1
|
1 |
/*
|
|
2 |
* Copyright 2001-2007 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/_graphKit.cpp.incl"
|
|
27 |
|
|
28 |
//----------------------------GraphKit-----------------------------------------
|
|
29 |
// Main utility constructor.
|
|
30 |
GraphKit::GraphKit(JVMState* jvms)
|
|
31 |
: Phase(Phase::Parser),
|
|
32 |
_env(C->env()),
|
|
33 |
_gvn(*C->initial_gvn())
|
|
34 |
{
|
|
35 |
_exceptions = jvms->map()->next_exception();
|
|
36 |
if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
|
|
37 |
set_jvms(jvms);
|
|
38 |
}
|
|
39 |
|
|
40 |
// Private constructor for parser.
|
|
41 |
GraphKit::GraphKit()
|
|
42 |
: Phase(Phase::Parser),
|
|
43 |
_env(C->env()),
|
|
44 |
_gvn(*C->initial_gvn())
|
|
45 |
{
|
|
46 |
_exceptions = NULL;
|
|
47 |
set_map(NULL);
|
|
48 |
debug_only(_sp = -99);
|
|
49 |
debug_only(set_bci(-99));
|
|
50 |
}
|
|
51 |
|
|
52 |
|
|
53 |
|
|
54 |
//---------------------------clean_stack---------------------------------------
|
|
55 |
// Clear away rubbish from the stack area of the JVM state.
|
|
56 |
// This destroys any arguments that may be waiting on the stack.
|
|
57 |
void GraphKit::clean_stack(int from_sp) {
|
|
58 |
SafePointNode* map = this->map();
|
|
59 |
JVMState* jvms = this->jvms();
|
|
60 |
int stk_size = jvms->stk_size();
|
|
61 |
int stkoff = jvms->stkoff();
|
|
62 |
Node* top = this->top();
|
|
63 |
for (int i = from_sp; i < stk_size; i++) {
|
|
64 |
if (map->in(stkoff + i) != top) {
|
|
65 |
map->set_req(stkoff + i, top);
|
|
66 |
}
|
|
67 |
}
|
|
68 |
}
|
|
69 |
|
|
70 |
|
|
71 |
//--------------------------------sync_jvms-----------------------------------
|
|
72 |
// Make sure our current jvms agrees with our parse state.
|
|
73 |
JVMState* GraphKit::sync_jvms() const {
|
|
74 |
JVMState* jvms = this->jvms();
|
|
75 |
jvms->set_bci(bci()); // Record the new bci in the JVMState
|
|
76 |
jvms->set_sp(sp()); // Record the new sp in the JVMState
|
|
77 |
assert(jvms_in_sync(), "jvms is now in sync");
|
|
78 |
return jvms;
|
|
79 |
}
|
|
80 |
|
|
81 |
#ifdef ASSERT
|
|
82 |
bool GraphKit::jvms_in_sync() const {
|
|
83 |
Parse* parse = is_Parse();
|
|
84 |
if (parse == NULL) {
|
|
85 |
if (bci() != jvms()->bci()) return false;
|
|
86 |
if (sp() != (int)jvms()->sp()) return false;
|
|
87 |
return true;
|
|
88 |
}
|
|
89 |
if (jvms()->method() != parse->method()) return false;
|
|
90 |
if (jvms()->bci() != parse->bci()) return false;
|
|
91 |
int jvms_sp = jvms()->sp();
|
|
92 |
if (jvms_sp != parse->sp()) return false;
|
|
93 |
int jvms_depth = jvms()->depth();
|
|
94 |
if (jvms_depth != parse->depth()) return false;
|
|
95 |
return true;
|
|
96 |
}
|
|
97 |
|
|
98 |
// Local helper checks for special internal merge points
|
|
99 |
// used to accumulate and merge exception states.
|
|
100 |
// They are marked by the region's in(0) edge being the map itself.
|
|
101 |
// Such merge points must never "escape" into the parser at large,
|
|
102 |
// until they have been handed to gvn.transform.
|
|
103 |
static bool is_hidden_merge(Node* reg) {
|
|
104 |
if (reg == NULL) return false;
|
|
105 |
if (reg->is_Phi()) {
|
|
106 |
reg = reg->in(0);
|
|
107 |
if (reg == NULL) return false;
|
|
108 |
}
|
|
109 |
return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
|
|
110 |
}
|
|
111 |
|
|
112 |
void GraphKit::verify_map() const {
|
|
113 |
if (map() == NULL) return; // null map is OK
|
|
114 |
assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
|
|
115 |
assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
|
|
116 |
assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
|
|
117 |
}
|
|
118 |
|
|
119 |
void GraphKit::verify_exception_state(SafePointNode* ex_map) {
|
|
120 |
assert(ex_map->next_exception() == NULL, "not already part of a chain");
|
|
121 |
assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
|
|
122 |
}
|
|
123 |
#endif
|
|
124 |
|
|
125 |
//---------------------------stop_and_kill_map---------------------------------
|
|
126 |
// Set _map to NULL, signalling a stop to further bytecode execution.
|
|
127 |
// First smash the current map's control to a constant, to mark it dead.
|
|
128 |
void GraphKit::stop_and_kill_map() {
|
|
129 |
SafePointNode* dead_map = stop();
|
|
130 |
if (dead_map != NULL) {
|
|
131 |
dead_map->disconnect_inputs(NULL); // Mark the map as killed.
|
|
132 |
assert(dead_map->is_killed(), "must be so marked");
|
|
133 |
}
|
|
134 |
}
|
|
135 |
|
|
136 |
|
|
137 |
//--------------------------------stopped--------------------------------------
|
|
138 |
// Tell if _map is NULL, or control is top.
|
|
139 |
bool GraphKit::stopped() {
|
|
140 |
if (map() == NULL) return true;
|
|
141 |
else if (control() == top()) return true;
|
|
142 |
else return false;
|
|
143 |
}
|
|
144 |
|
|
145 |
|
|
146 |
//-----------------------------has_ex_handler----------------------------------
|
|
147 |
// Tell if this method or any caller method has exception handlers.
|
|
148 |
bool GraphKit::has_ex_handler() {
|
|
149 |
for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
|
|
150 |
if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
|
|
151 |
return true;
|
|
152 |
}
|
|
153 |
}
|
|
154 |
return false;
|
|
155 |
}
|
|
156 |
|
|
157 |
//------------------------------save_ex_oop------------------------------------
|
|
158 |
// Save an exception without blowing stack contents or other JVM state.
|
|
159 |
void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
|
|
160 |
assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
|
|
161 |
ex_map->add_req(ex_oop);
|
|
162 |
debug_only(verify_exception_state(ex_map));
|
|
163 |
}
|
|
164 |
|
|
165 |
inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
|
|
166 |
assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
|
|
167 |
Node* ex_oop = ex_map->in(ex_map->req()-1);
|
|
168 |
if (clear_it) ex_map->del_req(ex_map->req()-1);
|
|
169 |
return ex_oop;
|
|
170 |
}
|
|
171 |
|
|
172 |
//-----------------------------saved_ex_oop------------------------------------
|
|
173 |
// Recover a saved exception from its map.
|
|
174 |
Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
|
|
175 |
return common_saved_ex_oop(ex_map, false);
|
|
176 |
}
|
|
177 |
|
|
178 |
//--------------------------clear_saved_ex_oop---------------------------------
|
|
179 |
// Erase a previously saved exception from its map.
|
|
180 |
Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
|
|
181 |
return common_saved_ex_oop(ex_map, true);
|
|
182 |
}
|
|
183 |
|
|
184 |
#ifdef ASSERT
|
|
185 |
//---------------------------has_saved_ex_oop----------------------------------
|
|
186 |
// Erase a previously saved exception from its map.
|
|
187 |
bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
|
|
188 |
return ex_map->req() == ex_map->jvms()->endoff()+1;
|
|
189 |
}
|
|
190 |
#endif
|
|
191 |
|
|
192 |
//-------------------------make_exception_state--------------------------------
|
|
193 |
// Turn the current JVM state into an exception state, appending the ex_oop.
|
|
194 |
SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
|
|
195 |
sync_jvms();
|
|
196 |
SafePointNode* ex_map = stop(); // do not manipulate this map any more
|
|
197 |
set_saved_ex_oop(ex_map, ex_oop);
|
|
198 |
return ex_map;
|
|
199 |
}
|
|
200 |
|
|
201 |
|
|
202 |
//--------------------------add_exception_state--------------------------------
|
|
203 |
// Add an exception to my list of exceptions.
|
|
204 |
void GraphKit::add_exception_state(SafePointNode* ex_map) {
|
|
205 |
if (ex_map == NULL || ex_map->control() == top()) {
|
|
206 |
return;
|
|
207 |
}
|
|
208 |
#ifdef ASSERT
|
|
209 |
verify_exception_state(ex_map);
|
|
210 |
if (has_exceptions()) {
|
|
211 |
assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
|
|
212 |
}
|
|
213 |
#endif
|
|
214 |
|
|
215 |
// If there is already an exception of exactly this type, merge with it.
|
|
216 |
// In particular, null-checks and other low-level exceptions common up here.
