
1679 | |
1680 } else { // Child is a new instruction | |
1681 // Reduce the instruction, and add a direct pointer from this | |
1682 // machine instruction to the newly reduced one. | |
1683 Node *mem1 = (Node*)1; | |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1684 debug_only(Node *save_mem_node = _mem_node;) |
0 | 1685 mach->add_req( ReduceInst( kid, newrule, mem1 ) ); |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
169
diff
changeset
|
1686 debug_only(_mem_node = save_mem_node;) |
0 | 1687 } |
1688 } | |
1689 } | |
1690 | |
1691 | |
1692 // ------------------------------------------------------------------------- | |
1693 // Java-Java calling convention | |
1694 // (what you use when Java calls Java) | |
1695 | |
1696 //------------------------------find_receiver---------------------------------- | |
1697 // For a given signature, return the OptoReg for parameter 0. | |
1698 OptoReg::Name Matcher::find_receiver( bool is_outgoing ) { | |
1699 VMRegPair regs; | |
1700 BasicType sig_bt = T_OBJECT; | |
1701 calling_convention(&sig_bt, ®s, 1, is_outgoing); | |
1702 // Return argument 0 register. In the LP64 build pointers | |
1703 // take 2 registers, but the VM wants only the 'main' name. | |
1704 return OptoReg::as_OptoReg(regs.first()); | |
1705 } | |
1706 | |
1707 // A method-klass-holder may be passed in the inline_cache_reg | |
1708 // and then expanded into the inline_cache_reg and a method_oop register | |
1709 // defined in ad_<arch>.cpp | |
1710 | |
1711 | |
1712 //------------------------------find_shared------------------------------------ | |
1713 // Set bits if Node is shared or otherwise a root | |
1714 void Matcher::find_shared( Node *n ) { | |
1715 // Allocate stack of size C->unique() * 2 to avoid frequent realloc | |
1716 MStack mstack(C->unique() * 2); | |
586
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1717 // Mark nodes as address_visited if they are inputs to an address expression |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1718 VectorSet address_visited(Thread::current()->resource_area()); |
0 | 1719 mstack.push(n, Visit); // Don't need to pre-visit root node |
1720 while (mstack.is_nonempty()) { | |
1721 n = mstack.node(); // Leave node on stack | |
1722 Node_State nstate = mstack.state(); | |
586
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1723 uint nop = n->Opcode(); |
0 | 1724 if (nstate == Pre_Visit) { |
586
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1725 if (address_visited.test(n->_idx)) { // Visited in address already? |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1726 // Flag as visited and shared now. |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1727 set_visited(n); |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1728 } |
0 | 1729 if (is_visited(n)) { // Visited already? |
1730 // Node is shared and has no reason to clone. Flag it as shared. | |
1731 // This causes it to match into a register for the sharing. | |
1732 set_shared(n); // Flag as shared and | |
1733 mstack.pop(); // remove node from stack | |
1734 continue; | |
1735 } | |
1736 nstate = Visit; // Not already visited; so visit now | |
1737 } | |
1738 if (nstate == Visit) { | |
1739 mstack.set_state(Post_Visit); | |
1740 set_visited(n); // Flag as visited now | |
1741 bool mem_op = false; | |
1742 | |
586
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1743 switch( nop ) { // Handle some opcodes special |
0 | 1744 case Op_Phi: // Treat Phis as shared roots |
1745 case Op_Parm: | |
1746 case Op_Proj: // All handled specially during matching | |
63
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
0
diff
changeset
|
1747 case Op_SafePointScalarObject: |
0 | 1748 set_shared(n); |
1749 set_dontcare(n); | |
1750 break; | |
1751 case Op_If: | |
1752 case Op_CountedLoopEnd: | |
