Mercurial > hg > graal-compiler
annotate src/share/vm/opto/macro.cpp @ 66:6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
Summary: Remove lock/unlock MemBar nodes and specify locks in debug info for deoptimization.
Reviewed-by: never
author | kvn |
---|---|
date | Fri, 14 Mar 2008 16:40:42 -0700 |
parents | eac007780a58 |
children | a8880a78d355 |
rev | line source |
---|---|
0 | 1 /* |
2 * Copyright 2005-2007 Sun Microsystems, Inc. All Rights Reserved. | |
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. | |
4 * | |
5 * This code is free software; you can redistribute it and/or modify it | |
6 * under the terms of the GNU General Public License version 2 only, as | |
7 * published by the Free Software Foundation. | |
8 * | |
9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
12 * version 2 for more details (a copy is included in the LICENSE file that | |
13 * accompanied this code). | |
14 * | |
15 * You should have received a copy of the GNU General Public License version | |
16 * 2 along with this work; if not, write to the Free Software Foundation, | |
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
18 * | |
19 * Please contact 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/_macro.cpp.incl" | |
27 | |
28 | |
29 // | |
30 // Replace any references to "oldref" in inputs to "use" with "newref". | |
31 // Returns the number of replacements made. | |
32 // | |
33 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { | |
34 int nreplacements = 0; | |
35 uint req = use->req(); | |
36 for (uint j = 0; j < use->len(); j++) { | |
37 Node *uin = use->in(j); | |
38 if (uin == oldref) { | |
39 if (j < req) | |
40 use->set_req(j, newref); | |
41 else | |
42 use->set_prec(j, newref); | |
43 nreplacements++; | |
44 } else if (j >= req && uin == NULL) { | |
45 break; | |
46 } | |
47 } | |
48 return nreplacements; | |
49 } | |
50 | |
51 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) { | |
52 // Copy debug information and adjust JVMState information | |
53 uint old_dbg_start = oldcall->tf()->domain()->cnt(); | |
54 uint new_dbg_start = newcall->tf()->domain()->cnt(); | |
55 int jvms_adj = new_dbg_start - old_dbg_start; | |
56 assert (new_dbg_start == newcall->req(), "argument count mismatch"); | |
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
|
57 |
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
0
diff
changeset
|
58 Dict* sosn_map = new Dict(cmpkey,hashkey); |
0 | 59 for (uint i = old_dbg_start; i < oldcall->req(); i++) { |
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
|
60 Node* old_in = oldcall->in(i); |
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
0
diff
changeset
|
61 // Clone old SafePointScalarObjectNodes, adjusting their field contents. |
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
0
diff
changeset
|
62 if (old_in->is_SafePointScalarObject()) { |
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
0
diff
changeset
|
63 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject(); |
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
0
diff
changeset
|
64 uint old_unique = C->unique(); |
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
0
diff
changeset
|
65 Node* new_in = old_sosn->clone(jvms_adj, sosn_map); |
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
0
diff
changeset
|
66 if (old_unique != C->unique()) { |
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
0
diff
changeset
|
67 new_in = transform_later(new_in); // Register new node. |
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
0
diff
changeset
|
68 } |
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
0
diff
changeset
|
69 old_in = new_in; |
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
0
diff
changeset
|
70 } |
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
0
diff
changeset
|
71 newcall->add_req(old_in); |
0 | 72 } |
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
|
73 |
0 | 74 newcall->set_jvms(oldcall->jvms()); |
75 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) { | |
76 jvms->set_map(newcall); | |
77 jvms->set_locoff(jvms->locoff()+jvms_adj); | |
78 jvms->set_stkoff(jvms->stkoff()+jvms_adj); | |
79 jvms->set_monoff(jvms->monoff()+jvms_adj); | |
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
|
80 jvms->set_scloff(jvms->scloff()+jvms_adj); |
0 | 81 jvms->set_endoff(jvms->endoff()+jvms_adj); |
82 } | |
83 } | |
84 | |
85 Node* PhaseMacroExpand::opt_iff(Node* region, Node* iff) { | |
86 IfNode *opt_iff = transform_later(iff)->as_If(); | |
87 | |
88 // Fast path taken; set region slot 2 | |
89 Node *fast_taken = transform_later( new (C, 1) IfFalseNode(opt_iff) ); | |
90 region->init_req(2,fast_taken); // Capture fast-control | |
91 | |
92 // Fast path not-taken, i.e. slow path | |
93 Node *slow_taken = transform_later( new (C, 1) IfTrueNode(opt_iff) ); | |
94 return slow_taken; | |
95 } | |
96 | |
97 //--------------------copy_predefined_input_for_runtime_call-------------------- | |
98 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { | |
99 // Set fixed predefined input arguments | |
100 call->init_req( TypeFunc::Control, ctrl ); | |
101 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); | |
102 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? | |
103 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); | |
104 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); | |
105 } | |
106 | |
107 //------------------------------make_slow_call--------------------------------- | |
108 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, address slow_call, const char* leaf_name, Node* slow_path, Node* parm0, Node* parm1) { | |
109 | |
110 // Slow-path call | |
111 int size = slow_call_type->domain()->cnt(); | |
112 CallNode *call = leaf_name | |
113 ? (CallNode*)new (C, size) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) | |
114 : (CallNode*)new (C, size) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM ); | |
115 | |
116 // Slow path call has no side-effects, uses few values | |
117 copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); | |
118 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0); | |
119 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1); | |
120 copy_call_debug_info(oldcall, call); | |
121 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. | |
122 _igvn.hash_delete(oldcall); | |
123 _igvn.subsume_node(oldcall, call); | |