|
|
217 |
Node* ex_oop = saved_ex_oop(ex_map);
|
|
218 |
const Type* ex_type = _gvn.type(ex_oop);
|
|
219 |
if (ex_oop == top()) {
|
|
220 |
// No action needed.
|
|
221 |
return;
|
|
222 |
}
|
|
223 |
assert(ex_type->isa_instptr(), "exception must be an instance");
|
|
224 |
for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
|
|
225 |
const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
|
|
226 |
// We check sp also because call bytecodes can generate exceptions
|
|
227 |
// both before and after arguments are popped!
|
|
228 |
if (ex_type2 == ex_type
|
|
229 |
&& e2->_jvms->sp() == ex_map->_jvms->sp()) {
|
|
230 |
combine_exception_states(ex_map, e2);
|
|
231 |
return;
|
|
232 |
}
|
|
233 |
}
|
|
234 |
|
|
235 |
// No pre-existing exception of the same type. Chain it on the list.
|
|
236 |
push_exception_state(ex_map);
|
|
237 |
}
|
|
238 |
|
|
239 |
//-----------------------add_exception_states_from-----------------------------
|
|
240 |
void GraphKit::add_exception_states_from(JVMState* jvms) {
|
|
241 |
SafePointNode* ex_map = jvms->map()->next_exception();
|
|
242 |
if (ex_map != NULL) {
|
|
243 |
jvms->map()->set_next_exception(NULL);
|
|
244 |
for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
|
|
245 |
next_map = ex_map->next_exception();
|
|
246 |
ex_map->set_next_exception(NULL);
|
|
247 |
add_exception_state(ex_map);
|
|
248 |
}
|
|
249 |
}
|
|
250 |
}
|
|
251 |
|
|
252 |
//-----------------------transfer_exceptions_into_jvms-------------------------
|
|
253 |
JVMState* GraphKit::transfer_exceptions_into_jvms() {
|
|
254 |
if (map() == NULL) {
|
|
255 |
// We need a JVMS to carry the exceptions, but the map has gone away.
|
|
256 |
// Create a scratch JVMS, cloned from any of the exception states...
|
|
257 |
if (has_exceptions()) {
|
|
258 |
_map = _exceptions;
|
|
259 |
_map = clone_map();
|
|
260 |
_map->set_next_exception(NULL);
|
|
261 |
clear_saved_ex_oop(_map);
|
|
262 |
debug_only(verify_map());
|
|
263 |
} else {
|
|
264 |
// ...or created from scratch
|
|
265 |
JVMState* jvms = new (C) JVMState(_method, NULL);
|
|
266 |
jvms->set_bci(_bci);
|
|
267 |
jvms->set_sp(_sp);
|
|
268 |
jvms->set_map(new (C, TypeFunc::Parms) SafePointNode(TypeFunc::Parms, jvms));
|
|
269 |
set_jvms(jvms);
|
|
270 |
for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
|
|
271 |
set_all_memory(top());
|
|
272 |
while (map()->req() < jvms->endoff()) map()->add_req(top());
|
|
273 |
}
|
|
274 |
// (This is a kludge, in case you didn't notice.)
|
|
275 |
set_control(top());
|
|
276 |
}
|
|
277 |
JVMState* jvms = sync_jvms();
|
|
278 |
assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
|
|
279 |
jvms->map()->set_next_exception(_exceptions);
|
|
280 |
_exceptions = NULL; // done with this set of exceptions
|
|
281 |
return jvms;
|
|
282 |
}
|
|
283 |
|
|
284 |
static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
|
|
285 |
assert(is_hidden_merge(dstphi), "must be a special merge node");
|
|
286 |
assert(is_hidden_merge(srcphi), "must be a special merge node");
|
|
287 |
uint limit = srcphi->req();
|
|
288 |
for (uint i = PhiNode::Input; i < limit; i++) {
|
|
289 |
dstphi->add_req(srcphi->in(i));
|
|
290 |
}
|
|
291 |
}
|
|
292 |
static inline void add_one_req(Node* dstphi, Node* src) {
|
|
293 |
assert(is_hidden_merge(dstphi), "must be a special merge node");
|
|
294 |
assert(!is_hidden_merge(src), "must not be a special merge node");
|
|
295 |
dstphi->add_req(src);
|
|
296 |
}
|
|
297 |
|
|
298 |
//-----------------------combine_exception_states------------------------------
|
|
299 |
// This helper function combines exception states by building phis on a
|
|
300 |
// specially marked state-merging region. These regions and phis are
|
|
301 |
// untransformed, and can build up gradually. The region is marked by
|
|
302 |
// having a control input of its exception map, rather than NULL. Such
|
|
303 |
// regions do not appear except in this function, and in use_exception_state.
|
|
304 |
void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
|
|
305 |
if (failing()) return; // dying anyway...
|
|
306 |
JVMState* ex_jvms = ex_map->_jvms;
|
|
307 |
assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
|
|
308 |
assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
|
|
309 |
assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
|
|
310 |
assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
|
|
311 |
assert(ex_map->req() == phi_map->req(), "matching maps");
|
|
312 |
uint tos = ex_jvms->stkoff() + ex_jvms->sp();
|
|
313 |
Node* hidden_merge_mark = root();
|
|
314 |
Node* region = phi_map->control();
|
|
315 |
MergeMemNode* phi_mem = phi_map->merged_memory();
|
|
316 |
MergeMemNode* ex_mem = ex_map->merged_memory();
|
|
317 |
if (region->in(0) != hidden_merge_mark) {
|
|
318 |
// The control input is not (yet) a specially-marked region in phi_map.
|
|
319 |
// Make it so, and build some phis.
|
|
320 |
region = new (C, 2) RegionNode(2);
|
|
321 |
_gvn.set_type(region, Type::CONTROL);
|
|
322 |
region->set_req(0, hidden_merge_mark); // marks an internal ex-state
|
|
323 |
region->init_req(1, phi_map->control());
|
|
324 |
phi_map->set_control(region);
|
|
325 |
Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
|
|
326 |
record_for_igvn(io_phi);
|
|
327 |
_gvn.set_type(io_phi, Type::ABIO);
|
|
328 |
phi_map->set_i_o(io_phi);
|
|
329 |
for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
|
|
330 |
Node* m = mms.memory();
|
|
331 |
Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
|
|
332 |
record_for_igvn(m_phi);
|
|
333 |
_gvn.set_type(m_phi, Type::MEMORY);
|
|
334 |
mms.set_memory(m_phi);
|
|
335 |
}
|
|
336 |
}
|
|
337 |
|
|
338 |
// Either or both of phi_map and ex_map might already be converted into phis.
|
|
339 |
Node* ex_control = ex_map->control();
|
|
340 |
// if there is special marking on ex_map also, we add multiple edges from src
|
|
341 |
bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
|
|
342 |
// how wide was the destination phi_map, originally?
|
|
343 |
uint orig_width = region->req();
|
|
344 |
|
|
345 |
if (add_multiple) {
|
|
346 |
add_n_reqs(region, ex_control);
|
|
347 |
add_n_reqs(phi_map->i_o(), ex_map->i_o());
|
|
348 |
} else {
|
|
349 |
// ex_map has no merges, so we just add single edges everywhere
|
|
350 |
add_one_req(region, ex_control);
|
|
351 |
add_one_req(phi_map->i_o(), ex_map->i_o());
|
|
352 |
}
|
|
353 |
for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
|
|
354 |
if (mms.is_empty()) {
|
|
355 |
// get a copy of the base memory, and patch some inputs into it
|
|
356 |
const TypePtr* adr_type = mms.adr_type(C);
|
|
357 |
Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
|
|
358 |
assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
|
|
359 |
mms.set_memory(phi);
|
|
360 |
// Prepare to append interesting stuff onto the newly sliced phi:
|
|
361 |
while (phi->req() > orig_width) phi->del_req(phi->req()-1);
|
|
362 |
}
|
|
363 |
// Append stuff from ex_map:
|
|
364 |
if (add_multiple) {
|
|
365 |
add_n_reqs(mms.memory(), mms.memory2());
|
|
366 |
} else {
|
|
367 |
add_one_req(mms.memory(), mms.memory2());
|
|
368 |
}
|
|
369 |
}
|
|
370 |
uint limit = ex_map->req();
|
|
371 |
for (uint i = TypeFunc::Parms; i < limit; i++) {
|
|
372 |
// Skip everything in the JVMS after tos. (The ex_oop follows.)
|
|
373 |
if (i == tos) i = ex_jvms->monoff();
|
|
374 |
Node* src = ex_map->in(i);
|
|
375 |
Node* dst = phi_map->in(i);
|
|
376 |
if (src != dst) {
|
|
377 |
PhiNode* phi;
|
|
378 |
if (dst->in(0) != region) {
|
|
379 |
dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
|
|
380 |
record_for_igvn(phi);
|
|
381 |
_gvn.set_type(phi, phi->type());
|
|
382 |
phi_map->set_req(i, dst);
|
|
383 |
// Prepare to append interesting stuff onto the new phi:
|
|
384 |
while (dst->req() > orig_width) dst->del_req(dst->req()-1);
|
|
385 |
} else {
|
|
386 |
assert(dst->is_Phi(), "nobody else uses a hidden region");
|
|
387 |
phi = (PhiNode*)dst;
|
|
388 |
}
|
|
389 |
if (add_multiple && src->in(0) == ex_control) {
|
|
390 |
// Both are phis.
