Mercurial > hg > truffle

comparison src/share/vm/opto/macro.cpp @ 0:a61af66fc99e jdk7-b24

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Initial load
author duke
date Sat, 01 Dec 2007 00:00:00 +0000
parents
children eac007780a58
comparison
equal deleted inserted replaced
-1:000000000000 0:a61af66fc99e
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");
57 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
58 newcall->add_req(oldcall->in(i));
59 }
60 newcall->set_jvms(oldcall->jvms());
61 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
62 jvms->set_map(newcall);
63 jvms->set_locoff(jvms->locoff()+jvms_adj);
64 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
65 jvms->set_monoff(jvms->monoff()+jvms_adj);
66 jvms->set_endoff(jvms->endoff()+jvms_adj);
67 }
68 }
69
70 Node* PhaseMacroExpand::opt_iff(Node* region, Node* iff) {
71 IfNode *opt_iff = transform_later(iff)->as_If();
72
73 // Fast path taken; set region slot 2
74 Node *fast_taken = transform_later( new (C, 1) IfFalseNode(opt_iff) );
75 region->init_req(2,fast_taken); // Capture fast-control
76
77 // Fast path not-taken, i.e. slow path
78 Node *slow_taken = transform_later( new (C, 1) IfTrueNode(opt_iff) );
79 return slow_taken;
80 }
81
82 //--------------------copy_predefined_input_for_runtime_call--------------------
83 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
84 // Set fixed predefined input arguments
85 call->init_req( TypeFunc::Control, ctrl );
86 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
87 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
88 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
89 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
90 }
91
92 //------------------------------make_slow_call---------------------------------
93 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) {
94
95 // Slow-path call
96 int size = slow_call_type->domain()->cnt();
97 CallNode *call = leaf_name
98 ? (CallNode*)new (C, size) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
99 : (CallNode*)new (C, size) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
100
101 // Slow path call has no side-effects, uses few values
102 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
103 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
104 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
105 copy_call_debug_info(oldcall, call);
106 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
107 _igvn.hash_delete(oldcall);
108 _igvn.subsume_node(oldcall, call);
109 transform_later(call);
110
111 return call;
112 }
113
114 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
115 _fallthroughproj = NULL;
116 _fallthroughcatchproj = NULL;
117 _ioproj_fallthrough = NULL;
118 _ioproj_catchall = NULL;
119 _catchallcatchproj = NULL;
120 _memproj_fallthrough = NULL;
121 _memproj_catchall = NULL;
122 _resproj = NULL;
123 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
124 ProjNode *pn = call->fast_out(i)->as_Proj();
125 switch (pn->_con) {
126 case TypeFunc::Control:
127 {
128 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
129 _fallthroughproj = pn;
130 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
131 const Node *cn = pn->fast_out(j);
132 if (cn->is_Catch()) {
133 ProjNode *cpn = NULL;
134 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
135 cpn = cn->fast_out(k)->as_Proj();
136 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
137 if (cpn->_con == CatchProjNode::fall_through_index)
138 _fallthroughcatchproj = cpn;
139 else {
140 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
141 _catchallcatchproj = cpn;
142 }
143 }
144 }
145 break;
146 }
147 case TypeFunc::I_O:
148 if (pn->_is_io_use)
149 _ioproj_catchall = pn;
150 else
151 _ioproj_fallthrough = pn;
152 break;
153 case TypeFunc::Memory:
154 if (pn->_is_io_use)
155 _memproj_catchall = pn;
156 else
157 _memproj_fallthrough = pn;
158 break;
159 case TypeFunc::Parms:
160 _resproj = pn;
161 break;
162 default:
163 assert(false, "unexpected projection from allocation node.");
164 }
165 }
166
167 }
168
169
170 //---------------------------set_eden_pointers-------------------------
171 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
172 if (UseTLAB) { // Private allocation: load from TLS
173 Node* thread = transform_later(new (C, 1) ThreadLocalNode());
174 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
175 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
176 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
177 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
178 } else { // Shared allocation: load from globals
179 CollectedHeap* ch = Universe::heap();
180 address top_adr = (address)ch->top_addr();
181 address end_adr = (address)ch->end_addr();
182 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
183 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
184 }
185 }
186
187
188 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
189 Node* adr = basic_plus_adr(base, offset);
190 const TypePtr* adr_type = TypeRawPtr::BOTTOM;
191 Node* value = LoadNode::make(C, ctl, mem, adr, adr_type, value_type, bt);
192 transform_later(value);
193 return value;
194 }
195
196
197 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
198 Node* adr = basic_plus_adr(base, offset);
199 mem = StoreNode::make(C, ctl, mem, adr, NULL, value, bt);
200 transform_later(mem);
201 return mem;
202 }
203
204 //=============================================================================
205 //
206 // A L L O C A T I O N
207 //
208 // Allocation attempts to be fast in the case of frequent small objects.