124 transform_later(call); | |
125 | |
126 return call; | |
127 } | |
128 | |
129 void PhaseMacroExpand::extract_call_projections(CallNode *call) { | |
130 _fallthroughproj = NULL; | |
131 _fallthroughcatchproj = NULL; | |
132 _ioproj_fallthrough = NULL; | |
133 _ioproj_catchall = NULL; | |
134 _catchallcatchproj = NULL; | |
135 _memproj_fallthrough = NULL; | |
136 _memproj_catchall = NULL; | |
137 _resproj = NULL; | |
138 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) { | |
139 ProjNode *pn = call->fast_out(i)->as_Proj(); | |
140 switch (pn->_con) { | |
141 case TypeFunc::Control: | |
142 { | |
143 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj | |
144 _fallthroughproj = pn; | |
145 DUIterator_Fast jmax, j = pn->fast_outs(jmax); | |
146 const Node *cn = pn->fast_out(j); | |
147 if (cn->is_Catch()) { | |
148 ProjNode *cpn = NULL; | |
149 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { | |
150 cpn = cn->fast_out(k)->as_Proj(); | |
151 assert(cpn->is_CatchProj(), "must be a CatchProjNode"); | |
152 if (cpn->_con == CatchProjNode::fall_through_index) | |
153 _fallthroughcatchproj = cpn; | |
154 else { | |
155 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); | |
156 _catchallcatchproj = cpn; | |
157 } | |
158 } | |
159 } | |
160 break; | |
161 } | |
162 case TypeFunc::I_O: | |
163 if (pn->_is_io_use) | |
164 _ioproj_catchall = pn; | |
165 else | |
166 _ioproj_fallthrough = pn; | |
167 break; | |
168 case TypeFunc::Memory: | |
169 if (pn->_is_io_use) | |
170 _memproj_catchall = pn; | |
171 else | |
172 _memproj_fallthrough = pn; | |
173 break; | |
174 case TypeFunc::Parms: | |
175 _resproj = pn; | |
176 break; | |
177 default: | |
178 assert(false, "unexpected projection from allocation node."); | |
179 } | |
180 } | |
181 | |
182 } | |
183 | |
184 | |
185 //---------------------------set_eden_pointers------------------------- | |
186 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { | |
187 if (UseTLAB) { // Private allocation: load from TLS | |
188 Node* thread = transform_later(new (C, 1) ThreadLocalNode()); | |
189 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); | |
190 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); | |
191 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); | |
192 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); | |
193 } else { // Shared allocation: load from globals | |
194 CollectedHeap* ch = Universe::heap(); | |
195 address top_adr = (address)ch->top_addr(); | |
196 address end_adr = (address)ch->end_addr(); | |
197 eden_top_adr = makecon(TypeRawPtr::make(top_adr)); | |
198 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); | |
199 } | |
200 } | |
201 | |
202 | |
203 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { | |
204 Node* adr = basic_plus_adr(base, offset); | |
205 const TypePtr* adr_type = TypeRawPtr::BOTTOM; | |
206 Node* value = LoadNode::make(C, ctl, mem, adr, adr_type, value_type, bt); | |
207 transform_later(value); | |
208 return value; | |
209 } | |
210 | |
211 | |
212 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { | |
213 Node* adr = basic_plus_adr(base, offset); | |
214 mem = StoreNode::make(C, ctl, mem, adr, NULL, value, bt); | |
215 transform_later(mem); | |
216 return mem; | |
217 } | |
218 | |
219 //============================================================================= | |
220 // | |
221 // A L L O C A T I O N | |
222 // | |
223 // Allocation attempts to be fast in the case of frequent small objects. | |
224 // It breaks down like this: | |
225 // | |
226 // 1) Size in doublewords is computed. This is a constant for objects and | |
227 // variable for most arrays. Doubleword units are used to avoid size | |
228 // overflow of huge doubleword arrays. We need doublewords in the end for | |
229 // rounding. | |
230 // | |
231 // 2) Size is checked for being 'too large'. Too-large allocations will go | |
232 // the slow path into the VM. The slow path can throw any required | |
233 // exceptions, and does all the special checks for very large arrays. The | |
234 // size test can constant-fold away for objects. For objects with | |
235 // finalizers it constant-folds the otherway: you always go slow with | |
236 // finalizers. | |
237 // | |
238 // 3) If NOT using TLABs, this is the contended loop-back point. | |
239 // Load-Locked the heap top. If using TLABs normal-load the heap top. | |
240 // | |
241 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. | |
242 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish | |
243 // "size*8" we always enter the VM, where "largish" is a constant picked small | |
244 // enough that there's always space between the eden max and 4Gig (old space is | |
245 // there so it's quite large) and large enough that the cost of entering the VM | |
246 // is dwarfed by the cost to initialize the space. | |
247 // | |
248 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back | |
249 // down. If contended, repeat at step 3. If using TLABs normal-store | |
250 // adjusted heap top back down; there is no contention. | |
251 // | |
252 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark | |
253 // fields. | |
254 // | |
255 // 7) Merge with the slow-path; cast the raw memory pointer to the correct | |
256 // oop flavor. | |
257 // | |
258 //============================================================================= | |
259 // FastAllocateSizeLimit value is in DOUBLEWORDS. | |
260 // Allocations bigger than this always go the slow route. | |
261 // This value must be small enough that allocation attempts that need to | |
262 // trigger exceptions go the slow route. Also, it must be small enough so | |
263 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. | |
264 //=============================================================================j// | |
265 // %%% Here is an old comment from parseHelper.cpp; is it outdated? | |
266 // The allocator will coalesce int->oop copies away. See comment in | |
267 // coalesce.cpp about how this works. It depends critically on the exact | |
268 // code shape produced here, so if you are changing this code shape | |
269 // make sure the GC info for the heap-top is correct in and around the | |
270 // slow-path call. | |
271 // | |
272 | |
273 void PhaseMacroExpand::expand_allocate_common( | |
274 AllocateNode* alloc, // allocation node to be expanded | |