|
|
391 |
add_n_reqs(dst, src);
|
|
392 |
} else {
|
|
393 |
while (dst->req() < region->req()) add_one_req(dst, src);
|
|
394 |
}
|
|
395 |
const Type* srctype = _gvn.type(src);
|
|
396 |
if (phi->type() != srctype) {
|
|
397 |
const Type* dsttype = phi->type()->meet(srctype);
|
|
398 |
if (phi->type() != dsttype) {
|
|
399 |
phi->set_type(dsttype);
|
|
400 |
_gvn.set_type(phi, dsttype);
|
|
401 |
}
|
|
402 |
}
|
|
403 |
}
|
|
404 |
}
|
|
405 |
}
|
|
406 |
|
|
407 |
//--------------------------use_exception_state--------------------------------
|
|
408 |
Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
|
|
409 |
if (failing()) { stop(); return top(); }
|
|
410 |
Node* region = phi_map->control();
|
|
411 |
Node* hidden_merge_mark = root();
|
|
412 |
assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
|
|
413 |
Node* ex_oop = clear_saved_ex_oop(phi_map);
|
|
414 |
if (region->in(0) == hidden_merge_mark) {
|
|
415 |
// Special marking for internal ex-states. Process the phis now.
|
|
416 |
region->set_req(0, region); // now it's an ordinary region
|
|
417 |
set_jvms(phi_map->jvms()); // ...so now we can use it as a map
|
|
418 |
// Note: Setting the jvms also sets the bci and sp.
|
|
419 |
set_control(_gvn.transform(region));
|
|
420 |
uint tos = jvms()->stkoff() + sp();
|
|
421 |
for (uint i = 1; i < tos; i++) {
|
|
422 |
Node* x = phi_map->in(i);
|
|
423 |
if (x->in(0) == region) {
|
|
424 |
assert(x->is_Phi(), "expected a special phi");
|
|
425 |
phi_map->set_req(i, _gvn.transform(x));
|
|
426 |
}
|
|
427 |
}
|
|
428 |
for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
|
|
429 |
Node* x = mms.memory();
|
|
430 |
if (x->in(0) == region) {
|
|
431 |
assert(x->is_Phi(), "nobody else uses a hidden region");
|
|
432 |
mms.set_memory(_gvn.transform(x));
|
|
433 |
}
|
|
434 |
}
|
|
435 |
if (ex_oop->in(0) == region) {
|
|
436 |
assert(ex_oop->is_Phi(), "expected a special phi");
|
|
437 |
ex_oop = _gvn.transform(ex_oop);
|
|
438 |
}
|
|
439 |
} else {
|
|
440 |
set_jvms(phi_map->jvms());
|
|
441 |
}
|
|
442 |
|
|
443 |
assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
|
|
444 |
assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
|
|
445 |
return ex_oop;
|
|
446 |
}
|
|
447 |
|
|
448 |
//---------------------------------java_bc-------------------------------------
|
|
449 |
Bytecodes::Code GraphKit::java_bc() const {
|
|
450 |
ciMethod* method = this->method();
|
|
451 |
int bci = this->bci();
|
|
452 |
if (method != NULL && bci != InvocationEntryBci)
|
|
453 |
return method->java_code_at_bci(bci);
|
|
454 |
else
|
|
455 |
return Bytecodes::_illegal;
|
|
456 |
}
|
|
457 |
|
|
458 |
//------------------------------builtin_throw----------------------------------
|
|
459 |
void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
|
|
460 |
bool must_throw = true;
|
|
461 |
|
|
462 |
if (JvmtiExport::can_post_exceptions()) {
|
|
463 |
// Do not try anything fancy if we're notifying the VM on every throw.
|
|
464 |
// Cf. case Bytecodes::_athrow in parse2.cpp.
|
|
465 |
uncommon_trap(reason, Deoptimization::Action_none,
|
|
466 |
(ciKlass*)NULL, (char*)NULL, must_throw);
|
|
467 |
return;
|
|
468 |
}
|
|
469 |
|
|
470 |
// If this particular condition has not yet happened at this
|
|
471 |
// bytecode, then use the uncommon trap mechanism, and allow for
|
|
472 |
// a future recompilation if several traps occur here.
|
|
473 |
// If the throw is hot, try to use a more complicated inline mechanism
|
|
474 |
// which keeps execution inside the compiled code.
|
|
475 |
bool treat_throw_as_hot = false;
|
|
476 |
ciMethodData* md = method()->method_data();
|
|
477 |
|
|
478 |
if (ProfileTraps) {
|
|
479 |
if (too_many_traps(reason)) {
|
|
480 |
treat_throw_as_hot = true;
|
|
481 |
}
|
|
482 |
// (If there is no MDO at all, assume it is early in
|
|
483 |
// execution, and that any deopts are part of the
|
|
484 |
// startup transient, and don't need to be remembered.)
|
|
485 |
|
|
486 |
// Also, if there is a local exception handler, treat all throws
|
|
487 |
// as hot if there has been at least one in this method.
|
|
488 |
if (C->trap_count(reason) != 0
|
|
489 |
&& method()->method_data()->trap_count(reason) != 0
|
|
490 |
&& has_ex_handler()) {
|
|
491 |
treat_throw_as_hot = true;
|
|
492 |
}
|
|
493 |
}
|
|
494 |
|
|
495 |
// If this throw happens frequently, an uncommon trap might cause
|
|
496 |
// a performance pothole. If there is a local exception handler,
|
|
497 |
// and if this particular bytecode appears to be deoptimizing often,
|
|
498 |
// let us handle the throw inline, with a preconstructed instance.
|
|
499 |
// Note: If the deopt count has blown up, the uncommon trap
|
|
500 |
// runtime is going to flush this nmethod, not matter what.
|
|
501 |
if (treat_throw_as_hot
|
|
502 |
&& (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
|
|
503 |
// If the throw is local, we use a pre-existing instance and
|
|
504 |
// punt on the backtrace. This would lead to a missing backtrace
|
|
505 |
// (a repeat of 4292742) if the backtrace object is ever asked
|
|
506 |
// for its backtrace.
|
|
507 |
// Fixing this remaining case of 4292742 requires some flavor of
|
|
508 |
// escape analysis. Leave that for the future.
|
|
509 |
ciInstance* ex_obj = NULL;
|
|
510 |
switch (reason) {
|
|
511 |
case Deoptimization::Reason_null_check:
|
|
512 |
ex_obj = env()->NullPointerException_instance();
|
|
513 |
break;
|
|
514 |
case Deoptimization::Reason_div0_check:
|
|
515 |
ex_obj = env()->ArithmeticException_instance();
|
|
516 |
break;
|
|
517 |
case Deoptimization::Reason_range_check:
|
|
518 |
ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
|
|
519 |
break;
|
|
520 |
case Deoptimization::Reason_class_check:
|
|
521 |
if (java_bc() == Bytecodes::_aastore) {
|
|
522 |
ex_obj = env()->ArrayStoreException_instance();
|
|
523 |
} else {
|
|
524 |
ex_obj = env()->ClassCastException_instance();
|
|
525 |
}
|
|
526 |
break;
|
|
527 |
}
|
|
528 |
if (failing()) { stop(); return; } // exception allocation might fail
|
|
529 |
if (ex_obj != NULL) {
|
|
530 |
// Cheat with a preallocated exception object.
|
|
531 |
if (C->log() != NULL)
|
|
532 |
C->log()->elem("hot_throw preallocated='1' reason='%s'",
|
|
533 |
Deoptimization::trap_reason_name(reason));
|
|
534 |
const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
|
|
535 |
Node* ex_node = _gvn.transform(new (C, 1) ConPNode(ex_con));
|
|
536 |
|
|
537 |
// Clear the detail message of the preallocated exception object.
|
|
538 |
// Weblogic sometimes mutates the detail message of exceptions
|
|
539 |
// using reflection.
|
|
540 |
int offset = java_lang_Throwable::get_detailMessage_offset();
|
|
541 |
const TypePtr* adr_typ = ex_con->add_offset(offset);
|
|
542 |
|
|
543 |
Node *adr = basic_plus_adr(ex_node, ex_node, offset);
|
|
544 |
Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), ex_con, T_OBJECT);
|
|
545 |
|
|
546 |
add_exception_state(make_exception_state(ex_node));
|
|
547 |
return;
|
|
548 |
}
|
|
549 |
}
|
|
550 |
|
|
551 |
// %%% Maybe add entry to OptoRuntime which directly throws the exc.?
|
|
552 |
// It won't be much cheaper than bailing to the interp., since we'll
|
|
553 |
// have to pass up all the debug-info, and the runtime will have to
|
|
554 |
// create the stack trace.
|
|
555 |
|
|
556 |
// Usual case: Bail to interpreter.
|
|
557 |
// Reserve the right to recompile if we haven't seen anything yet.