209 // It breaks down like this:
210 //
211 // 1) Size in doublewords is computed. This is a constant for objects and
212 // variable for most arrays. Doubleword units are used to avoid size
213 // overflow of huge doubleword arrays. We need doublewords in the end for
214 // rounding.
215 //
216 // 2) Size is checked for being 'too large'. Too-large allocations will go
217 // the slow path into the VM. The slow path can throw any required
218 // exceptions, and does all the special checks for very large arrays. The
219 // size test can constant-fold away for objects. For objects with
220 // finalizers it constant-folds the otherway: you always go slow with
221 // finalizers.
222 //
223 // 3) If NOT using TLABs, this is the contended loop-back point.
224 // Load-Locked the heap top. If using TLABs normal-load the heap top.
225 //
226 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
227 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
228 // "size*8" we always enter the VM, where "largish" is a constant picked small
229 // enough that there's always space between the eden max and 4Gig (old space is
230 // there so it's quite large) and large enough that the cost of entering the VM
231 // is dwarfed by the cost to initialize the space.
232 //
233 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
234 // down. If contended, repeat at step 3. If using TLABs normal-store
235 // adjusted heap top back down; there is no contention.
236 //
237 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
238 // fields.
239 //
240 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
241 // oop flavor.
242 //
243 //=============================================================================
244 // FastAllocateSizeLimit value is in DOUBLEWORDS.
245 // Allocations bigger than this always go the slow route.
246 // This value must be small enough that allocation attempts that need to
247 // trigger exceptions go the slow route. Also, it must be small enough so
248 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
249 //=============================================================================j//
250 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
251 // The allocator will coalesce int->oop copies away. See comment in
252 // coalesce.cpp about how this works. It depends critically on the exact
253 // code shape produced here, so if you are changing this code shape
254 // make sure the GC info for the heap-top is correct in and around the
255 // slow-path call.
256 //
257
258 void PhaseMacroExpand::expand_allocate_common(
259 AllocateNode* alloc, // allocation node to be expanded
260 Node* length, // array length for an array allocation
261 const TypeFunc* slow_call_type, // Type of slow call
262 address slow_call_address // Address of slow call
263 )
264 {
265
266 Node* ctrl = alloc->in(TypeFunc::Control);
267 Node* mem = alloc->in(TypeFunc::Memory);
268 Node* i_o = alloc->in(TypeFunc::I_O);
269 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
270 Node* klass_node = alloc->in(AllocateNode::KlassNode);
271 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
272
273 Node* eden_top_adr;
274 Node* eden_end_adr;
275 set_eden_pointers(eden_top_adr, eden_end_adr);
276
277 uint raw_idx = C->get_alias_index(TypeRawPtr::BOTTOM);
278 assert(ctrl != NULL, "must have control");
279
280 // Load Eden::end. Loop invariant and hoisted.
281 //
282 // Note: We set the control input on "eden_end" and "old_eden_top" when using
283 // a TLAB to work around a bug where these values were being moved across
284 // a safepoint. These are not oops, so they cannot be include in the oop
285 // map, but the can be changed by a GC. The proper way to fix this would
286 // be to set the raw memory state when generating a SafepointNode. However
287 // this will require extensive changes to the loop optimization in order to
288 // prevent a degradation of the optimization.
289 // See comment in memnode.hpp, around line 227 in class LoadPNode.