1753 mstack.set_state(Alt_Post_Visit); // Alternative way | |
1754 // Convert (If (Bool (CmpX A B))) into (If (Bool) (CmpX A B)). Helps | |
1755 // with matching cmp/branch in 1 instruction. The Matcher needs the | |
1756 // Bool and CmpX side-by-side, because it can only get at constants | |
1757 // that are at the leaves of Match trees, and the Bool's condition acts | |
1758 // as a constant here. | |
1759 mstack.push(n->in(1), Visit); // Clone the Bool | |
1760 mstack.push(n->in(0), Pre_Visit); // Visit control input | |
1761 continue; // while (mstack.is_nonempty()) | |
1762 case Op_ConvI2D: // These forms efficiently match with a prior | |
1763 case Op_ConvI2F: // Load but not a following Store | |
1764 if( n->in(1)->is_Load() && // Prior load | |
1765 n->outcnt() == 1 && // Not already shared | |
1766 n->unique_out()->is_Store() ) // Following store | |
1767 set_shared(n); // Force it to be a root | |
1768 break; | |
1769 case Op_ReverseBytesI: | |
1770 case Op_ReverseBytesL: | |
1771 if( n->in(1)->is_Load() && // Prior load | |
1772 n->outcnt() == 1 ) // Not already shared | |
1773 set_shared(n); // Force it to be a root | |
1774 break; | |
1775 case Op_BoxLock: // Cant match until we get stack-regs in ADLC | |
1776 case Op_IfFalse: | |
1777 case Op_IfTrue: | |
1778 case Op_MachProj: | |
1779 case Op_MergeMem: | |
1780 case Op_Catch: | |
1781 case Op_CatchProj: | |
1782 case Op_CProj: | |
1783 case Op_JumpProj: | |
1784 case Op_JProj: | |
1785 case Op_NeverBranch: | |
1786 set_dontcare(n); | |
1787 break; | |
1788 case Op_Jump: | |
1789 mstack.push(n->in(1), Visit); // Switch Value | |
1790 mstack.push(n->in(0), Pre_Visit); // Visit Control input | |
1791 continue; // while (mstack.is_nonempty()) | |
1792 case Op_StrComp: | |
681 | 1793 case Op_StrEquals: |
1794 case Op_StrIndexOf: | |
169
9148c65abefc
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
168
diff
changeset
|
1795 case Op_AryEq: |
0 | 1796 set_shared(n); // Force result into register (it will be anyways) |
1797 break; | |
1798 case Op_ConP: { // Convert pointers above the centerline to NUL | |
1799 TypeNode *tn = n->as_Type(); // Constants derive from type nodes | |
1800 const TypePtr* tp = tn->type()->is_ptr(); | |
1801 if (tp->_ptr == TypePtr::AnyNull) { | |
1802 tn->set_type(TypePtr::NULL_PTR); | |
1803 } | |
1804 break; | |
1805 } | |
163 | 1806 case Op_ConN: { // Convert narrow pointers above the centerline to NUL |
1807 TypeNode *tn = n->as_Type(); // Constants derive from type nodes | |
221
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
216
diff
changeset
|
1808 const TypePtr* tp = tn->type()->make_ptr(); |
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
216
diff
changeset
|
1809 if (tp && tp->_ptr == TypePtr::AnyNull) { |
163 | 1810 tn->set_type(TypeNarrowOop::NULL_PTR); |
1811 } | |
1812 break; | |
1813 } | |
0 | 1814 case Op_Binary: // These are introduced in the Post_Visit state. |
1815 ShouldNotReachHere(); | |
1816 break; | |
1817 case Op_StoreB: // Do match these, despite no ideal reg | |
1818 case Op_StoreC: | |
1819 case Op_StoreCM: | |
1820 case Op_StoreD: | |
1821 case Op_StoreF: | |
1822 case Op_StoreI: | |
1823 case Op_StoreL: | |
1824 case Op_StoreP: | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
1825 case Op_StoreN: |
0 | 1826 case Op_Store16B: |
1827 case Op_Store8B: | |
1828 case Op_Store4B: | |
1829 case Op_Store8C: | |
1830 case Op_Store4C: | |
1831 case Op_Store2C: | |
1832 case Op_Store4I: | |
1833 case Op_Store2I: | |
1834 case Op_Store2L: | |
1835 case Op_Store4F: | |
1836 case Op_Store2F: | |
1837 case Op_Store2D: | |
1838 case Op_ClearArray: | |
1839 case Op_SafePoint: | |
1840 mem_op = true; | |
1841 break; | |
1842 case Op_LoadB: | |
558
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
420
diff