|
|
558 |
|
|
559 |
Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
|
|
560 |
if (treat_throw_as_hot
|
|
561 |
&& (method()->method_data()->trap_recompiled_at(bci())
|
|
562 |
|| C->too_many_traps(reason))) {
|
|
563 |
// We cannot afford to take more traps here. Suffer in the interpreter.
|
|
564 |
if (C->log() != NULL)
|
|
565 |
C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
|
|
566 |
Deoptimization::trap_reason_name(reason),
|
|
567 |
C->trap_count(reason));
|
|
568 |
action = Deoptimization::Action_none;
|
|
569 |
}
|
|
570 |
|
|
571 |
// "must_throw" prunes the JVM state to include only the stack, if there
|
|
572 |
// are no local exception handlers. This should cut down on register
|
|
573 |
// allocation time and code size, by drastically reducing the number
|
|
574 |
// of in-edges on the call to the uncommon trap.
|
|
575 |
|
|
576 |
uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
|
|
577 |
}
|
|
578 |
|
|
579 |
|
|
580 |
//----------------------------PreserveJVMState---------------------------------
|
|
581 |
PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
|
|
582 |
debug_only(kit->verify_map());
|
|
583 |
_kit = kit;
|
|
584 |
_map = kit->map(); // preserve the map
|
|
585 |
_sp = kit->sp();
|
|
586 |
kit->set_map(clone_map ? kit->clone_map() : NULL);
|
|
587 |
#ifdef ASSERT
|
|
588 |
_bci = kit->bci();
|
|
589 |
Parse* parser = kit->is_Parse();
|
|
590 |
int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->pre_order();
|
|
591 |
_block = block;
|
|
592 |
#endif
|
|
593 |
}
|
|
594 |
PreserveJVMState::~PreserveJVMState() {
|
|
595 |
GraphKit* kit = _kit;
|
|
596 |
#ifdef ASSERT
|
|
597 |
assert(kit->bci() == _bci, "bci must not shift");
|
|
598 |
Parse* parser = kit->is_Parse();
|
|
599 |
int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->pre_order();
|
|
600 |
assert(block == _block, "block must not shift");
|
|
601 |
#endif
|
|
602 |
kit->set_map(_map);
|
|
603 |
kit->set_sp(_sp);
|
|
604 |
}
|
|
605 |
|
|
606 |
|
|
607 |
//-----------------------------BuildCutout-------------------------------------
|
|
608 |
BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
|
|
609 |
: PreserveJVMState(kit)
|
|
610 |
{
|
|
611 |
assert(p->is_Con() || p->is_Bool(), "test must be a bool");
|
|
612 |
SafePointNode* outer_map = _map; // preserved map is caller's
|
|
613 |
SafePointNode* inner_map = kit->map();
|
|
614 |
IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
|
|
615 |
outer_map->set_control(kit->gvn().transform( new (kit->C, 1) IfTrueNode(iff) ));
|
|
616 |
inner_map->set_control(kit->gvn().transform( new (kit->C, 1) IfFalseNode(iff) ));
|
|
617 |
}
|
|
618 |
BuildCutout::~BuildCutout() {
|
|
619 |
GraphKit* kit = _kit;
|
|
620 |
assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
|
|
621 |
}
|
|
622 |
|
|
623 |
|
|
624 |
//------------------------------clone_map--------------------------------------
|
|
625 |
// Implementation of PreserveJVMState
|
|
626 |
//
|
|
627 |
// Only clone_map(...) here. If this function is only used in the
|
|
628 |
// PreserveJVMState class we may want to get rid of this extra
|
|
629 |
// function eventually and do it all there.
|
|
630 |
|
|
631 |
SafePointNode* GraphKit::clone_map() {
|
|
632 |
if (map() == NULL) return NULL;
|
|
633 |
|
|
634 |
// Clone the memory edge first
|
|
635 |
Node* mem = MergeMemNode::make(C, map()->memory());
|
|
636 |
gvn().set_type_bottom(mem);
|
|
637 |
|
|
638 |
SafePointNode *clonemap = (SafePointNode*)map()->clone();
|
|
639 |
JVMState* jvms = this->jvms();
|
|
640 |
JVMState* clonejvms = jvms->clone_shallow(C);
|
|
641 |
clonemap->set_memory(mem);
|
|
642 |
clonemap->set_jvms(clonejvms);
|
|
643 |
clonejvms->set_map(clonemap);
|
|
644 |
record_for_igvn(clonemap);
|
|
645 |
gvn().set_type_bottom(clonemap);
|
|
646 |
return clonemap;
|
|
647 |
}
|
|
648 |
|
|
649 |
|
|
650 |
//-----------------------------set_map_clone-----------------------------------
|
|
651 |
void GraphKit::set_map_clone(SafePointNode* m) {
|
|
652 |
_map = m;
|
|
653 |
_map = clone_map();
|
|
654 |
_map->set_next_exception(NULL);
|
|
655 |
debug_only(verify_map());
|
|
656 |
}
|
|
657 |
|
|
658 |
|
|
659 |
//----------------------------kill_dead_locals---------------------------------
|
|
660 |
// Detect any locals which are known to be dead, and force them to top.
|
|
661 |
void GraphKit::kill_dead_locals() {
|
|
662 |
// Consult the liveness information for the locals. If any
|
|
663 |
// of them are unused, then they can be replaced by top(). This
|
|
664 |
// should help register allocation time and cut down on the size
|
|
665 |
// of the deoptimization information.
|
|
666 |
|
|
667 |
// This call is made from many of the bytecode handling
|
|
668 |
// subroutines called from the Big Switch in do_one_bytecode.
|
|
669 |
// Every bytecode which might include a slow path is responsible
|
|
670 |
// for killing its dead locals. The more consistent we
|
|
671 |
// are about killing deads, the fewer useless phis will be
|
|
672 |
// constructed for them at various merge points.
|
|
673 |
|
|
674 |
// bci can be -1 (InvocationEntryBci). We return the entry
|
|
675 |
// liveness for the method.
|
|
676 |
|
|
677 |
if (method() == NULL || method()->code_size() == 0) {
|
|
678 |
// We are building a graph for a call to a native method.
|
|
679 |
// All locals are live.
|
|
680 |
return;
|
|
681 |
}
|
|
682 |
|
|
683 |
ResourceMark rm;
|
|
684 |
|
|
685 |
// Consult the liveness information for the locals. If any
|
|
686 |
// of them are unused, then they can be replaced by top(). This
|
|
687 |
// should help register allocation time and cut down on the size
|
|
688 |
// of the deoptimization information.
|
|
689 |
MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
|
|
690 |
|
|
691 |
int len = (int)live_locals.size();
|
|
692 |
assert(len <= jvms()->loc_size(), "too many live locals");
|
|
693 |
for (int local = 0; local < len; local++) {
|
|
694 |
if (!live_locals.at(local)) {
|
|
695 |
set_local(local, top());
|
|
696 |
}
|
|
697 |
}
|
|
698 |
}
|
|
699 |
|
|
700 |
#ifdef ASSERT
|
|
701 |
//-------------------------dead_locals_are_killed------------------------------
|
|
702 |
// Return true if all dead locals are set to top in the map.
|
|
703 |
// Used to assert "clean" debug info at various points.
|
|
704 |
bool GraphKit::dead_locals_are_killed() {
|
|
705 |
if (method() == NULL || method()->code_size() == 0) {
|
|
706 |
// No locals need to be dead, so all is as it should be.
|
|
707 |
return true;
|
|
708 |
}
|
|
709 |
|
|
710 |
// Make sure somebody called kill_dead_locals upstream.
|
|
711 |
ResourceMark rm;
|
|
712 |
for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
|
|
713 |
if (jvms->loc_size() == 0) continue; // no locals to consult
|
|
714 |
SafePointNode* map = jvms->map();
|
|
715 |
ciMethod* method = jvms->method();
|
|
716 |
int bci = jvms->bci();
|
|
717 |
if (jvms == this->jvms()) {
|
|
718 |
bci = this->bci(); // it might not yet be synched
|
|
719 |
}
|
|
720 |
MethodLivenessResult live_locals = method->liveness_at_bci(bci);
|
|
721 |
int len = (int)live_locals.size();
|
|
722 |
if (!live_locals.is_valid() || len == 0)
|
|
723 |
// This method is trivial, or is poisoned by a breakpoint.
|
|
724 |
return true;
|
|
725 |
assert(len == jvms->loc_size(), "live map consistent with locals map");
|
|
726 |
for (int local = 0; local < len; local++) {
|
|
727 |
if (!live_locals.at(local) && map->local(jvms, local) != top()) {
|
|
728 |
if (PrintMiscellaneous && (Verbose || WizardMode)) {
|
|
729 |
tty->print_cr("Zombie local %d: ", local);
|
|
730 |
jvms->dump();
|
|
731 |
}
|
|
732 |
return false;
|
|
733 |
}
|
|
734 |
}
|
|
735 |
}
|
|
736 |
return true;
|
|
737 |
}
|
|
738 |
|
|
739 |
#endif //ASSERT
|
|
740 |
|
|
741 |
// Helper function for adding JVMState and debug information to node
|
|
742 |
void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
|
|
743 |
// Add the safepoint edges to the call (or other safepoint).