290 Node* eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
291
292 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
293 // they will not be used if "always_slow" is set
294 enum { slow_result_path = 1, fast_result_path = 2 };
295 Node *result_region;
296 Node *result_phi_rawmem;
297 Node *result_phi_rawoop;
298 Node *result_phi_i_o;
299
300 // The initial slow comparison is a size check, the comparison
301 // we want to do is a BoolTest::gt
302 bool always_slow = false;
303 int tv = _igvn.find_int_con(initial_slow_test, -1);
304 if (tv >= 0) {
305 always_slow = (tv == 1);
306 initial_slow_test = NULL;
307 } else {
308 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
309 }
310
311 if (DTraceAllocProbes) {
312 // Force slow-path allocation
313 always_slow = true;
314 initial_slow_test = NULL;
315 }
316
317 enum { too_big_or_final_path = 1, need_gc_path = 2 };
318 Node *slow_region = NULL;
319 Node *toobig_false = ctrl;
320
321 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
322 // generate the initial test if necessary
323 if (initial_slow_test != NULL ) {
324 slow_region = new (C, 3) RegionNode(3);
325
326 // Now make the initial failure test. Usually a too-big test but
327 // might be a TRUE for finalizers or a fancy class check for
328 // newInstance0.
329 IfNode *toobig_iff = new (C, 2) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
330 transform_later(toobig_iff);
331 // Plug the failing-too-big test into the slow-path region
332 Node *toobig_true = new (C, 1) IfTrueNode( toobig_iff );
333 transform_later(toobig_true);
334 slow_region ->init_req( too_big_or_final_path, toobig_true );
335 toobig_false = new (C, 1) IfFalseNode( toobig_iff );
336 transform_later(toobig_false);
337 } else { // No initial test, just fall into next case
338 toobig_false = ctrl;
339 debug_only(slow_region = NodeSentinel);
340 }
341
342 Node *slow_mem = mem; // save the current memory state for slow path
343 // generate the fast allocation code unless we know that the initial test will always go slow
344 if (!always_slow) {
345 // allocate the Region and Phi nodes for the result
346 result_region = new (C, 3) RegionNode(3);
347 result_phi_rawmem = new (C, 3) PhiNode( result_region, Type::MEMORY, TypeRawPtr::BOTTOM );
348 result_phi_rawoop = new (C, 3) PhiNode( result_region, TypeRawPtr::BOTTOM );
349 result_phi_i_o = new (C, 3) PhiNode( result_region, Type::ABIO ); // I/O is used for Prefetch
350
351 // We need a Region for the loop-back contended case.
352 enum { fall_in_path = 1, contended_loopback_path = 2 };
353 Node *contended_region;
354 Node *contended_phi_rawmem;
355 if( UseTLAB ) {
356 contended_region = toobig_false;
357 contended_phi_rawmem = mem;
358 } else {
359 contended_region = new (C, 3) RegionNode(3);
360 contended_phi_rawmem = new (C, 3) PhiNode( contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
361 // Now handle the passing-too-big test. We fall into the contended
362 // loop-back merge point.
363 contended_region ->init_req( fall_in_path, toobig_false );
364 contended_phi_rawmem->init_req( fall_in_path, mem );
365 transform_later(contended_region);
366 transform_later(contended_phi_rawmem);
367 }
368
369 // Load(-locked) the heap top.
370 // See note above concerning the control input when using a TLAB
371 Node *old_eden_top = UseTLAB
372 ? new (C, 3) LoadPNode ( ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM )
373 : new (C, 3) LoadPLockedNode( contended_region, contended_phi_rawmem, eden_top_adr );
374
375 transform_later(old_eden_top);
376 // Add to heap top to get a new heap top
377 Node *new_eden_top = new (C, 4) AddPNode( top(), old_eden_top, size_in_bytes );
378 transform_later(new_eden_top);
379 // Check for needing a GC; compare against heap end
380 Node *needgc_cmp = new (C, 3) CmpPNode( new_eden_top, eden_end );
381 transform_later(needgc_cmp);
382 Node *needgc_bol = new (C, 2) BoolNode( needgc_cmp, BoolTest::ge );
383 transform_later(needgc_bol);
384 IfNode *needgc_iff = new (C, 2) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
385 transform_later(needgc_iff);
386
387 // Plug the failing-heap-space-need-gc test into the slow-path region
388 Node *needgc_true = new (C, 1) IfTrueNode( needgc_iff );
389 transform_later(needgc_true);
390 if( initial_slow_test ) {
391 slow_region ->init_req( need_gc_path, needgc_true );
392 // This completes all paths into the slow merge point
393 transform_later(slow_region);
394 } else { // No initial slow path needed!
395 // Just fall from the need-GC path straight into the VM call.
396 slow_region = needgc_true;
397 }
398 // No need for a GC. Setup for the Store-Conditional
399 Node *needgc_false = new (C, 1) IfFalseNode( needgc_iff );
400 transform_later(needgc_false);
401
402 // Grab regular I/O before optional prefetch may change it.