275 Node* length, // array length for an array allocation | |
276 const TypeFunc* slow_call_type, // Type of slow call | |
277 address slow_call_address // Address of slow call | |
278 ) | |
279 { | |
280 | |
281 Node* ctrl = alloc->in(TypeFunc::Control); | |
282 Node* mem = alloc->in(TypeFunc::Memory); | |
283 Node* i_o = alloc->in(TypeFunc::I_O); | |
284 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); | |
285 Node* klass_node = alloc->in(AllocateNode::KlassNode); | |
286 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); | |
287 | |
288 Node* eden_top_adr; | |
289 Node* eden_end_adr; | |
290 set_eden_pointers(eden_top_adr, eden_end_adr); | |
291 | |
292 uint raw_idx = C->get_alias_index(TypeRawPtr::BOTTOM); | |
293 assert(ctrl != NULL, "must have control"); | |
294 | |
295 // Load Eden::end. Loop invariant and hoisted. | |
296 // | |
297 // Note: We set the control input on "eden_end" and "old_eden_top" when using | |
298 // a TLAB to work around a bug where these values were being moved across | |
299 // a safepoint. These are not oops, so they cannot be include in the oop | |
300 // map, but the can be changed by a GC. The proper way to fix this would | |
301 // be to set the raw memory state when generating a SafepointNode. However | |
302 // this will require extensive changes to the loop optimization in order to | |
303 // prevent a degradation of the optimization. | |
304 // See comment in memnode.hpp, around line 227 in class LoadPNode. | |
305 Node* eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); | |
306 | |
307 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. | |
308 // they will not be used if "always_slow" is set | |
309 enum { slow_result_path = 1, fast_result_path = 2 }; | |
310 Node *result_region; | |
311 Node *result_phi_rawmem; | |
312 Node *result_phi_rawoop; | |
313 Node *result_phi_i_o; | |
314 | |
315 // The initial slow comparison is a size check, the comparison | |
316 // we want to do is a BoolTest::gt | |
317 bool always_slow = false; | |
318 int tv = _igvn.find_int_con(initial_slow_test, -1); | |
319 if (tv >= 0) { | |
320 always_slow = (tv == 1); | |
321 initial_slow_test = NULL; | |
322 } else { | |
323 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); | |
324 } | |
325 | |
326 if (DTraceAllocProbes) { | |
327 // Force slow-path allocation | |
328 always_slow = true; | |
329 initial_slow_test = NULL; | |
330 } | |
331 | |
332 enum { too_big_or_final_path = 1, need_gc_path = 2 }; | |
333 Node *slow_region = NULL; | |
334 Node *toobig_false = ctrl; | |
335 | |
336 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); | |
337 // generate the initial test if necessary | |
338 if (initial_slow_test != NULL ) { | |
339 slow_region = new (C, 3) RegionNode(3); | |
340 | |
341 // Now make the initial failure test. Usually a too-big test but | |
342 // might be a TRUE for finalizers or a fancy class check for | |
343 // newInstance0. | |
344 IfNode *toobig_iff = new (C, 2) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); | |
345 transform_later(toobig_iff); | |
346 // Plug the failing-too-big test into the slow-path region | |
347 Node *toobig_true = new (C, 1) IfTrueNode( toobig_iff ); | |
348 transform_later(toobig_true); | |
349 slow_region ->init_req( too_big_or_final_path, toobig_true ); | |
350 toobig_false = new (C, 1) IfFalseNode( toobig_iff ); | |
351 transform_later(toobig_false); | |
352 } else { // No initial test, just fall into next case | |
353 toobig_false = ctrl; | |
354 debug_only(slow_region = NodeSentinel); | |
355 } | |
356 | |
357 Node *slow_mem = mem; // save the current memory state for slow path | |
358 // generate the fast allocation code unless we know that the initial test will always go slow | |
359 if (!always_slow) { | |
360 // allocate the Region and Phi nodes for the result | |
361 result_region = new (C, 3) RegionNode(3); | |
362 result_phi_rawmem = new (C, 3) PhiNode( result_region, Type::MEMORY, TypeRawPtr::BOTTOM ); | |
363 result_phi_rawoop = new (C, 3) PhiNode( result_region, TypeRawPtr::BOTTOM ); | |
364 result_phi_i_o = new (C, 3) PhiNode( result_region, Type::ABIO ); // I/O is used for Prefetch | |
365 | |
366 // We need a Region for the loop-back contended case. | |
367 enum { fall_in_path = 1, contended_loopback_path = 2 }; | |
368 Node *contended_region; | |
369 Node *contended_phi_rawmem; | |
370 if( UseTLAB ) { | |
371 contended_region = toobig_false; | |
372 contended_phi_rawmem = mem; | |
373 } else { | |
374 contended_region = new (C, 3) RegionNode(3); | |
375 contended_phi_rawmem = new (C, 3) PhiNode( contended_region, Type::MEMORY, TypeRawPtr::BOTTOM); | |
376 // Now handle the passing-too-big test. We fall into the contended | |
377 // loop-back merge point. | |
378 contended_region ->init_req( fall_in_path, toobig_false ); | |
379 contended_phi_rawmem->init_req( fall_in_path, mem ); | |
380 transform_later(contended_region); | |
381 transform_later(contended_phi_rawmem); | |
382 } | |
383 | |
384 // Load(-locked) the heap top. | |
385 // See note above concerning the control input when using a TLAB | |
386 Node *old_eden_top = UseTLAB | |
387 ? new (C, 3) LoadPNode ( ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM ) | |
388 : new (C, 3) LoadPLockedNode( contended_region, contended_phi_rawmem, eden_top_adr ); | |
389 | |
390 transform_later(old_eden_top); | |
391 // Add to heap top to get a new heap top | |
392 Node *new_eden_top = new (C, 4) AddPNode( top(), old_eden_top, size_in_bytes ); | |
393 transform_later(new_eden_top); | |
394 // Check for needing a GC; compare against heap end | |
395 Node *needgc_cmp = new (C, 3) CmpPNode( new_eden_top, eden_end ); | |
396 transform_later(needgc_cmp); | |
397 Node *needgc_bol = new (C, 2) BoolNode( needgc_cmp, BoolTest::ge ); | |
398 transform_later(needgc_bol); | |
399 IfNode *needgc_iff = new (C, 2) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); | |
400 transform_later(needgc_iff); | |
401 | |
402 // Plug the failing-heap-space-need-gc test into the slow-path region | |
403 Node *needgc_true = new (C, 1) IfTrueNode( needgc_iff ); | |
404 transform_later(needgc_true); | |
405 if( initial_slow_test ) { | |
406 slow_region ->init_req( need_gc_path, needgc_true ); | |
407 // This completes all paths into the slow merge point | |
408 transform_later(slow_region); | |