changeset
|
1843 case Op_LoadUS: |
0 | 1844 case Op_LoadD: |
1845 case Op_LoadF: | |
1846 case Op_LoadI: | |
1847 case Op_LoadKlass: | |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1848 case Op_LoadNKlass: |
0 | 1849 case Op_LoadL: |
1850 case Op_LoadS: | |
1851 case Op_LoadP: | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
1852 case Op_LoadN: |
0 | 1853 case Op_LoadRange: |
1854 case Op_LoadD_unaligned: | |
1855 case Op_LoadL_unaligned: | |
1856 case Op_Load16B: | |
1857 case Op_Load8B: | |
1858 case Op_Load4B: | |
1859 case Op_Load4C: | |
1860 case Op_Load2C: | |
1861 case Op_Load8C: | |
1862 case Op_Load8S: | |
1863 case Op_Load4S: | |
1864 case Op_Load2S: | |
1865 case Op_Load4I: | |
1866 case Op_Load2I: | |
1867 case Op_Load2L: | |
1868 case Op_Load4F: | |
1869 case Op_Load2F: | |
1870 case Op_Load2D: | |
1871 mem_op = true; | |
1872 // Must be root of match tree due to prior load conflict | |
1873 if( C->subsume_loads() == false ) { | |
1874 set_shared(n); | |
1875 } | |
1876 // Fall into default case | |
1877 default: | |
1878 if( !n->ideal_reg() ) | |
1879 set_dontcare(n); // Unmatchable Nodes | |
1880 } // end_switch | |
1881 | |
1882 for(int i = n->req() - 1; i >= 0; --i) { // For my children | |
1883 Node *m = n->in(i); // Get ith input | |
1884 if (m == NULL) continue; // Ignore NULLs | |
1885 uint mop = m->Opcode(); | |
1886 | |
1887 // Must clone all producers of flags, or we will not match correctly. | |
1888 // Suppose a compare setting int-flags is shared (e.g., a switch-tree) | |
1889 // then it will match into an ideal Op_RegFlags. Alas, the fp-flags | |
1890 // are also there, so we may match a float-branch to int-flags and | |
1891 // expect the allocator to haul the flags from the int-side to the | |
1892 // fp-side. No can do. | |
1893 if( _must_clone[mop] ) { | |
1894 mstack.push(m, Visit); | |
1895 continue; // for(int i = ...) | |
1896 } | |
1897 | |
1898 // Clone addressing expressions as they are "free" in most instructions | |
1899 if( mem_op && i == MemNode::Address && mop == Op_AddP ) { | |
309
eaf496ad4a14
6732698: crash with dead code from compressed oops in gcm
never
parents:
235
diff
changeset
|
1900 if (m->in(AddPNode::Base)->Opcode() == Op_DecodeN) { |
eaf496ad4a14
6732698: crash with dead code from compressed oops in gcm
never
parents:
235
diff
changeset
|
1901 // Bases used in addresses must be shared but since |
eaf496ad4a14
6732698: crash with dead code from compressed oops in gcm
never
parents:
235
diff
changeset
|
1902 // they are shared through a DecodeN they may appear |
eaf496ad4a14
6732698: crash with dead code from compressed oops in gcm
never
parents:
235
diff
changeset
|
1903 // to have a single use so force sharing here. |
eaf496ad4a14
6732698: crash with dead code from compressed oops in gcm
never
parents:
235
diff
changeset
|
1904 set_shared(m->in(AddPNode::Base)->in(1)); |
eaf496ad4a14
6732698: crash with dead code from compressed oops in gcm
never
parents:
235
diff
changeset
|
1905 } |
586
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1906 |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1907 // Some inputs for address expression are not put on stack |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1908 // to avoid marking them as shared and forcing them into register |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1909 // if they are used only in address expressions. |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1910 // But they should be marked as shared if there are other uses |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1911 // besides address expressions. |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1912 |
0 | 1913 Node *off = m->in(AddPNode::Offset); |
586