403 // Slow-path does no I/O so just set it to the original I/O.
404 result_phi_i_o->init_req( slow_result_path, i_o );
405
406 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
407 old_eden_top, new_eden_top, length);
408
409 // Store (-conditional) the modified eden top back down.
410 // StorePConditional produces flags for a test PLUS a modified raw
411 // memory state.
412 Node *store_eden_top;
413 Node *fast_oop_ctrl;
414 if( UseTLAB ) {
415 store_eden_top = new (C, 4) StorePNode( needgc_false, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, new_eden_top );
416 transform_later(store_eden_top);
417 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
418 } else {
419 store_eden_top = new (C, 5) StorePConditionalNode( needgc_false, contended_phi_rawmem, eden_top_adr, new_eden_top, old_eden_top );
420 transform_later(store_eden_top);
421 Node *contention_check = new (C, 2) BoolNode( store_eden_top, BoolTest::ne );
422 transform_later(contention_check);
423 store_eden_top = new (C, 1) SCMemProjNode(store_eden_top);
424 transform_later(store_eden_top);
425
426 // If not using TLABs, check to see if there was contention.
427 IfNode *contention_iff = new (C, 2) IfNode ( needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN );
428 transform_later(contention_iff);
429 Node *contention_true = new (C, 1) IfTrueNode( contention_iff );
430 transform_later(contention_true);
431 // If contention, loopback and try again.
432 contended_region->init_req( contended_loopback_path, contention_true );
433 contended_phi_rawmem->init_req( contended_loopback_path, store_eden_top );
434
435 // Fast-path succeeded with no contention!
436 Node *contention_false = new (C, 1) IfFalseNode( contention_iff );
437 transform_later(contention_false);
438 fast_oop_ctrl = contention_false;
439 }
440
441 // Rename successful fast-path variables to make meaning more obvious
442 Node* fast_oop = old_eden_top;
443 Node* fast_oop_rawmem = store_eden_top;
444 fast_oop_rawmem = initialize_object(alloc,
445 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
446 klass_node, length, size_in_bytes);
447
448 if (ExtendedDTraceProbes) {
449 // Slow-path call
450 int size = TypeFunc::Parms + 2;
451 CallLeafNode *call = new (C, size) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
452 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
453 "dtrace_object_alloc",
454 TypeRawPtr::BOTTOM);
455
456 // Get base of thread-local storage area
457 Node* thread = new (C, 1) ThreadLocalNode();
458 transform_later(thread);
459
460 call->init_req(TypeFunc::Parms+0, thread);
461 call->init_req(TypeFunc::Parms+1, fast_oop);
462 call->init_req( TypeFunc::Control, fast_oop_ctrl );
463 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
464 call->init_req( TypeFunc::Memory , fast_oop_rawmem );
465 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
466 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
467 transform_later(call);
468 fast_oop_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control);
469 transform_later(fast_oop_ctrl);
470 fast_oop_rawmem = new (C, 1) ProjNode(call,TypeFunc::Memory);
471 transform_later(fast_oop_rawmem);
472 }
473
474 // Plug in the successful fast-path into the result merge point
475 result_region ->init_req( fast_result_path, fast_oop_ctrl );
476 result_phi_rawoop->init_req( fast_result_path, fast_oop );
477 result_phi_i_o ->init_req( fast_result_path, i_o );
478 result_phi_rawmem->init_req( fast_result_path, fast_oop_rawmem );
479 } else {
480 slow_region = ctrl;
481 }
482
483 // Generate slow-path call
484 CallNode *call = new (C, slow_call_type->domain()->cnt())
485 CallStaticJavaNode(slow_call_type, slow_call_address,
486 OptoRuntime::stub_name(slow_call_address),
487 alloc->jvms()->bci(),
488 TypePtr::BOTTOM);
489 call->init_req( TypeFunc::Control, slow_region );
490 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
491 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
492 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
493 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
494
495 call->init_req(TypeFunc::Parms+0, klass_node);
496 if (length != NULL) {
497 call->init_req(TypeFunc::Parms+1, length);
498 }
499
500 // Copy debug information and adjust JVMState information, then replace
501 // allocate node with the call
502 copy_call_debug_info((CallNode *) alloc, call);
503 if (!always_slow) {
504 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
505 }
506 _igvn.hash_delete(alloc);
507 _igvn.subsume_node(alloc, call);
508 transform_later(call);
509
510 // Identify the output projections from the allocate node and
511 // adjust any references to them.