|
|
744 |
|
|
745 |
// Make sure dead locals are set to top. This
|
|
746 |
// should help register allocation time and cut down on the size
|
|
747 |
// of the deoptimization information.
|
|
748 |
assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
|
|
749 |
|
|
750 |
// Walk the inline list to fill in the correct set of JVMState's
|
|
751 |
// Also fill in the associated edges for each JVMState.
|
|
752 |
|
|
753 |
JVMState* youngest_jvms = sync_jvms();
|
|
754 |
|
|
755 |
// Do we need debug info here? If it is a SafePoint and this method
|
|
756 |
// cannot de-opt, then we do NOT need any debug info.
|
|
757 |
bool full_info = (C->deopt_happens() || call->Opcode() != Op_SafePoint);
|
|
758 |
|
|
759 |
// If we are guaranteed to throw, we can prune everything but the
|
|
760 |
// input to the current bytecode.
|
|
761 |
bool can_prune_locals = false;
|
|
762 |
uint stack_slots_not_pruned = 0;
|
|
763 |
int inputs = 0, depth = 0;
|
|
764 |
if (must_throw) {
|
|
765 |
assert(method() == youngest_jvms->method(), "sanity");
|
|
766 |
if (compute_stack_effects(inputs, depth)) {
|
|
767 |
can_prune_locals = true;
|
|
768 |
stack_slots_not_pruned = inputs;
|
|
769 |
}
|
|
770 |
}
|
|
771 |
|
|
772 |
if (JvmtiExport::can_examine_or_deopt_anywhere()) {
|
|
773 |
// At any safepoint, this method can get breakpointed, which would
|
|
774 |
// then require an immediate deoptimization.
|
|
775 |
full_info = true;
|
|
776 |
can_prune_locals = false; // do not prune locals
|
|
777 |
stack_slots_not_pruned = 0;
|
|
778 |
}
|
|
779 |
|
|
780 |
// do not scribble on the input jvms
|
|
781 |
JVMState* out_jvms = youngest_jvms->clone_deep(C);
|
|
782 |
call->set_jvms(out_jvms); // Start jvms list for call node
|
|
783 |
|
|
784 |
// Presize the call:
|
|
785 |
debug_only(uint non_debug_edges = call->req());
|
|
786 |
call->add_req_batch(top(), youngest_jvms->debug_depth());
|
|
787 |
assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
|
|
788 |
|
|
789 |
// Set up edges so that the call looks like this:
|
|
790 |
// Call [state:] ctl io mem fptr retadr
|
|
791 |
// [parms:] parm0 ... parmN
|
|
792 |
// [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
|
|
793 |
// [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
|
|
794 |
// [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
|
|
795 |
// Note that caller debug info precedes callee debug info.
|
|
796 |
|
|
797 |
// Fill pointer walks backwards from "young:" to "root:" in the diagram above:
|
|
798 |
uint debug_ptr = call->req();
|
|
799 |
|
|
800 |
// Loop over the map input edges associated with jvms, add them
|
|
801 |
// to the call node, & reset all offsets to match call node array.
|
|
802 |
for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
|
|
803 |
uint debug_end = debug_ptr;
|
|
804 |
uint debug_start = debug_ptr - in_jvms->debug_size();
|
|
805 |
debug_ptr = debug_start; // back up the ptr
|
|
806 |
|
|
807 |
uint p = debug_start; // walks forward in [debug_start, debug_end)
|
|
808 |
uint j, k, l;
|
|
809 |
SafePointNode* in_map = in_jvms->map();
|
|
810 |
out_jvms->set_map(call);
|
|
811 |
|
|
812 |
if (can_prune_locals) {
|
|
813 |
assert(in_jvms->method() == out_jvms->method(), "sanity");
|
|
814 |
// If the current throw can reach an exception handler in this JVMS,
|
|
815 |
// then we must keep everything live that can reach that handler.
|
|
816 |
// As a quick and dirty approximation, we look for any handlers at all.
|
|
817 |
if (in_jvms->method()->has_exception_handlers()) {
|
|
818 |
can_prune_locals = false;
|
|
819 |
}
|
|
820 |
}
|
|
821 |
|
|
822 |
// Add the Locals
|
|
823 |
k = in_jvms->locoff();
|
|
824 |
l = in_jvms->loc_size();
|
|
825 |
out_jvms->set_locoff(p);
|
|
826 |
if (full_info && !can_prune_locals) {
|
|
827 |
for (j = 0; j < l; j++)
|
|
828 |
call->set_req(p++, in_map->in(k+j));
|
|
829 |
} else {
|
|
830 |
p += l; // already set to top above by add_req_batch
|
|
831 |
}
|
|
832 |
|
|
833 |
// Add the Expression Stack
|
|
834 |
k = in_jvms->stkoff();
|
|
835 |
l = in_jvms->sp();
|
|
836 |
out_jvms->set_stkoff(p);
|
|
837 |
if (full_info && !can_prune_locals) {
|
|
838 |
for (j = 0; j < l; j++)
|
|
839 |
call->set_req(p++, in_map->in(k+j));
|
|
840 |
} else if (can_prune_locals && stack_slots_not_pruned != 0) {
|
|
841 |
// Divide stack into {S0,...,S1}, where S0 is set to top.
|
|
842 |
uint s1 = stack_slots_not_pruned;
|
|
843 |
stack_slots_not_pruned = 0; // for next iteration
|
|
844 |
if (s1 > l) s1 = l;
|
|
845 |
uint s0 = l - s1;
|
|
846 |
p += s0; // skip the tops preinstalled by add_req_batch
|
|
847 |
for (j = s0; j < l; j++)
|
|
848 |
call->set_req(p++, in_map->in(k+j));
|
|
849 |
} else {
|
|
850 |
p += l; // already set to top above by add_req_batch
|
|
851 |
}
|
|
852 |
|
|
853 |
// Add the Monitors
|
|
854 |
k = in_jvms->monoff();
|
|
855 |
l = in_jvms->mon_size();
|
|
856 |
out_jvms->set_monoff(p);
|
|
857 |
for (j = 0; j < l; j++)
|
|
858 |
call->set_req(p++, in_map->in(k+j));
|
|
859 |
|
|
860 |
// Finish the new jvms.
|
|
861 |
out_jvms->set_endoff(p);
|
|
862 |
|
|
863 |
assert(out_jvms->endoff() == debug_end, "fill ptr must match");
|
|
864 |
assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
|
|
865 |
assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
|
|
866 |
assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
|
|
867 |
assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
|
|
868 |
|
|
869 |
// Update the two tail pointers in parallel.