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1914 if( off->is_Con() && |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1915 // When there are other uses besides address expressions |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1916 // put it on stack and mark as shared. |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1917 !is_visited(m) ) { |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1918 address_visited.test_set(m->_idx); // Flag as address_visited |
0 | 1919 Node *adr = m->in(AddPNode::Address); |
1920 | |
1921 // Intel, ARM and friends can handle 2 adds in addressing mode | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1922 if( clone_shift_expressions && adr->is_AddP() && |
0 | 1923 // AtomicAdd is not an addressing expression. |
1924 // Cheap to find it by looking for screwy base. | |
586
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1925 !adr->in(AddPNode::Base)->is_top() && |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1926 // Are there other uses besides address expressions? |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1927 !is_visited(adr) ) { |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1928 address_visited.set(adr->_idx); // Flag as address_visited |
0 | 1929 Node *shift = adr->in(AddPNode::Offset); |
1930 // Check for shift by small constant as well | |
1931 if( shift->Opcode() == Op_LShiftX && shift->in(2)->is_Con() && | |
586
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1932 shift->in(2)->get_int() <= 3 && |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1933 // Are there other uses besides address expressions? |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1934 !is_visited(shift) ) { |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1935 address_visited.set(shift->_idx); // Flag as address_visited |
0 | 1936 mstack.push(shift->in(2), Visit); |
586
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1937 Node *conv = shift->in(1); |
0 | 1938 #ifdef _LP64 |
1939 // Allow Matcher to match the rule which bypass | |
1940 // ConvI2L operation for an array index on LP64 | |
1941 // if the index value is positive. | |
586
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1942 if( conv->Opcode() == Op_ConvI2L && |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1943 conv->as_Type()->type()->is_long()->_lo >= 0 && |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1944 // Are there other uses besides address expressions? |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1945 !is_visited(conv) ) { |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1946 address_visited.set(conv->_idx); // Flag as address_visited |
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1947 mstack.push(conv->in(1), Pre_Visit); |
0 | 1948 } else |
1949 #endif | |
586
6bea93606c11
6791572: assert("duplicating node that's already been matched")
kvn
parents:
558
diff
changeset
|
1950 mstack.push(conv, Pre_Visit); |
0 | 1951 } else { |
1952 mstack.push(shift, Pre_Visit); | |
1953 } | |
1954 mstack.push(adr->in(AddPNode::Address), Pre_Visit); | |
1955 mstack.push(adr->in(AddPNode::Base), Pre_Visit); | |
1956 } else { // Sparc, Alpha, PPC and friends | |
1957 mstack.push(adr, Pre_Visit); | |
1958 } | |
1959 | |
1960 // Clone X+offset as it also folds into most addressing expressions | |
1961 mstack.push(off, Visit); | |
1962 mstack.push(m->in(AddPNode::Base), Pre_Visit); | |
1963 continue; // for(int i = ...) | |
1964 } // if( off->is_Con() ) | |
1965 } // if( mem_op && | |
1966 mstack.push(m, Pre_Visit); | |
1967 } // for(int i = ...) | |
1968 } | |
1969 else if (nstate == Alt_Post_Visit) { | |
1970 mstack.pop(); // Remove node from stack | |
1971 // We cannot remove the Cmp input from the Bool here, as the Bool may be | |
1972 // shared and all users of the Bool need to move the Cmp in parallel. | |