512 // The control and io projections look like:
513 //
514 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
515 // Allocate Catch
516 // ^---Proj(io) <-------+ ^---CatchProj(io)
517 //
518 // We are interested in the CatchProj nodes.
519 //
520 extract_call_projections(call);
521
522 // An allocate node has separate memory projections for the uses on the control and i_o paths
523 // Replace uses of the control memory projection with result_phi_rawmem (unless we are only generating a slow call)
524 if (!always_slow && _memproj_fallthrough != NULL) {
525 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
526 Node *use = _memproj_fallthrough->fast_out(i);
527 _igvn.hash_delete(use);
528 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
529 _igvn._worklist.push(use);
530 // back up iterator
531 --i;
532 }
533 }
534 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete _memproj_catchall so
535 // we end up with a call that has only 1 memory projection
536 if (_memproj_catchall != NULL ) {
537 if (_memproj_fallthrough == NULL) {
538 _memproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::Memory);
539 transform_later(_memproj_fallthrough);
540 }
541 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
542 Node *use = _memproj_catchall->fast_out(i);
543 _igvn.hash_delete(use);
544 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
545 _igvn._worklist.push(use);
546 // back up iterator
547 --i;
548 }
549 }
550
551 mem = result_phi_rawmem;
552
553 // An allocate node has separate i_o projections for the uses on the control and i_o paths
554 // Replace uses of the control i_o projection with result_phi_i_o (unless we are only generating a slow call)
555 if (_ioproj_fallthrough == NULL) {
556 _ioproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::I_O);
557 transform_later(_ioproj_fallthrough);
558 } else if (!always_slow) {
559 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
560 Node *use = _ioproj_fallthrough->fast_out(i);
561
562 _igvn.hash_delete(use);
563 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
564 _igvn._worklist.push(use);
565 // back up iterator
566 --i;
567 }
568 }
569 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete _ioproj_catchall so
570 // we end up with a call that has only 1 control projection
571 if (_ioproj_catchall != NULL ) {
572 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
573 Node *use = _ioproj_catchall->fast_out(i);
574 _igvn.hash_delete(use);
575 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
576 _igvn._worklist.push(use);
577 // back up iterator
578 --i;
579 }
580 }
581
582 // if we generated only a slow call, we are done
583 if (always_slow)
584 return;
585
586
587 if (_fallthroughcatchproj != NULL) {
588 ctrl = _fallthroughcatchproj->clone();
589 transform_later(ctrl);
590 _igvn.hash_delete(_fallthroughcatchproj);
591 _igvn.subsume_node(_fallthroughcatchproj, result_region);
592 } else {
593 ctrl = top();
594 }
595 Node *slow_result;
596 if (_resproj == NULL) {
597 // no uses of the allocation result
598 slow_result = top();
599 } else {
600 slow_result = _resproj->clone();
601 transform_later(slow_result);
602 _igvn.hash_delete(_resproj);
603 _igvn.subsume_node(_resproj, result_phi_rawoop);
604 }
605
606 // Plug slow-path into result merge point
607 result_region ->init_req( slow_result_path, ctrl );
608 result_phi_rawoop->init_req( slow_result_path, slow_result);
609 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
610 transform_later(result_region);
611 transform_later(result_phi_rawoop);
612 transform_later(result_phi_rawmem);
613 transform_later(result_phi_i_o);
614 // This completes all paths into the result merge point
615 }
616
617
618 // Helper for PhaseMacroExpand::expand_allocate_common.
619 // Initializes the newly-allocated storage.
620 Node*
621 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
622 Node* control, Node* rawmem, Node* object,
623 Node* klass_node, Node* length,
624 Node* size_in_bytes) {
625 InitializeNode* init = alloc->initialization();
626 // Store the klass & mark bits
627 Node* mark_node = NULL;
628 // For now only enable fast locking for non-array types
629 if (UseBiasedLocking && (length == NULL)) {
630 mark_node = make_load(NULL, rawmem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeRawPtr::BOTTOM, T_ADDRESS);
631 } else {
632 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
633 }
634 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
635 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_OBJECT);
636 int header_size = alloc->minimum_header_size(); // conservatively small
637
638 // Array length
639 if (length != NULL) { // Arrays need length field
640 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
641 // conservatively small header size:
642 header_size = sizeof(arrayOopDesc);
643 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
644 if (k->is_array_klass()) // we know the exact header size in most cases:
645 header_size = Klass::layout_helper_header_size(k->layout_helper());
646 }
647
648 // Clear the object body, if necessary.