|
|
870 |
out_jvms = out_jvms->caller();
|
|
871 |
in_jvms = in_jvms->caller();
|
|
872 |
}
|
|
873 |
|
|
874 |
assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
|
|
875 |
|
|
876 |
// Test the correctness of JVMState::debug_xxx accessors:
|
|
877 |
assert(call->jvms()->debug_start() == non_debug_edges, "");
|
|
878 |
assert(call->jvms()->debug_end() == call->req(), "");
|
|
879 |
assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
|
|
880 |
}
|
|
881 |
|
|
882 |
bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
|
|
883 |
Bytecodes::Code code = java_bc();
|
|
884 |
if (code == Bytecodes::_wide) {
|
|
885 |
code = method()->java_code_at_bci(bci() + 1);
|
|
886 |
}
|
|
887 |
|
|
888 |
BasicType rtype = T_ILLEGAL;
|
|
889 |
int rsize = 0;
|
|
890 |
|
|
891 |
if (code != Bytecodes::_illegal) {
|
|
892 |
depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
|
|
893 |
rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
|
|
894 |
if (rtype < T_CONFLICT)
|
|
895 |
rsize = type2size[rtype];
|
|
896 |
}
|
|
897 |
|
|
898 |
switch (code) {
|
|
899 |
case Bytecodes::_illegal:
|
|
900 |
return false;
|
|
901 |
|
|
902 |
case Bytecodes::_ldc:
|
|
903 |
case Bytecodes::_ldc_w:
|
|
904 |
case Bytecodes::_ldc2_w:
|
|
905 |
inputs = 0;
|
|
906 |
break;
|
|
907 |
|
|
908 |
case Bytecodes::_dup: inputs = 1; break;
|
|
909 |
case Bytecodes::_dup_x1: inputs = 2; break;
|
|
910 |
case Bytecodes::_dup_x2: inputs = 3; break;
|
|
911 |
case Bytecodes::_dup2: inputs = 2; break;
|
|
912 |
case Bytecodes::_dup2_x1: inputs = 3; break;
|
|
913 |
case Bytecodes::_dup2_x2: inputs = 4; break;
|
|
914 |
case Bytecodes::_swap: inputs = 2; break;
|
|
915 |
case Bytecodes::_arraylength: inputs = 1; break;
|
|
916 |
|
|
917 |
case Bytecodes::_getstatic:
|
|
918 |
case Bytecodes::_putstatic:
|
|
919 |
case Bytecodes::_getfield:
|
|
920 |
case Bytecodes::_putfield:
|
|
921 |
{
|
|
922 |
bool is_get = (depth >= 0), is_static = (depth & 1);
|
|
923 |
bool ignore;
|
|
924 |
ciBytecodeStream iter(method());
|
|
925 |
iter.reset_to_bci(bci());
|
|
926 |
iter.next();
|
|
927 |
ciField* field = iter.get_field(ignore);
|
|
928 |
int size = field->type()->size();
|
|
929 |
inputs = (is_static ? 0 : 1);
|
|
930 |
if (is_get) {
|
|
931 |
depth = size - inputs;
|
|
932 |
} else {
|
|
933 |
inputs += size; // putxxx pops the value from the stack
|
|
934 |
depth = - inputs;
|
|
935 |
}
|
|
936 |
}
|
|
937 |
break;
|
|
938 |
|
|
939 |
case Bytecodes::_invokevirtual:
|
|
940 |
case Bytecodes::_invokespecial:
|
|
941 |
case Bytecodes::_invokestatic:
|
|
942 |
case Bytecodes::_invokeinterface:
|
|
943 |
{
|
|
944 |
bool is_static = (depth == 0);
|
|
945 |
bool ignore;
|
|
946 |
ciBytecodeStream iter(method());
|
|
947 |
iter.reset_to_bci(bci());
|
|
948 |
iter.next();
|
|
949 |
ciMethod* method = iter.get_method(ignore);
|
|
950 |
inputs = method->arg_size_no_receiver();
|
|
951 |
if (!is_static) inputs += 1;
|
|
952 |
int size = method->return_type()->size();
|
|
953 |
depth = size - inputs;
|
|
954 |
}
|
|
955 |
break;
|
|
956 |
|
|
957 |
case Bytecodes::_multianewarray:
|
|
958 |
{
|
|
959 |
ciBytecodeStream iter(method());
|
|
960 |
iter.reset_to_bci(bci());
|
|
961 |
iter.next();
|
|
962 |
inputs = iter.get_dimensions();
|
|
963 |
assert(rsize == 1, "");
|
|
964 |
depth = rsize - inputs;
|
|
965 |
}
|
|
966 |
break;
|
|
967 |
|
|
968 |
case Bytecodes::_ireturn:
|
|
969 |
case Bytecodes::_lreturn:
|
|
970 |
case Bytecodes::_freturn:
|
|
971 |
case Bytecodes::_dreturn:
|
|
972 |
case Bytecodes::_areturn:
|
|
973 |
assert(rsize = -depth, "");
|
|
974 |
inputs = rsize;
|
|
975 |
break;
|
|
976 |
|
|
977 |
case Bytecodes::_jsr:
|
|
978 |
case Bytecodes::_jsr_w:
|
|
979 |
inputs = 0;
|
|
980 |
depth = 1; // S.B. depth=1, not zero
|
|
981 |
break;
|
|
982 |
|
|
983 |
default:
|
|
984 |
// bytecode produces a typed result
|
|
985 |
inputs = rsize - depth;
|
|
986 |
assert(inputs >= 0, "");
|
|
987 |
break;
|
|
988 |
}
|
|
989 |
|
|
990 |
#ifdef ASSERT
|
|
991 |
// spot check
|
|
992 |
int outputs = depth + inputs;
|
|
993 |
assert(outputs >= 0, "sanity");
|
|
994 |
switch (code) {
|
|
995 |
case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
|
|
996 |
case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
|
|
997 |
case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
|
|
998 |
case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
|
|
999 |
case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
|
|
1000 |
}
|
|
1001 |
#endif //ASSERT
|
|
1002 |
|
|
1003 |
return true;
|
|
1004 |
}
|
|
1005 |
|
|
1006 |
|
|
1007 |
|
|
1008 |
//------------------------------basic_plus_adr---------------------------------
|
|
1009 |
Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
|
|
1010 |
// short-circuit a common case
|
|
1011 |
if (offset == intcon(0)) return ptr;
|
|
1012 |
return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) );
|
|
1013 |
}
|
|
1014 |
|
|
1015 |
Node* GraphKit::ConvI2L(Node* offset) {
|
|
1016 |
// short-circuit a common case
|
|
1017 |
jint offset_con = find_int_con(offset, Type::OffsetBot);
|
|
1018 |
if (offset_con != Type::OffsetBot) {
|
|
1019 |
return longcon((long) offset_con);
|
|
1020 |
}
|
|
1021 |
return _gvn.transform( new (C, 2) ConvI2LNode(offset));
|
|
1022 |
}
|
|
1023 |
Node* GraphKit::ConvL2I(Node* offset) {
|
|
1024 |
// short-circuit a common case
|
|
1025 |
jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
|
|
1026 |
if (offset_con != (jlong)Type::OffsetBot) {
|
|
1027 |
return intcon((int) offset_con);
|
|
1028 |
}
|
|
1029 |
return _gvn.transform( new (C, 2) ConvL2INode(offset));
|
|
1030 |
}
|
|
1031 |
|
|
1032 |
//-------------------------load_object_klass-----------------------------------
|
|
1033 |
Node* GraphKit::load_object_klass(Node* obj) {
|
|
1034 |
// Special-case a fresh allocation to avoid building nodes:
|
|
1035 |
Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
|
|
1036 |
if (akls != NULL) return akls;
|
|
1037 |
Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
|
|
1038 |
return _gvn.transform( new (C, 3) LoadKlassNode(0, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
|
|
1039 |
}
|
|
1040 |
|
|
1041 |
//-------------------------load_array_length-----------------------------------
|
|
1042 |
Node* GraphKit::load_array_length(Node* array) {
|
|
1043 |
// Special-case a fresh allocation to avoid building nodes:
|
|
1044 |
Node* alen = AllocateArrayNode::Ideal_length(array, &_gvn);
|
|
1045 |
if (alen != NULL) return alen;
|
|
1046 |
Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
|
|
1047 |
return _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
|
|
1048 |
}
|
|
1049 |
|
|
1050 |
//------------------------------do_null_check----------------------------------
|
|
1051 |
// Helper function to do a NULL pointer check. Returned value is
|
|
1052 |
// the incoming address with NULL casted away. You are allowed to use the
|
|
1053 |
// not-null value only if you are control dependent on the test.
|
|
1054 |
extern int explicit_null_checks_inserted,
|
|
1055 |
explicit_null_checks_elided;
|
|
1056 |
Node* GraphKit::null_check_common(Node* value, BasicType type,
|
|
1057 |
// optional arguments for variations:
|
|
1058 |
bool assert_null,
|
|
1059 |
Node* *null_control) {
|
|
1060 |
assert(!assert_null || null_control == NULL, "not both at once");
|
|
1061 |
if (stopped()) return top();
|
|
1062 |
if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
|
|
1063 |
// For some performance testing, we may wish to suppress null checking.
|
|
1064 |
value = cast_not_null(value); // Make it appear to be non-null (4962416).
|
|
1065 |
return value;
|
|
1066 |
}
|
|
1067 |
explicit_null_checks_inserted++;
|
|
1068 |
|
|
1069 |
// Construct NULL check
|
|
1070 |
Node *chk = NULL;
|
|
1071 |
switch(type) {
|
|
1072 |
case T_LONG : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
|
|
1073 |
case T_INT : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break;
|
|
1074 |
case T_ARRAY : // fall through
|
|
1075 |
type = T_OBJECT; // simplify further tests
|
|
1076 |
case T_OBJECT : {
|
|
1077 |
const Type *t = _gvn.type( value );
|
|
1078 |
|
|
1079 |
const TypeInstPtr* tp = t->isa_instptr();
|
|
1080 |
if (tp != NULL && !tp->klass()->is_loaded()
|
|
1081 |
// Only for do_null_check, not any of its siblings:
|
|
1082 |
&& !assert_null && null_control == NULL) {
|
|
1083 |
// Usually, any field access or invocation on an unloaded oop type
|
|
1084 |
// will simply fail to link, since the statically linked class is
|
|
1085 |
// likely also to be unloaded. However, in -Xcomp mode, sometimes
|
|
1086 |
// the static class is loaded but the sharper oop type is not.
|
|
1087 |
// Rather than checking for this obscure case in lots of places,
|
|
1088 |
// we simply observe that a null check on an unloaded class
|
|
1089 |
// will always be followed by a nonsense operation, so we
|
|
1090 |
// can just issue the uncommon trap here.
|
|
1091 |
// Our access to the unloaded class will only be correct
|
|
1092 |
// after it has been loaded and initialized, which requires
|
|
1093 |
// a trip through the interpreter.
|
|
1094 |
#ifndef PRODUCT
|
|
1095 |
if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
|
|
1096 |
#endif
|
|
1097 |
uncommon_trap(Deoptimization::Reason_unloaded,
|
|
1098 |
Deoptimization::Action_reinterpret,
|
|
1099 |
tp->klass(), "!loaded");
|
|
1100 |
return top();
|
|
1101 |
}
|
|
1102 |
|
|
1103 |
if (assert_null) {
|
|
1104 |
// See if the type is contained in NULL_PTR.