409 } else { // No initial slow path needed! | |
410 // Just fall from the need-GC path straight into the VM call. | |
411 slow_region = needgc_true; | |
412 } | |
413 // No need for a GC. Setup for the Store-Conditional | |
414 Node *needgc_false = new (C, 1) IfFalseNode( needgc_iff ); | |
415 transform_later(needgc_false); | |
416 | |
417 // Grab regular I/O before optional prefetch may change it. | |
418 // Slow-path does no I/O so just set it to the original I/O. | |
419 result_phi_i_o->init_req( slow_result_path, i_o ); | |
420 | |
421 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem, | |
422 old_eden_top, new_eden_top, length); | |
423 | |
424 // Store (-conditional) the modified eden top back down. | |
425 // StorePConditional produces flags for a test PLUS a modified raw | |
426 // memory state. | |
427 Node *store_eden_top; | |
428 Node *fast_oop_ctrl; | |
429 if( UseTLAB ) { | |
430 store_eden_top = new (C, 4) StorePNode( needgc_false, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, new_eden_top ); | |
431 transform_later(store_eden_top); | |
432 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path | |
433 } else { | |
434 store_eden_top = new (C, 5) StorePConditionalNode( needgc_false, contended_phi_rawmem, eden_top_adr, new_eden_top, old_eden_top ); | |
435 transform_later(store_eden_top); | |
436 Node *contention_check = new (C, 2) BoolNode( store_eden_top, BoolTest::ne ); | |
437 transform_later(contention_check); | |
438 store_eden_top = new (C, 1) SCMemProjNode(store_eden_top); | |
439 transform_later(store_eden_top); | |
440 | |
441 // If not using TLABs, check to see if there was contention. | |
442 IfNode *contention_iff = new (C, 2) IfNode ( needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN ); | |
443 transform_later(contention_iff); | |
444 Node *contention_true = new (C, 1) IfTrueNode( contention_iff ); | |
445 transform_later(contention_true); | |
446 // If contention, loopback and try again. | |
447 contended_region->init_req( contended_loopback_path, contention_true ); | |
448 contended_phi_rawmem->init_req( contended_loopback_path, store_eden_top ); | |
449 | |
450 // Fast-path succeeded with no contention! | |
451 Node *contention_false = new (C, 1) IfFalseNode( contention_iff ); | |
452 transform_later(contention_false); | |
453 fast_oop_ctrl = contention_false; | |
454 } | |
455 | |
456 // Rename successful fast-path variables to make meaning more obvious | |
457 Node* fast_oop = old_eden_top; | |
458 Node* fast_oop_rawmem = store_eden_top; | |
459 fast_oop_rawmem = initialize_object(alloc, | |
460 fast_oop_ctrl, fast_oop_rawmem, fast_oop, | |
461 klass_node, length, size_in_bytes); | |
462 | |
463 if (ExtendedDTraceProbes) { | |
464 // Slow-path call | |
465 int size = TypeFunc::Parms + 2; | |
466 CallLeafNode *call = new (C, size) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), | |
467 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), | |
468 "dtrace_object_alloc", | |
469 TypeRawPtr::BOTTOM); | |
470 | |
471 // Get base of thread-local storage area | |
472 Node* thread = new (C, 1) ThreadLocalNode(); | |
473 transform_later(thread); | |
474 | |
475 call->init_req(TypeFunc::Parms+0, thread); | |
476 call->init_req(TypeFunc::Parms+1, fast_oop); | |
477 call->init_req( TypeFunc::Control, fast_oop_ctrl ); | |
478 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o | |
479 call->init_req( TypeFunc::Memory , fast_oop_rawmem ); | |
480 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); | |
481 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); | |
482 transform_later(call); | |
483 fast_oop_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control); | |
484 transform_later(fast_oop_ctrl); | |
485 fast_oop_rawmem = new (C, 1) ProjNode(call,TypeFunc::Memory); | |
486 transform_later(fast_oop_rawmem); | |
487 } | |
488 | |
489 // Plug in the successful fast-path into the result merge point | |
490 result_region ->init_req( fast_result_path, fast_oop_ctrl ); | |
491 result_phi_rawoop->init_req( fast_result_path, fast_oop ); | |
492 result_phi_i_o ->init_req( fast_result_path, i_o ); | |
493 result_phi_rawmem->init_req( fast_result_path, fast_oop_rawmem ); | |
494 } else { | |
495 slow_region = ctrl; | |
496 } | |
497 | |
498 // Generate slow-path call | |
499 CallNode *call = new (C, slow_call_type->domain()->cnt()) | |
500 CallStaticJavaNode(slow_call_type, slow_call_address, | |
501 OptoRuntime::stub_name(slow_call_address), | |
502 alloc->jvms()->bci(), | |
503 TypePtr::BOTTOM); | |
504 call->init_req( TypeFunc::Control, slow_region ); | |
505 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o | |
506 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs | |
507 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); | |
508 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); | |
509 | |
510 call->init_req(TypeFunc::Parms+0, klass_node); | |
511 if (length != NULL) { | |
512 call->init_req(TypeFunc::Parms+1, length); | |
513 } | |
514 | |
515 // Copy debug information and adjust JVMState information, then replace | |
516 // allocate node with the call | |
517 copy_call_debug_info((CallNode *) alloc, call); | |
518 if (!always_slow) { | |
519 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. | |
520 } | |
521 _igvn.hash_delete(alloc); | |
522 _igvn.subsume_node(alloc, call); | |
523 transform_later(call); | |
524 | |
525 // Identify the output projections from the allocate node and | |
526 // adjust any references to them. | |
527 // The control and io projections look like: | |
528 // | |
529 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) | |
530 // Allocate Catch | |
531 // ^---Proj(io) <-------+ ^---CatchProj(io) | |
532 // | |
533 // We are interested in the CatchProj nodes. | |
534 // | |
535 extract_call_projections(call); | |
536 | |
537 // An allocate node has separate memory projections for the uses on the control and i_o paths | |
538 // Replace uses of the control memory projection with result_phi_rawmem (unless we are only generating a slow call) | |
539 if (!always_slow && _memproj_fallthrough != NULL) { | |
540 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { | |
541 Node *use = _memproj_fallthrough->fast_out(i); | |
542 _igvn.hash_delete(use); | |
543 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); | |
544 _igvn._worklist.push(use); | |