649 if (init == NULL) {
650 // The init has somehow disappeared; be cautious and clear everything.
651 //
652 // This can happen if a node is allocated but an uncommon trap occurs
653 // immediately. In this case, the Initialize gets associated with the
654 // trap, and may be placed in a different (outer) loop, if the Allocate
655 // is in a loop. If (this is rare) the inner loop gets unrolled, then
656 // there can be two Allocates to one Initialize. The answer in all these
657 // edge cases is safety first. It is always safe to clear immediately
658 // within an Allocate, and then (maybe or maybe not) clear some more later.
659 if (!ZeroTLAB)
660 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
661 header_size, size_in_bytes,
662 &_igvn);
663 } else {
664 if (!init->is_complete()) {
665 // Try to win by zeroing only what the init does not store.
666 // We can also try to do some peephole optimizations,
667 // such as combining some adjacent subword stores.
668 rawmem = init->complete_stores(control, rawmem, object,
669 header_size, size_in_bytes, &_igvn);
670 }
671
672 // We have no more use for this link, since the AllocateNode goes away:
673 init->set_req(InitializeNode::RawAddress, top());
674 // (If we keep the link, it just confuses the register allocator,
675 // who thinks he sees a real use of the address by the membar.)
676 }
677
678 return rawmem;
679 }
680
681 // Generate prefetch instructions for next allocations.
682 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
683 Node*& contended_phi_rawmem,
684 Node* old_eden_top, Node* new_eden_top,
685 Node* length) {
686 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
687 // Generate prefetch allocation with watermark check.
688 // As an allocation hits the watermark, we will prefetch starting
689 // at a "distance" away from watermark.
690 enum { fall_in_path = 1, pf_path = 2 };
691
692 Node *pf_region = new (C, 3) RegionNode(3);
693 Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY,
694 TypeRawPtr::BOTTOM );
695 // I/O is used for Prefetch
696 Node *pf_phi_abio = new (C, 3) PhiNode( pf_region, Type::ABIO );
697
698 Node *thread = new (C, 1) ThreadLocalNode();
699 transform_later(thread);
700
701 Node *eden_pf_adr = new (C, 4) AddPNode( top()/*not oop*/, thread,
702 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
703 transform_later(eden_pf_adr);
704
705 Node *old_pf_wm = new (C, 3) LoadPNode( needgc_false,
706 contended_phi_rawmem, eden_pf_adr,
707 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM );
708 transform_later(old_pf_wm);
709
710 // check against new_eden_top
711 Node *need_pf_cmp = new (C, 3) CmpPNode( new_eden_top, old_pf_wm );
712 transform_later(need_pf_cmp);
713 Node *need_pf_bol = new (C, 2) BoolNode( need_pf_cmp, BoolTest::ge );
714 transform_later(need_pf_bol);
715 IfNode *need_pf_iff = new (C, 2) IfNode( needgc_false, need_pf_bol,
716 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
717 transform_later(need_pf_iff);
718
719 // true node, add prefetchdistance
720 Node *need_pf_true = new (C, 1) IfTrueNode( need_pf_iff );
721 transform_later(need_pf_true);
722
723 Node *need_pf_false = new (C, 1) IfFalseNode( need_pf_iff );
724 transform_later(need_pf_false);
725
726 Node *new_pf_wmt = new (C, 4) AddPNode( top(), old_pf_wm,
727 _igvn.MakeConX(AllocatePrefetchDistance) );
728 transform_later(new_pf_wmt );
729 new_pf_wmt->set_req(0, need_pf_true);
730
731 Node *store_new_wmt = new (C, 4) StorePNode( need_pf_true,
732 contended_phi_rawmem, eden_pf_adr,
733 TypeRawPtr::BOTTOM, new_pf_wmt );
734 transform_later(store_new_wmt);
735
736 // adding prefetches
737 pf_phi_abio->init_req( fall_in_path, i_o );
738
739 Node *prefetch_adr;
740 Node *prefetch;
741 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
742 uint step_size = AllocatePrefetchStepSize;
743 uint distance = 0;
744
745 for ( uint i = 0; i < lines; i++ ) {
746 prefetch_adr = new (C, 4) AddPNode( old_pf_wm, new_pf_wmt,
747 _igvn.MakeConX(distance) );
748 transform_later(prefetch_adr);
749 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr );
750 transform_later(prefetch);
751 distance += step_size;
752 i_o = prefetch;
753 }
754 pf_phi_abio->set_req( pf_path, i_o );
755
756 pf_region->init_req( fall_in_path, need_pf_false );
757 pf_region->init_req( pf_path, need_pf_true );
758
759 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
760 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
761
762 transform_later(pf_region);
763 transform_later(pf_phi_rawmem);
764 transform_later(pf_phi_abio);
765
766 needgc_false = pf_region;
767 contended_phi_rawmem = pf_phi_rawmem;
768 i_o = pf_phi_abio;
769 } else if( AllocatePrefetchStyle > 0 ) {
770 // Insert a prefetch for each allocation only on the fast-path
771 Node *prefetch_adr;
772 Node *prefetch;
773 // Generate several prefetch instructions only for arrays.