1973 // This leaves both the Bool and the If pointing at the Cmp. To | |
1974 // prevent the Matcher from trying to Match the Cmp along both paths | |
1975 // BoolNode::match_edge always returns a zero. | |
1976 | |
1977 // We reorder the Op_If in a pre-order manner, so we can visit without | |
605 | 1978 // accidentally sharing the Cmp (the Bool and the If make 2 users). |
0 | 1979 n->add_req( n->in(1)->in(1) ); // Add the Cmp next to the Bool |
1980 } | |
1981 else if (nstate == Post_Visit) { | |
1982 mstack.pop(); // Remove node from stack | |
1983 | |
1984 // Now hack a few special opcodes | |
1985 switch( n->Opcode() ) { // Handle some opcodes special | |
1986 case Op_StorePConditional: | |
420
a1980da045cc
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
368
diff
changeset
|
1987 case Op_StoreIConditional: |
0 | 1988 case Op_StoreLConditional: |
1989 case Op_CompareAndSwapI: | |
1990 case Op_CompareAndSwapL: | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
1991 case Op_CompareAndSwapP: |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
1992 case Op_CompareAndSwapN: { // Convert trinary to binary-tree |
0 | 1993 Node *newval = n->in(MemNode::ValueIn ); |
1994 Node *oldval = n->in(LoadStoreNode::ExpectedIn); | |
1995 Node *pair = new (C, 3) BinaryNode( oldval, newval ); | |
1996 n->set_req(MemNode::ValueIn,pair); | |
1997 n->del_req(LoadStoreNode::ExpectedIn); | |
1998 break; | |
1999 } | |
2000 case Op_CMoveD: // Convert trinary to binary-tree | |
2001 case Op_CMoveF: | |
2002 case Op_CMoveI: | |
2003 case Op_CMoveL: | |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2004 case Op_CMoveN: |
0 | 2005 case Op_CMoveP: { |
2006 // Restructure into a binary tree for Matching. It's possible that | |
2007 // we could move this code up next to the graph reshaping for IfNodes | |
2008 // or vice-versa, but I do not want to debug this for Ladybird. | |
2009 // 10/2/2000 CNC. | |
2010 Node *pair1 = new (C, 3) BinaryNode(n->in(1),n->in(1)->in(1)); | |
2011 n->set_req(1,pair1); | |
2012 Node *pair2 = new (C, 3) BinaryNode(n->in(2),n->in(3)); | |
2013 n->set_req(2,pair2); | |
2014 n->del_req(3); | |
2015 break; | |
2016 } | |
2017 default: | |
2018 break; | |
2019 } | |
2020 } | |
2021 else { | |
2022 ShouldNotReachHere(); | |
2023 } | |
2024 } // end of while (mstack.is_nonempty()) | |
2025 } | |
2026 | |
2027 #ifdef ASSERT | |
2028 // machine-independent root to machine-dependent root | |
2029 void Matcher::dump_old2new_map() { | |
2030 _old2new_map.dump(); | |
2031 } | |
2032 #endif | |
2033 | |
2034 //---------------------------collect_null_checks------------------------------- | |
2035 // Find null checks in the ideal graph; write a machine-specific node for | |
2036 // it. Used by later implicit-null-check handling. Actually collects | |
2037 // either an IfTrue or IfFalse for the common NOT-null path, AND the ideal | |
2038 // value being tested. | |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2039 void Matcher::collect_null_checks( Node *proj, Node *orig_proj ) { |
0 | 2040 Node *iff = proj->in(0); |
2041 if( iff->Opcode() == Op_If ) { | |
2042 // During matching If's have Bool & Cmp side-by-side | |
2043 BoolNode *b = iff->in(1)->as_Bool(); | |
2044 Node *cmp = iff->in(2); | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
2045 int opc = cmp->Opcode(); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
2046 if (opc != Op_CmpP && opc != Op_CmpN) return; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
2047 |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
2048 const Type* ct = cmp->in(2)->bottom_type(); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
2049 if (ct == TypePtr::NULL_PTR || |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
2050 (opc == Op_CmpN && ct == TypeNarrowOop::NULL_PTR)) { |