545 // back up iterator | |
546 --i; | |
547 } | |
548 } | |
549 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete _memproj_catchall so | |
550 // we end up with a call that has only 1 memory projection | |
551 if (_memproj_catchall != NULL ) { | |
552 if (_memproj_fallthrough == NULL) { | |
553 _memproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::Memory); | |
554 transform_later(_memproj_fallthrough); | |
555 } | |
556 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { | |
557 Node *use = _memproj_catchall->fast_out(i); | |
558 _igvn.hash_delete(use); | |
559 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); | |
560 _igvn._worklist.push(use); | |
561 // back up iterator | |
562 --i; | |
563 } | |
564 } | |
565 | |
566 mem = result_phi_rawmem; | |
567 | |
568 // An allocate node has separate i_o projections for the uses on the control and i_o paths | |
569 // Replace uses of the control i_o projection with result_phi_i_o (unless we are only generating a slow call) | |
570 if (_ioproj_fallthrough == NULL) { | |
571 _ioproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::I_O); | |
572 transform_later(_ioproj_fallthrough); | |
573 } else if (!always_slow) { | |
574 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { | |
575 Node *use = _ioproj_fallthrough->fast_out(i); | |
576 | |
577 _igvn.hash_delete(use); | |
578 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); | |
579 _igvn._worklist.push(use); | |
580 // back up iterator | |
581 --i; | |
582 } | |
583 } | |
584 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete _ioproj_catchall so | |
585 // we end up with a call that has only 1 control projection | |
586 if (_ioproj_catchall != NULL ) { | |
587 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { | |
588 Node *use = _ioproj_catchall->fast_out(i); | |
589 _igvn.hash_delete(use); | |
590 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); | |
591 _igvn._worklist.push(use); | |
592 // back up iterator | |
593 --i; | |
594 } | |
595 } | |
596 | |
597 // if we generated only a slow call, we are done | |
598 if (always_slow) | |
599 return; | |
600 | |
601 | |
602 if (_fallthroughcatchproj != NULL) { | |
603 ctrl = _fallthroughcatchproj->clone(); | |
604 transform_later(ctrl); | |
605 _igvn.hash_delete(_fallthroughcatchproj); | |
606 _igvn.subsume_node(_fallthroughcatchproj, result_region); | |
607 } else { | |
608 ctrl = top(); | |
609 } | |
610 Node *slow_result; | |
611 if (_resproj == NULL) { | |
612 // no uses of the allocation result | |
613 slow_result = top(); | |
614 } else { | |
615 slow_result = _resproj->clone(); | |
616 transform_later(slow_result); | |
617 _igvn.hash_delete(_resproj); | |
618 _igvn.subsume_node(_resproj, result_phi_rawoop); | |
619 } | |
620 | |
621 // Plug slow-path into result merge point | |
622 result_region ->init_req( slow_result_path, ctrl ); | |
623 result_phi_rawoop->init_req( slow_result_path, slow_result); | |
624 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); | |
625 transform_later(result_region); | |
626 transform_later(result_phi_rawoop); | |
627 transform_later(result_phi_rawmem); | |
628 transform_later(result_phi_i_o); | |
629 // This completes all paths into the result merge point | |
630 } | |
631 | |
632 | |
633 // Helper for PhaseMacroExpand::expand_allocate_common. | |
634 // Initializes the newly-allocated storage. | |
635 Node* | |
636 PhaseMacroExpand::initialize_object(AllocateNode* alloc, | |
637 Node* control, Node* rawmem, Node* object, | |
638 Node* klass_node, Node* length, | |
639 Node* size_in_bytes) { | |
640 InitializeNode* init = alloc->initialization(); | |
641 // Store the klass & mark bits | |
642 Node* mark_node = NULL; | |
643 // For now only enable fast locking for non-array types | |
644 if (UseBiasedLocking && (length == NULL)) { | |
645 mark_node = make_load(NULL, rawmem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeRawPtr::BOTTOM, T_ADDRESS); | |
646 } else { | |
647 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); | |
648 } | |
649 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); | |
650 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_OBJECT); | |
651 int header_size = alloc->minimum_header_size(); // conservatively small | |
652 | |
653 // Array length | |
654 if (length != NULL) { // Arrays need length field | |
655 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); | |
656 // conservatively small header size: | |
657 header_size = sizeof(arrayOopDesc); | |
658 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); | |
659 if (k->is_array_klass()) // we know the exact header size in most cases: | |
660 header_size = Klass::layout_helper_header_size(k->layout_helper()); | |
661 } | |
662 | |
663 // Clear the object body, if necessary. | |
664 if (init == NULL) { | |
665 // The init has somehow disappeared; be cautious and clear everything. | |
666 // | |
667 // This can happen if a node is allocated but an uncommon trap occurs | |
668 // immediately. In this case, the Initialize gets associated with the | |
669 // trap, and may be placed in a different (outer) loop, if the Allocate | |
670 // is in a loop. If (this is rare) the inner loop gets unrolled, then | |
671 // there can be two Allocates to one Initialize. The answer in all these | |
672 // edge cases is safety first. It is always safe to clear immediately | |
673 // within an Allocate, and then (maybe or maybe not) clear some more later. | |
674 if (!ZeroTLAB) | |
675 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, | |
676 header_size, size_in_bytes, | |
677 &_igvn); | |
678 } else { | |
679 if (!init->is_complete()) { | |
680 // Try to win by zeroing only what the init does not store. | |
681 // We can also try to do some peephole optimizations, | |
682 // such as combining some adjacent subword stores. | |
683 rawmem = init->complete_stores(control, rawmem, object, | |
684 header_size, size_in_bytes, &_igvn); | |
685 } | |
686 | |
687 // We have no more use for this link, since the AllocateNode goes away: | |
688 init->set_req(InitializeNode::RawAddress, top()); | |
689 // (If we keep the link, it just confuses the register allocator, | |
690 // who thinks he sees a real use of the address by the membar.) | |
691 } | |
692 | |
693 return rawmem; | |
694 } | |
695 | |
696 // Generate prefetch instructions for next allocations. | |