774 uint lines = (length != NULL) ? AllocatePrefetchLines : 1;
775 uint step_size = AllocatePrefetchStepSize;
776 uint distance = AllocatePrefetchDistance;
777 for ( uint i = 0; i < lines; i++ ) {
778 prefetch_adr = new (C, 4) AddPNode( old_eden_top, new_eden_top,
779 _igvn.MakeConX(distance) );
780 transform_later(prefetch_adr);
781 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr );
782 // Do not let it float too high, since if eden_top == eden_end,
783 // both might be null.
784 if( i == 0 ) { // Set control for first prefetch, next follows it
785 prefetch->init_req(0, needgc_false);
786 }
787 transform_later(prefetch);
788 distance += step_size;
789 i_o = prefetch;
790 }
791 }
792 return i_o;
793 }
794
795
796 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
797 expand_allocate_common(alloc, NULL,
798 OptoRuntime::new_instance_Type(),
799 OptoRuntime::new_instance_Java());
800 }
801
802 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
803 Node* length = alloc->in(AllocateNode::ALength);
804 expand_allocate_common(alloc, length,
805 OptoRuntime::new_array_Type(),
806 OptoRuntime::new_array_Java());
807 }
808
809
810 // we have determined that this lock/unlock can be eliminated, we simply
811 // eliminate the node without expanding it.
812 //
813 // Note: The membar's associated with the lock/unlock are currently not
814 // eliminated. This should be investigated as a future enhancement.
815 //
816 void PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
817 Node* mem = alock->in(TypeFunc::Memory);
818
819 // The memory projection from a lock/unlock is RawMem
820 // The input to a Lock is merged memory, so extract its RawMem input
821 // (unless the MergeMem has been optimized away.)
822 if (alock->is_Lock()) {
823 if (mem->is_MergeMem())
824 mem = mem->as_MergeMem()->in(Compile::AliasIdxRaw);
825 }
826
827 extract_call_projections(alock);
828 // There are 2 projections from the lock. The lock node will
829 // be deleted when its last use is subsumed below.
830 assert(alock->outcnt() == 2 && _fallthroughproj != NULL &&
831 _memproj_fallthrough != NULL, "Unexpected projections from Lock/Unlock");
832 _igvn.hash_delete(_fallthroughproj);
833 _igvn.subsume_node(_fallthroughproj, alock->in(TypeFunc::Control));
834 _igvn.hash_delete(_memproj_fallthrough);
835 _igvn.subsume_node(_memproj_fallthrough, mem);
836 return;
837 }
838
839
840 //------------------------------expand_lock_node----------------------
841 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
842
843 Node* ctrl = lock->in(TypeFunc::Control);
844 Node* mem = lock->in(TypeFunc::Memory);
845 Node* obj = lock->obj_node();
846 Node* box = lock->box_node();
847 Node *flock = lock->fastlock_node();
848
849 if (lock->is_eliminated()) {
850 eliminate_locking_node(lock);
851 return;
852 }
853
854 // Make the merge point
855 Node *region = new (C, 3) RegionNode(3);
856
857 Node *bol = transform_later(new (C, 2) BoolNode(flock,BoolTest::ne));
858 Node *iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
859 // Optimize test; set region slot 2
860 Node *slow_path = opt_iff(region,iff);
861
862 // Make slow path call
863 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box );
864
865 extract_call_projections(call);
866
867 // Slow path can only throw asynchronous exceptions, which are always
868 // de-opted. So the compiler thinks the slow-call can never throw an
869 // exception. If it DOES throw an exception we would need the debug
870 // info removed first (since if it throws there is no monitor).