0 | 2051 |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2052 bool push_it = false; |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
2053 if( proj->Opcode() == Op_IfTrue ) { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
2054 extern int all_null_checks_found; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
2055 all_null_checks_found++; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
2056 if( b->_test._test == BoolTest::ne ) { |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2057 push_it = true; |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
2058 } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
2059 } else { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
2060 assert( proj->Opcode() == Op_IfFalse, "" ); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
2061 if( b->_test._test == BoolTest::eq ) { |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2062 push_it = true; |
0 | 2063 } |
2064 } | |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2065 if( push_it ) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2066 _null_check_tests.push(proj); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2067 Node* val = cmp->in(1); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2068 #ifdef _LP64 |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2069 if (UseCompressedOops && !Matcher::clone_shift_expressions && |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2070 val->bottom_type()->isa_narrowoop()) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2071 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2072 // Look for DecodeN node which should be pinned to orig_proj. |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2073 // On platforms (Sparc) which can not handle 2 adds |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2074 // in addressing mode we have to keep a DecodeN node and |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2075 // use it to do implicit NULL check in address. |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2076 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2077 // DecodeN node was pinned to non-null path (orig_proj) during |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2078 // CastPP transformation in final_graph_reshaping_impl(). |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2079 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2080 uint cnt = orig_proj->outcnt(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2081 for (uint i = 0; i < orig_proj->outcnt(); i++) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2082 Node* d = orig_proj->raw_out(i); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2083 if (d->is_DecodeN() && d->in(1) == val) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2084 val = d; |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2085 val->set_req(0, NULL); // Unpin now. |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2086 break; |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2087 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2088 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2089 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2090 #endif |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2091 _null_check_tests.push(val); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2092 } |
0 | 2093 } |
2094 } | |
2095 } | |
2096 | |
2097 //---------------------------validate_null_checks------------------------------ | |
2098 // Its possible that the value being NULL checked is not the root of a match | |
2099 // tree. If so, I cannot use the value in an implicit null check. | |
2100 void Matcher::validate_null_checks( ) { | |
2101 uint cnt = _null_check_tests.size(); | |
2102 for( uint i=0; i < cnt; i+=2 ) { | |
2103 Node *test = _null_check_tests[i]; | |