|
|
1105 |
// If so, then the value is already null.
|
|
1106 |
if (t->higher_equal(TypePtr::NULL_PTR)) {
|
|
1107 |
explicit_null_checks_elided++;
|
|
1108 |
return value; // Elided null assert quickly!
|
|
1109 |
}
|
|
1110 |
} else {
|
|
1111 |
// See if mixing in the NULL pointer changes type.
|
|
1112 |
// If so, then the NULL pointer was not allowed in the original
|
|
1113 |
// type. In other words, "value" was not-null.
|
|
1114 |
if (t->meet(TypePtr::NULL_PTR) != t) {
|
|
1115 |
// same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
|
|
1116 |
explicit_null_checks_elided++;
|
|
1117 |
return value; // Elided null check quickly!
|
|
1118 |
}
|
|
1119 |
}
|
|
1120 |
chk = new (C, 3) CmpPNode( value, null() );
|
|
1121 |
break;
|
|
1122 |
}
|
|
1123 |
|
|
1124 |
default : ShouldNotReachHere();
|
|
1125 |
}
|
|
1126 |
assert(chk != NULL, "sanity check");
|
|
1127 |
chk = _gvn.transform(chk);
|
|
1128 |
|
|
1129 |
BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
|
|
1130 |
BoolNode *btst = new (C, 2) BoolNode( chk, btest);
|
|
1131 |
Node *tst = _gvn.transform( btst );
|
|
1132 |
|
|
1133 |
//-----------
|
|
1134 |
// if peephole optimizations occured, a prior test existed.
|
|
1135 |
// If a prior test existed, maybe it dominates as we can avoid this test.
|
|
1136 |
if (tst != btst && type == T_OBJECT) {
|
|
1137 |
// At this point we want to scan up the CFG to see if we can
|
|
1138 |
// find an identical test (and so avoid this test altogether).
|
|
1139 |
Node *cfg = control();
|
|
1140 |
int depth = 0;
|
|
1141 |
while( depth < 16 ) { // Limit search depth for speed
|
|
1142 |
if( cfg->Opcode() == Op_IfTrue &&
|
|
1143 |
cfg->in(0)->in(1) == tst ) {
|
|
1144 |
// Found prior test. Use "cast_not_null" to construct an identical
|
|
1145 |
// CastPP (and hence hash to) as already exists for the prior test.
|
|
1146 |
// Return that casted value.
|
|
1147 |
if (assert_null) {
|
|
1148 |
replace_in_map(value, null());
|
|
1149 |
return null(); // do not issue the redundant test
|
|
1150 |
}
|
|
1151 |
Node *oldcontrol = control();
|
|
1152 |
set_control(cfg);
|
|
1153 |
Node *res = cast_not_null(value);
|
|
1154 |
set_control(oldcontrol);
|
|
1155 |
explicit_null_checks_elided++;
|
|
1156 |
return res;
|
|
1157 |
}
|
|
1158 |
cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
|
|
1159 |
if (cfg == NULL) break; // Quit at region nodes
|
|
1160 |
depth++;
|
|
1161 |
}
|
|
1162 |
}
|
|
1163 |
|
|
1164 |
//-----------
|
|
1165 |
// Branch to failure if null
|
|
1166 |
float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
|
|
1167 |
Deoptimization::DeoptReason reason;
|
|
1168 |
if (assert_null)
|
|
1169 |
reason = Deoptimization::Reason_null_assert;
|
|
1170 |
else if (type == T_OBJECT)
|
|
1171 |
reason = Deoptimization::Reason_null_check;
|
|
1172 |
else
|
|
1173 |
reason = Deoptimization::Reason_div0_check;
|
|
1174 |
|
|
1175 |
// To cause an implicit null check, we set the not-null probability
|
|
1176 |
// to the maximum (PROB_MAX). For an explicit check the probablity
|
|
1177 |
// is set to a smaller value.
|
|
1178 |
if (null_control != NULL || too_many_traps(reason)) {
|
|
1179 |
// probability is less likely
|
|
1180 |
ok_prob = PROB_LIKELY_MAG(3);
|
|
1181 |
} else if (!assert_null &&
|
|
1182 |
(ImplicitNullCheckThreshold > 0) &&
|
|
1183 |
method() != NULL &&
|
|
1184 |
(method()->method_data()->trap_count(reason)
|
|
1185 |
>= (uint)ImplicitNullCheckThreshold)) {
|
|
1186 |
ok_prob = PROB_LIKELY_MAG(3);
|
|
1187 |
}
|
|
1188 |
|
|
1189 |
if (null_control != NULL) {
|
|
1190 |
IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
|
|
1191 |
Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff));
|
|
1192 |
set_control( _gvn.transform( new (C, 1) IfTrueNode(iff)));
|
|
1193 |
if (null_true == top())
|
|
1194 |
explicit_null_checks_elided++;
|
|
1195 |
(*null_control) = null_true;
|
|
1196 |
} else {
|
|
1197 |
BuildCutout unless(this, tst, ok_prob);
|
|
1198 |
// Check for optimizer eliding test at parse time
|
|
1199 |
if (stopped()) {
|
|
1200 |
// Failure not possible; do not bother making uncommon trap.
|
|
1201 |
explicit_null_checks_elided++;
|
|
1202 |
} else if (assert_null) {
|
|
1203 |
uncommon_trap(reason,
|
|
1204 |
Deoptimization::Action_make_not_entrant,
|
|
1205 |
NULL, "assert_null");
|
|
1206 |
} else {
|
|
1207 |
builtin_throw(reason);
|
|
1208 |
}
|
|
1209 |
}
|
|
1210 |
|
|
1211 |
// Must throw exception, fall-thru not possible?
|
|
1212 |
if (stopped()) {
|
|
1213 |
return top(); // No result
|
|
1214 |
}
|
|
1215 |
|
|
1216 |
if (assert_null) {
|
|
1217 |
// Cast obj to null on this path.
|
|
1218 |
replace_in_map(value, zerocon(type));
|
|
1219 |
return zerocon(type);
|
|
1220 |
}
|
|
1221 |
|
|
1222 |
// Cast obj to not-null on this path, if there is no null_control.
|
|
1223 |
// (If there is a null_control, a non-null value may come back to haunt us.)
|
|
1224 |
if (type == T_OBJECT) {
|
|
1225 |
Node* cast = cast_not_null(value, false);
|
|
1226 |
if (null_control == NULL || (*null_control) == top())
|
|
1227 |
replace_in_map(value, cast);
|
|
1228 |
value = cast;
|
|
1229 |
}
|
|
1230 |
|
|
1231 |
return value;
|
|
1232 |
}
|
|
1233 |
|
|
1234 |
|
|
1235 |
//------------------------------cast_not_null----------------------------------
|
|
1236 |
// Cast obj to not-null on this path
|
|
1237 |
Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
|
|
1238 |
const Type *t = _gvn.type(obj);
|
|
1239 |
const Type *t_not_null = t->join(TypePtr::NOTNULL);
|
|
1240 |
// Object is already not-null?
|
|
1241 |
if( t == t_not_null ) return obj;
|
|
1242 |
|
|
1243 |
Node *cast = new (C, 2) CastPPNode(obj,t_not_null);
|
|
1244 |
cast->init_req(0, control());
|
|
1245 |
cast = _gvn.transform( cast );
|
|
1246 |
|
|
1247 |
// Scan for instances of 'obj' in the current JVM mapping.
|
|
1248 |
// These instances are known to be not-null after the test.
|
|
1249 |
if (do_replace_in_map)
|
|
1250 |
replace_in_map(obj, cast);
|
|
1251 |
|
|
1252 |
return cast; // Return casted value
|
|
1253 |
}
|
|
1254 |
|
|
1255 |
|
|
1256 |
//--------------------------replace_in_map-------------------------------------
|
|
1257 |
void GraphKit::replace_in_map(Node* old, Node* neww) {
|
|
1258 |
this->map()->replace_edge(old, neww);
|
|
1259 |
|
|
1260 |
// Note: This operation potentially replaces any edge
|
|
1261 |
// on the map. This includes locals, stack, and monitors
|
|
1262 |
// of the current (innermost) JVM state.
|
|
1263 |
|
|
1264 |
// We can consider replacing in caller maps.
|
|
1265 |
// The idea would be that an inlined function's null checks
|
|
1266 |
// can be shared with the entire inlining tree.
|
|
1267 |
// The expense of doing this is that the PreserveJVMState class
|
|
1268 |
// would have to preserve caller states too, with a deep copy.
|
|
1269 |
}
|
|
1270 |
|
|
1271 |
|
|
1272 |
|
|
1273 |
//=============================================================================
|
|
1274 |
//--------------------------------memory---------------------------------------
|
|
1275 |
Node* GraphKit::memory(uint alias_idx) {
|
|
1276 |
MergeMemNode* mem = merged_memory();
|
|
1277 |
Node* p = mem->memory_at(alias_idx);
|
|
1278 |
_gvn.set_type(p, Type::MEMORY); // must be mapped
|
|
1279 |
return p;
|
|
1280 |
}
|
|
1281 |
|
|
1282 |
//-----------------------------reset_memory------------------------------------
|
|
1283 |
Node* GraphKit::reset_memory() {
|
|
1284 |
Node* mem = map()->memory();
|
|
1285 |
// do not use this node for any more parsing!