871 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
872 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
873
874 // Capture slow path
875 // disconnect fall-through projection from call and create a new one
876 // hook up users of fall-through projection to region
877 Node *slow_ctrl = _fallthroughproj->clone();
878 transform_later(slow_ctrl);
879 _igvn.hash_delete(_fallthroughproj);
880 _fallthroughproj->disconnect_inputs(NULL);
881 region->init_req(1, slow_ctrl);
882 // region inputs are now complete
883 transform_later(region);
884 _igvn.subsume_node(_fallthroughproj, region);
885
886 // create a Phi for the memory state
887 Node *mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
888 Node *memproj = transform_later( new (C, 1) ProjNode(call, TypeFunc::Memory) );
889 mem_phi->init_req(1, memproj );
890 mem_phi->init_req(2, mem);
891 transform_later(mem_phi);
892 _igvn.hash_delete(_memproj_fallthrough);
893 _igvn.subsume_node(_memproj_fallthrough, mem_phi);
894
895
896 }
897
898 //------------------------------expand_unlock_node----------------------
899 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
900
901 Node *ctrl = unlock->in(TypeFunc::Control);
902 Node* mem = unlock->in(TypeFunc::Memory);
903 Node* obj = unlock->obj_node();
904 Node* box = unlock->box_node();
905
906
907 if (unlock->is_eliminated()) {
908 eliminate_locking_node(unlock);
909 return;
910 }
911
912 // No need for a null check on unlock
913
914 // Make the merge point
915 RegionNode *region = new (C, 3) RegionNode(3);
916
917 FastUnlockNode *funlock = new (C, 3) FastUnlockNode( ctrl, obj, box );
918 funlock = transform_later( funlock )->as_FastUnlock();
919 Node *bol = transform_later(new (C, 2) BoolNode(funlock,BoolTest::ne));
920 Node *iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
921 // Optimize test; set region slot 2
922 Node *slow_path = opt_iff(region,iff);
923
924 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 );
925
926 extract_call_projections(call);
927
928 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
929 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
930
931 // No exceptions for unlocking
932 // Capture slow path
933 // disconnect fall-through projection from call and create a new one
934 // hook up users of fall-through projection to region
935 Node *slow_ctrl = _fallthroughproj->clone();
936 transform_later(slow_ctrl);
937 _igvn.hash_delete(_fallthroughproj);
938 _fallthroughproj->disconnect_inputs(NULL);
939 region->init_req(1, slow_ctrl);
940 // region inputs are now complete
941 transform_later(region);
942 _igvn.subsume_node(_fallthroughproj, region);
943
944 // create a Phi for the memory state
945 Node *mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
946 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) );
947 mem_phi->init_req(1, memproj );
948 mem_phi->init_req(2, mem);
949 transform_later(mem_phi);
950 _igvn.hash_delete(_memproj_fallthrough);
951 _igvn.subsume_node(_memproj_fallthrough, mem_phi);
952
953
954 }
955
956 //------------------------------expand_macro_nodes----------------------
957 // Returns true if a failure occurred.
958 bool PhaseMacroExpand::expand_macro_nodes() {
959 if (C->macro_count() == 0)
960 return false;
961 // Make sure expansion will not cause node limit to be exceeded. Worst case is a
962 // macro node gets expanded into about 50 nodes. Allow 50% more for optimization
963 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) )
964 return true;
965 // expand "macro" nodes
966 // nodes are removed from the macro list as they are processed
967 while (C->macro_count() > 0) {
968 Node * n = C->macro_node(0);
969 assert(n->is_macro(), "only macro nodes expected here");
970 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
971 // node is unreachable, so don't try to expand it
972 C->remove_macro_node(n);
973 continue;
974 }
975 switch (n->class_id()) {
976 case Node::Class_Allocate:
977 expand_allocate(n->as_Allocate());
978 break;
979 case Node::Class_AllocateArray:
980 expand_allocate_array(n->as_AllocateArray());
981 break;
982 case Node::Class_Lock:
983 expand_lock_node(n->as_Lock());
984 break;
985 case Node::Class_Unlock:
986 expand_unlock_node(n->as_Unlock());
987 break;
988 default:
989 assert(false, "unknown node type in macro list");
990 }
991 if (C->failing()) return true;
992 }
993 _igvn.optimize();
994 return false;
995 }