2104 Node *val = _null_check_tests[i+1]; | |
2105 if (has_new_node(val)) { | |
2106 // Is a match-tree root, so replace with the matched value | |
2107 _null_check_tests.map(i+1, new_node(val)); | |
2108 } else { | |
2109 // Yank from candidate list | |
2110 _null_check_tests.map(i+1,_null_check_tests[--cnt]); | |
2111 _null_check_tests.map(i,_null_check_tests[--cnt]); | |
2112 _null_check_tests.pop(); | |
2113 _null_check_tests.pop(); | |
2114 i-=2; | |
2115 } | |
2116 } | |
2117 } | |
2118 | |
2119 | |
2120 // Used by the DFA in dfa_sparc.cpp. Check for a prior FastLock | |
2121 // acting as an Acquire and thus we don't need an Acquire here. We | |
2122 // retain the Node to act as a compiler ordering barrier. | |
2123 bool Matcher::prior_fast_lock( const Node *acq ) { | |
2124 Node *r = acq->in(0); | |
2125 if( !r->is_Region() || r->req() <= 1 ) return false; | |
2126 Node *proj = r->in(1); | |
2127 if( !proj->is_Proj() ) return false; | |
2128 Node *call = proj->in(0); | |
2129 if( !call->is_Call() || call->as_Call()->entry_point() != OptoRuntime::complete_monitor_locking_Java() ) | |
2130 return false; | |
2131 | |
2132 return true; | |
2133 } | |
2134 | |
2135 // Used by the DFA in dfa_sparc.cpp. Check for a following FastUnLock | |
2136 // acting as a Release and thus we don't need a Release here. We | |
2137 // retain the Node to act as a compiler ordering barrier. | |
2138 bool Matcher::post_fast_unlock( const Node *rel ) { | |
2139 Compile *C = Compile::current(); | |
2140 assert( rel->Opcode() == Op_MemBarRelease, "" ); | |
2141 const MemBarReleaseNode *mem = (const MemBarReleaseNode*)rel; | |
2142 DUIterator_Fast imax, i = mem->fast_outs(imax); | |
2143 Node *ctrl = NULL; | |
2144 while( true ) { | |
2145 ctrl = mem->fast_out(i); // Throw out-of-bounds if proj not found | |
2146 assert( ctrl->is_Proj(), "only projections here" ); | |
2147 ProjNode *proj = (ProjNode*)ctrl; | |
2148 if( proj->_con == TypeFunc::Control && | |
2149 !C->node_arena()->contains(ctrl) ) // Unmatched old-space only | |
2150 break; | |
2151 i++; | |
2152 } | |
2153 Node *iff = NULL; | |
2154 for( DUIterator_Fast jmax, j = ctrl->fast_outs(jmax); j < jmax; j++ ) { | |
2155 Node *x = ctrl->fast_out(j); | |
2156 if( x->is_If() && x->req() > 1 && | |
2157 !C->node_arena()->contains(x) ) { // Unmatched old-space only | |
2158 iff = x; | |
2159 break; | |
2160 } | |
2161 } | |
2162 if( !iff ) return false; | |
2163 Node *bol = iff->in(1); | |
2164 // The iff might be some random subclass of If or bol might be Con-Top | |
2165 if (!bol->is_Bool()) return false; | |
2166 assert( bol->req() > 1, "" ); | |
2167 return (bol->in(1)->Opcode() == Op_FastUnlock); | |
2168 } | |
2169 | |
2170 // Used by the DFA in dfa_xxx.cpp. Check for a following barrier or | |
2171 // atomic instruction acting as a store_load barrier without any | |
2172 // intervening volatile load, and thus we don't need a barrier here. | |
2173 // We retain the Node to act as a compiler ordering barrier. | |
2174 bool Matcher::post_store_load_barrier(const Node *vmb) { | |
2175 Compile *C = Compile::current(); | |
2176 assert( vmb->is_MemBar(), "" ); | |
2177 assert( vmb->Opcode() != Op_MemBarAcquire, "" ); | |
2178 const MemBarNode *mem = (const MemBarNode*)vmb; | |
2179 | |
2180 // Get the Proj node, ctrl, that can be used to iterate forward | |
2181 Node *ctrl = NULL; | |
2182 DUIterator_Fast imax, i = mem->fast_outs(imax); | |
2183 while( true ) { | |
2184 ctrl = mem->fast_out(i); // Throw out-of-bounds if proj not found | |
2185 assert( ctrl->is_Proj(), "only projections here" ); | |
2186 ProjNode *proj = (ProjNode*)ctrl; | |
2187 if( proj->_con == TypeFunc::Control && | |
2188 !C->node_arena()->contains(ctrl) ) // Unmatched old-space only | |
2189 break; | |
2190 i++; | |