697 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, | |
698 Node*& contended_phi_rawmem, | |
699 Node* old_eden_top, Node* new_eden_top, | |
700 Node* length) { | |
701 if( UseTLAB && AllocatePrefetchStyle == 2 ) { | |
702 // Generate prefetch allocation with watermark check. | |
703 // As an allocation hits the watermark, we will prefetch starting | |
704 // at a "distance" away from watermark. | |
705 enum { fall_in_path = 1, pf_path = 2 }; | |
706 | |
707 Node *pf_region = new (C, 3) RegionNode(3); | |
708 Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY, | |
709 TypeRawPtr::BOTTOM ); | |
710 // I/O is used for Prefetch | |
711 Node *pf_phi_abio = new (C, 3) PhiNode( pf_region, Type::ABIO ); | |
712 | |
713 Node *thread = new (C, 1) ThreadLocalNode(); | |
714 transform_later(thread); | |
715 | |
716 Node *eden_pf_adr = new (C, 4) AddPNode( top()/*not oop*/, thread, | |
717 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); | |
718 transform_later(eden_pf_adr); | |
719 | |
720 Node *old_pf_wm = new (C, 3) LoadPNode( needgc_false, | |
721 contended_phi_rawmem, eden_pf_adr, | |
722 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM ); | |
723 transform_later(old_pf_wm); | |
724 | |
725 // check against new_eden_top | |
726 Node *need_pf_cmp = new (C, 3) CmpPNode( new_eden_top, old_pf_wm ); | |
727 transform_later(need_pf_cmp); | |
728 Node *need_pf_bol = new (C, 2) BoolNode( need_pf_cmp, BoolTest::ge ); | |
729 transform_later(need_pf_bol); | |
730 IfNode *need_pf_iff = new (C, 2) IfNode( needgc_false, need_pf_bol, | |
731 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); | |
732 transform_later(need_pf_iff); | |
733 | |
734 // true node, add prefetchdistance | |
735 Node *need_pf_true = new (C, 1) IfTrueNode( need_pf_iff ); | |
736 transform_later(need_pf_true); | |
737 | |
738 Node *need_pf_false = new (C, 1) IfFalseNode( need_pf_iff ); | |
739 transform_later(need_pf_false); | |
740 | |
741 Node *new_pf_wmt = new (C, 4) AddPNode( top(), old_pf_wm, | |
742 _igvn.MakeConX(AllocatePrefetchDistance) ); | |
743 transform_later(new_pf_wmt ); | |
744 new_pf_wmt->set_req(0, need_pf_true); | |
745 | |
746 Node *store_new_wmt = new (C, 4) StorePNode( need_pf_true, | |
747 contended_phi_rawmem, eden_pf_adr, | |
748 TypeRawPtr::BOTTOM, new_pf_wmt ); | |
749 transform_later(store_new_wmt); | |
750 | |
751 // adding prefetches | |
752 pf_phi_abio->init_req( fall_in_path, i_o ); | |
753 | |
754 Node *prefetch_adr; | |
755 Node *prefetch; | |
756 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize; | |
757 uint step_size = AllocatePrefetchStepSize; | |
758 uint distance = 0; | |
759 | |
760 for ( uint i = 0; i < lines; i++ ) { | |
761 prefetch_adr = new (C, 4) AddPNode( old_pf_wm, new_pf_wmt, | |
762 _igvn.MakeConX(distance) ); | |
763 transform_later(prefetch_adr); | |
764 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr ); | |
765 transform_later(prefetch); | |
766 distance += step_size; | |
767 i_o = prefetch; | |
768 } | |
769 pf_phi_abio->set_req( pf_path, i_o ); | |
770 | |
771 pf_region->init_req( fall_in_path, need_pf_false ); | |
772 pf_region->init_req( pf_path, need_pf_true ); | |
773 | |
774 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); | |
775 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); | |
776 | |
777 transform_later(pf_region); | |
778 transform_later(pf_phi_rawmem); | |
779 transform_later(pf_phi_abio); | |
780 | |
781 needgc_false = pf_region; | |
782 contended_phi_rawmem = pf_phi_rawmem; | |
783 i_o = pf_phi_abio; | |
784 } else if( AllocatePrefetchStyle > 0 ) { | |
785 // Insert a prefetch for each allocation only on the fast-path | |
786 Node *prefetch_adr; | |
787 Node *prefetch; | |
788 // Generate several prefetch instructions only for arrays. | |
789 uint lines = (length != NULL) ? AllocatePrefetchLines : 1; | |
790 uint step_size = AllocatePrefetchStepSize; | |
791 uint distance = AllocatePrefetchDistance; | |
792 for ( uint i = 0; i < lines; i++ ) { | |
793 prefetch_adr = new (C, 4) AddPNode( old_eden_top, new_eden_top, | |
794 _igvn.MakeConX(distance) ); | |
795 transform_later(prefetch_adr); | |
796 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr ); | |
797 // Do not let it float too high, since if eden_top == eden_end, | |
798 // both might be null. | |
799 if( i == 0 ) { // Set control for first prefetch, next follows it | |
800 prefetch->init_req(0, needgc_false); | |
801 } | |
802 transform_later(prefetch); | |
803 distance += step_size; | |
804 i_o = prefetch; | |
805 } | |
806 } | |
807 return i_o; | |
808 } | |
809 | |
810 | |
811 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { | |
812 expand_allocate_common(alloc, NULL, | |
813 OptoRuntime::new_instance_Type(), | |
814 OptoRuntime::new_instance_Java()); | |
815 } | |
816 | |
817 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { | |
818 Node* length = alloc->in(AllocateNode::ALength); | |
819 expand_allocate_common(alloc, length, | |
820 OptoRuntime::new_array_Type(), | |
821 OptoRuntime::new_array_Java()); | |
822 } | |
823 | |
824 | |
825 // we have determined that this lock/unlock can be eliminated, we simply | |
826 // eliminate the node without expanding it. | |
827 // | |
828 // Note: The membar's associated with the lock/unlock are currently not | |
829 // eliminated. This should be investigated as a future enhancement. | |
830 // | |
66
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
831 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
832 |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
833 if (!alock->is_eliminated()) { |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
834 return false; |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
835 } |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
836 // Mark the box lock as eliminated if all correspondent locks are eliminated |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
837 // to construct correct debug info. |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
838 BoxLockNode* box = alock->box_node()->as_BoxLock(); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
839 if (!box->is_eliminated()) { |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
840 bool eliminate = true; |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
841 for (DUIterator_Fast imax, i = box->fast_outs(imax); i < imax; i++) { |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