2191 } | |
2192 | |
2193 for( DUIterator_Fast jmax, j = ctrl->fast_outs(jmax); j < jmax; j++ ) { | |
2194 Node *x = ctrl->fast_out(j); | |
2195 int xop = x->Opcode(); | |
2196 | |
2197 // We don't need current barrier if we see another or a lock | |
2198 // before seeing volatile load. | |
2199 // | |
2200 // Op_Fastunlock previously appeared in the Op_* list below. | |
2201 // With the advent of 1-0 lock operations we're no longer guaranteed | |
2202 // that a monitor exit operation contains a serializing instruction. | |
2203 | |
2204 if (xop == Op_MemBarVolatile || | |
2205 xop == Op_FastLock || | |
2206 xop == Op_CompareAndSwapL || | |
2207 xop == Op_CompareAndSwapP || | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
63
diff
changeset
|
2208 xop == Op_CompareAndSwapN || |
0 | 2209 xop == Op_CompareAndSwapI) |
2210 return true; | |
2211 | |
2212 if (x->is_MemBar()) { | |
2213 // We must retain this membar if there is an upcoming volatile | |
2214 // load, which will be preceded by acquire membar. | |
2215 if (xop == Op_MemBarAcquire) | |
2216 return false; | |
2217 // For other kinds of barriers, check by pretending we | |
2218 // are them, and seeing if we can be removed. | |
2219 else | |
2220 return post_store_load_barrier((const MemBarNode*)x); | |
2221 } | |
2222 | |
2223 // Delicate code to detect case of an upcoming fastlock block | |
2224 if( x->is_If() && x->req() > 1 && | |
2225 !C->node_arena()->contains(x) ) { // Unmatched old-space only | |
2226 Node *iff = x; | |
2227 Node *bol = iff->in(1); | |
2228 // The iff might be some random subclass of If or bol might be Con-Top | |
2229 if (!bol->is_Bool()) return false; | |
2230 assert( bol->req() > 1, "" ); | |
2231 return (bol->in(1)->Opcode() == Op_FastUnlock); | |
2232 } | |
2233 // probably not necessary to check for these | |
2234 if (x->is_Call() || x->is_SafePoint() || x->is_block_proj()) | |
2235 return false; | |
2236 } | |
2237 return false; | |
2238 } | |
2239 | |
2240 //============================================================================= | |
2241 //---------------------------State--------------------------------------------- | |
2242 State::State(void) { | |
2243 #ifdef ASSERT | |
2244 _id = 0; | |
2245 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe); | |
2246 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d); | |
2247 //memset(_cost, -1, sizeof(_cost)); | |
2248 //memset(_rule, -1, sizeof(_rule)); | |
2249 #endif | |
2250 memset(_valid, 0, sizeof(_valid)); | |
2251 } | |
2252 | |
2253 #ifdef ASSERT | |
2254 State::~State() { | |
2255 _id = 99; | |
2256 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe); | |
2257 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d); | |
2258 memset(_cost, -3, sizeof(_cost)); | |
2259 memset(_rule, -3, sizeof(_rule)); | |
2260 } | |
2261 #endif | |
2262 | |
2263 #ifndef PRODUCT | |
2264 //---------------------------dump---------------------------------------------- | |
2265 void State::dump() { | |
2266 tty->print("\n"); | |
2267 dump(0); | |
2268 } | |
2269 | |
2270 void State::dump(int depth) { | |
2271 for( int j = 0; j < depth; j++ ) | |
2272 tty->print(" "); | |
2273 tty->print("--N: "); | |
2274 _leaf->dump(); | |
2275 uint i; | |
2276 for( i = 0; i < _LAST_MACH_OPER; i++ ) | |
2277 // Check for valid entry | |
2278 if( valid(i) ) { | |
2279 for( int j = 0; j < depth; j++ ) | |
2280 tty->print(" "); | |
2281 assert(_cost[i] != max_juint, "cost must be a valid value"); | |
2282 assert(_rule[i] < _last_Mach_Node, "rule[i] must be valid rule"); | |
2283 tty->print_cr("%s %d %s", | |
2284 ruleName[i], _cost[i], ruleName[_rule[i]] ); | |
2285 } | |
2286 tty->print_cr(""); | |
2287 | |
2288 for( i=0; i<2; i++ ) | |
2289 if( _kids[i] ) | |
2290 _kids[i]->dump(depth+1); | |
2291 } | |
2292 #endif |