842 Node *lck = box->fast_out(i); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
843 if (lck->is_Lock() && !lck->as_AbstractLock()->is_eliminated()) { |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
844 eliminate = false; |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
845 break; |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
846 } |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
847 } |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
848 if (eliminate) |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
849 box->set_eliminated(); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
850 } |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
851 |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
852 #ifndef PRODUCT |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
853 if (PrintEliminateLocks) { |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
854 if (alock->is_Lock()) { |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
855 tty->print_cr("++++ Eliminating: %d Lock", alock->_idx); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
856 } else { |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
857 tty->print_cr("++++ Eliminating: %d Unlock", alock->_idx); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
858 } |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
859 } |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
860 #endif |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
861 |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
862 Node* mem = alock->in(TypeFunc::Memory); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
863 Node* ctrl = alock->in(TypeFunc::Control); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
864 |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
865 extract_call_projections(alock); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
866 // There are 2 projections from the lock. The lock node will |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
867 // be deleted when its last use is subsumed below. |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
868 assert(alock->outcnt() == 2 && |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
869 _fallthroughproj != NULL && |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
870 _memproj_fallthrough != NULL, |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
871 "Unexpected projections from Lock/Unlock"); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
872 |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
873 Node* fallthroughproj = _fallthroughproj; |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
874 Node* memproj_fallthrough = _memproj_fallthrough; |
0 | 875 |
876 // The memory projection from a lock/unlock is RawMem | |
877 // The input to a Lock is merged memory, so extract its RawMem input | |
878 // (unless the MergeMem has been optimized away.) | |
879 if (alock->is_Lock()) { | |
66
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
880 // Seach for MemBarAcquire node and delete it also. |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
881 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
882 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquire, ""); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
883 Node* ctrlproj = membar->proj_out(TypeFunc::Control); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
884 Node* memproj = membar->proj_out(TypeFunc::Memory); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
885 _igvn.hash_delete(ctrlproj); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
886 _igvn.subsume_node(ctrlproj, fallthroughproj); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
887 _igvn.hash_delete(memproj); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
888 _igvn.subsume_node(memproj, memproj_fallthrough); |
0 | 889 } |
890 | |
66
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
891 // Seach for MemBarRelease node and delete it also. |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
892 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() && |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
893 ctrl->in(0)->is_MemBar()) { |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
894 MemBarNode* membar = ctrl->in(0)->as_MemBar(); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
895 assert(membar->Opcode() == Op_MemBarRelease && |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
896 mem->is_Proj() && membar == mem->in(0), ""); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
897 _igvn.hash_delete(fallthroughproj); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
898 _igvn.subsume_node(fallthroughproj, ctrl); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
899 _igvn.hash_delete(memproj_fallthrough); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
900 _igvn.subsume_node(memproj_fallthrough, mem); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
901 fallthroughproj = ctrl; |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
902 memproj_fallthrough = mem; |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
903 ctrl = membar->in(TypeFunc::Control); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
904 mem = membar->in(TypeFunc::Memory); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
905 } |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
906 |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
907 _igvn.hash_delete(fallthroughproj); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
908 _igvn.subsume_node(fallthroughproj, ctrl); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
909 _igvn.hash_delete(memproj_fallthrough); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
910 _igvn.subsume_node(memproj_fallthrough, mem); |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
911 return true; |
0 | 912 } |
913 | |
914 | |
915 //------------------------------expand_lock_node---------------------- | |
916 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { | |
917 | |
66
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
918 if (eliminate_locking_node(lock)) { |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
919 return; |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
920 } |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
921 |
0 | 922 Node* ctrl = lock->in(TypeFunc::Control); |
923 Node* mem = lock->in(TypeFunc::Memory); | |
924 Node* obj = lock->obj_node(); | |
925 Node* box = lock->box_node(); | |
66
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
926 Node* flock = lock->fastlock_node(); |
0 | 927 |
928 // Make the merge point | |
929 Node *region = new (C, 3) RegionNode(3); | |
930 | |
931 Node *bol = transform_later(new (C, 2) BoolNode(flock,BoolTest::ne)); | |
932 Node *iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); | |
933 // Optimize test; set region slot 2 | |
934 Node *slow_path = opt_iff(region,iff); | |
935 | |
936 // Make slow path call | |
937 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box ); | |
938 | |
939 extract_call_projections(call); | |
940 | |
941 // Slow path can only throw asynchronous exceptions, which are always | |
942 // de-opted. So the compiler thinks the slow-call can never throw an | |
943 // exception. If it DOES throw an exception we would need the debug | |
944 // info removed first (since if it throws there is no monitor). | |
945 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && | |
946 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); | |
947 | |
948 // Capture slow path | |
949 // disconnect fall-through projection from call and create a new one | |
950 // hook up users of fall-through projection to region | |
951 Node *slow_ctrl = _fallthroughproj->clone(); | |
952 transform_later(slow_ctrl); | |
953 _igvn.hash_delete(_fallthroughproj); | |
954 _fallthroughproj->disconnect_inputs(NULL); | |
955 region->init_req(1, slow_ctrl); | |
956 // region inputs are now complete | |
957 transform_later(region); | |
958 _igvn.subsume_node(_fallthroughproj, region); | |
959 | |
960 // create a Phi for the memory state | |
961 Node *mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); | |
962 Node *memproj = transform_later( new (C, 1) ProjNode(call, TypeFunc::Memory) ); | |
963 mem_phi->init_req(1, memproj ); | |
964 mem_phi->init_req(2, mem); | |
965 transform_later(mem_phi); | |
966 _igvn.hash_delete(_memproj_fallthrough); | |
967 _igvn.subsume_node(_memproj_fallthrough, mem_phi); | |
968 | |
969 | |
970 } | |
971 | |
972 //------------------------------expand_unlock_node---------------------- | |
973 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { | |
974 | |
66
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
975 if (eliminate_locking_node(unlock)) { |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
976 return; |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
977 } |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
978 |
6dbf1a175d6b
6672848: (Escape Analysis) improve lock elimination with EA
kvn
parents:
63
diff
changeset
|
979 Node* ctrl = unlock->in(TypeFunc::Control); |
0 | 980 Node* mem = unlock->in(TypeFunc::Memory); |
981 Node* obj = unlock->obj_node(); | |
982 Node* box = unlock->box_node(); | |
983 | |
984 // No need for a null check on unlock | |
985 | |
986 // Make the merge point | |
987 RegionNode *region = new (C, 3) RegionNode(3); | |
988 | |
989 FastUnlockNode *funlock = new (C, 3) FastUnlockNode( ctrl, obj, box ); | |
990 funlock = transform_later( funlock )->as_FastUnlock(); | |
991 Node *bol = transform_later(new (C, 2) BoolNode(funlock,BoolTest::ne)); | |
992 Node *iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); | |
993 // Optimize test; set region slot 2 | |
994 Node *slow_path = opt_iff(region,iff); | |
995 | |
996 CallNode *call = make_slow_call( (CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), "complete_monitor_unlocking_C", slow_path, obj, box ); | |
997 | |
998 extract_call_projections(call); | |
999 | |
1000 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && | |
1001 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); | |
1002 | |
1003 // No exceptions for unlocking | |
1004 // Capture slow path | |
1005 // disconnect fall-through projection from call and create a new one | |
1006 // hook up users of fall-through projection to region | |
1007 Node *slow_ctrl = _fallthroughproj->clone(); | |
1008 transform_later(slow_ctrl); | |
1009 _igvn.hash_delete(_fallthroughproj); | |
1010 _fallthroughproj->disconnect_inputs(NULL); | |
1011 region->init_req(1, slow_ctrl); | |
1012 // region inputs are now complete | |
1013 transform_later(region); | |
1014 _igvn.subsume_node(_fallthroughproj, region); | |
1015 | |
1016 // create a Phi for the memory state | |
1017 Node *mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); | |
1018 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) ); | |
1019 mem_phi->init_req(1, memproj ); | |
1020 mem_phi->init_req(2, mem); | |
1021 transform_later(mem_phi); | |
1022 _igvn.hash_delete(_memproj_fallthrough); | |
1023 _igvn.subsume_node(_memproj_fallthrough, mem_phi); | |
1024 | |
1025 | |
1026 } | |
1027 | |
1028 //------------------------------expand_macro_nodes---------------------- | |
1029 // Returns true if a failure occurred. | |
1030 bool PhaseMacroExpand::expand_macro_nodes() { | |
1031 if (C->macro_count() == 0) | |
1032 return false; | |
1033 // Make sure expansion will not cause node limit to be exceeded. Worst case is a | |
1034 // macro node gets expanded into about 50 nodes. Allow 50% more for optimization | |
1035 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) ) | |
1036 return true; | |
1037 // expand "macro" nodes | |
1038 // nodes are removed from the macro list as they are processed | |
1039 while (C->macro_count() > 0) { | |
1040 Node * n = C->macro_node(0); | |
1041 assert(n->is_macro(), "only macro nodes expected here"); | |
1042 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { | |
1043 // node is unreachable, so don't try to expand it | |
1044 C->remove_macro_node(n); | |
1045 continue; | |
1046 } | |
1047 switch (n->class_id()) { | |
1048 case Node::Class_Allocate: | |
1049 expand_allocate(n->as_Allocate()); | |
1050 break; | |
1051 case Node::Class_AllocateArray: | |
1052 expand_allocate_array(n->as_AllocateArray()); | |
1053 break; | |
1054 case Node::Class_Lock: | |
1055 expand_lock_node(n->as_Lock()); | |
1056 break; | |
1057 case Node::Class_Unlock: | |
1058 expand_unlock_node(n->as_Unlock()); | |
1059 break; | |
1060 default: | |
1061 assert(false, "unknown node type in macro list"); | |
1062 } | |
1063 if (C->failing()) return true; | |
1064 } | |
1065 _igvn.optimize(); | |
1066 return false; | |
1067 } |