|
|
1286 |
debug_only( map()->set_memory((Node*)NULL) );
|
|
1287 |
return _gvn.transform( mem );
|
|
1288 |
}
|
|
1289 |
|
|
1290 |
//------------------------------set_all_memory---------------------------------
|
|
1291 |
void GraphKit::set_all_memory(Node* newmem) {
|
|
1292 |
Node* mergemem = MergeMemNode::make(C, newmem);
|
|
1293 |
gvn().set_type_bottom(mergemem);
|
|
1294 |
map()->set_memory(mergemem);
|
|
1295 |
}
|
|
1296 |
|
|
1297 |
//------------------------------set_all_memory_call----------------------------
|
|
1298 |
void GraphKit::set_all_memory_call(Node* call) {
|
|
1299 |
Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) );
|
|
1300 |
set_all_memory(newmem);
|
|
1301 |
}
|
|
1302 |
|
|
1303 |
//=============================================================================
|
|
1304 |
//
|
|
1305 |
// parser factory methods for MemNodes
|
|
1306 |
//
|
|
1307 |
// These are layered on top of the factory methods in LoadNode and StoreNode,
|
|
1308 |
// and integrate with the parser's memory state and _gvn engine.
|
|
1309 |
//
|
|
1310 |
|
|
1311 |
// factory methods in "int adr_idx"
|
|
1312 |
Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
|
|
1313 |
int adr_idx,
|
|
1314 |
bool require_atomic_access) {
|
|
1315 |
assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
|
|
1316 |
const TypePtr* adr_type = NULL; // debug-mode-only argument
|
|
1317 |
debug_only(adr_type = C->get_adr_type(adr_idx));
|
|
1318 |
Node* mem = memory(adr_idx);
|
|
1319 |
Node* ld;
|
|
1320 |
if (require_atomic_access && bt == T_LONG) {
|
|
1321 |
ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t);
|
|
1322 |
} else {
|
|
1323 |
ld = LoadNode::make(C, ctl, mem, adr, adr_type, t, bt);
|
|
1324 |
}
|
|
1325 |
return _gvn.transform(ld);
|
|
1326 |
}
|
|
1327 |
|
|
1328 |
Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
|
|
1329 |
int adr_idx,
|
|
1330 |
bool require_atomic_access) {
|
|
1331 |
assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
|
|
1332 |
const TypePtr* adr_type = NULL;
|
|
1333 |
debug_only(adr_type = C->get_adr_type(adr_idx));
|
|
1334 |
Node *mem = memory(adr_idx);
|
|
1335 |
Node* st;
|
|
1336 |
if (require_atomic_access && bt == T_LONG) {
|
|
1337 |
st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val);
|
|
1338 |
} else {
|
|
1339 |
st = StoreNode::make(C, ctl, mem, adr, adr_type, val, bt);
|
|
1340 |
}
|
|
1341 |
st = _gvn.transform(st);
|
|
1342 |
set_memory(st, adr_idx);
|
|
1343 |
// Back-to-back stores can only remove intermediate store with DU info
|
|
1344 |
// so push on worklist for optimizer.
|
|
1345 |
if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
|
|
1346 |
record_for_igvn(st);
|
|
1347 |
|
|
1348 |
return st;
|
|
1349 |
}
|
|
1350 |
|
|
1351 |
void GraphKit::pre_barrier(Node* ctl,
|
|
1352 |
Node* obj,
|
|
1353 |
Node* adr,
|
|
1354 |
uint adr_idx,
|
|
1355 |
Node *val,
|
|
1356 |
const Type* val_type,
|
|
1357 |
BasicType bt) {
|
|
1358 |
BarrierSet* bs = Universe::heap()->barrier_set();
|
|
1359 |
set_control(ctl);
|
|
1360 |
switch (bs->kind()) {
|
|
1361 |
|
|
1362 |
case BarrierSet::CardTableModRef:
|
|
1363 |
case BarrierSet::CardTableExtension:
|
|
1364 |
case BarrierSet::ModRef:
|
|
1365 |
break;
|
|
1366 |
|
|
1367 |
case BarrierSet::Other:
|
|
1368 |
default :
|
|
1369 |
ShouldNotReachHere();
|
|
1370 |
|
|
1371 |
}
|
|
1372 |
}
|
|
1373 |
|
|
1374 |
void GraphKit::post_barrier(Node* ctl,
|
|
1375 |
Node* store,
|
|
1376 |
Node* obj,
|
|
1377 |
Node* adr,
|
|
1378 |
uint adr_idx,
|
|
1379 |
Node *val,
|
|
1380 |
BasicType bt,
|
|
1381 |
bool use_precise) {
|
|
1382 |
BarrierSet* bs = Universe::heap()->barrier_set();
|
|
1383 |
set_control(ctl);
|
|
1384 |
switch (bs->kind()) {
|
|
1385 |
|
|
1386 |
case BarrierSet::CardTableModRef:
|
|
1387 |
case BarrierSet::CardTableExtension:
|
|
1388 |
write_barrier_post(store, obj, adr, val, use_precise);
|
|
1389 |
break;
|
|
1390 |
|
|
1391 |
case BarrierSet::ModRef:
|
|
1392 |
break;
|
|
1393 |
|
|
1394 |
case BarrierSet::Other:
|
|
1395 |
default :
|
|
1396 |
ShouldNotReachHere();
|
|
1397 |
|
|
1398 |
}
|
|
1399 |
}
|
|
1400 |
|
|
1401 |
Node* GraphKit::store_oop_to_object(Node* ctl,
|
|
1402 |
Node* obj,
|
|
1403 |
Node* adr,
|
|
1404 |
const TypePtr* adr_type,
|
|
1405 |
Node *val,
|
|
1406 |
const Type* val_type,
|
|
1407 |
BasicType bt) {
|
|
1408 |
uint adr_idx = C->get_alias_index(adr_type);
|
|
1409 |
Node* store;
|
|
1410 |
pre_barrier(ctl, obj, adr, adr_idx, val, val_type, bt);
|
|
1411 |
store = store_to_memory(control(), adr, val, bt, adr_idx);
|
|
1412 |
post_barrier(control(), store, obj, adr, adr_idx, val, bt, false);
|
|
1413 |
return store;
|
|
1414 |
}
|
|
1415 |
|
|
1416 |
Node* GraphKit::store_oop_to_array(Node* ctl,
|
|
1417 |
Node* obj,
|
|
1418 |
Node* adr,
|
|
1419 |
const TypePtr* adr_type,
|
|
1420 |
Node *val,
|
|
1421 |
const Type* val_type,
|
|
1422 |
BasicType bt) {
|
|
1423 |
uint adr_idx = C->get_alias_index(adr_type);
|
|
1424 |
Node* store;
|
|
1425 |
pre_barrier(ctl, obj, adr, adr_idx, val, val_type, bt);
|
|
1426 |
store = store_to_memory(control(), adr, val, bt, adr_idx);
|
|
1427 |
post_barrier(control(), store, obj, adr, adr_idx, val, bt, true);
|
|
1428 |
return store;
|
|
1429 |
}
|
|
1430 |
|
|
1431 |
Node* GraphKit::store_oop_to_unknown(Node* ctl,
|
|
1432 |
Node* obj,
|
|
1433 |
Node* adr,
|
|
1434 |
const TypePtr* adr_type,
|
|
1435 |
Node *val,
|
|
1436 |
const Type* val_type,
|
|
1437 |
BasicType bt) {
|
|
1438 |
uint adr_idx = C->get_alias_index(adr_type);
|
|
1439 |
Node* store;
|
|
1440 |
pre_barrier(ctl, obj, adr, adr_idx, val, val_type, bt);
|
|
1441 |
store = store_to_memory(control(), adr, val, bt, adr_idx);
|
|
1442 |
post_barrier(control(), store, obj, adr, adr_idx, val, bt, true);
|
|
1443 |
return store;
|
|
1444 |
}
|
|
1445 |
|
|
1446 |
|
|
1447 |
//-------------------------array_element_address-------------------------
|
|
1448 |
Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
|
|
1449 |
const TypeInt* sizetype) {
|
|
1450 |
uint shift = exact_log2(type2aelembytes[elembt]);
|
|
1451 |
uint header = arrayOopDesc::base_offset_in_bytes(elembt);
|
|
1452 |
|
|
1453 |
// short-circuit a common case (saves lots of confusing waste motion)
|
|
1454 |
jint idx_con = find_int_con(idx, -1);
|
|
1455 |
if (idx_con >= 0) {
|
|
1456 |
intptr_t offset = header + ((intptr_t)idx_con << shift);
|
|
1457 |
return basic_plus_adr(ary, offset);
|
|
1458 |
}
|
|
1459 |
|
|
1460 |
// must be correct type for alignment purposes
|
|
1461 |
Node* base = basic_plus_adr(ary, header);
|
|
1462 |
#ifdef _LP64
|
|
1463 |
// The scaled index operand to AddP must be a clean 64-bit value.
|
|
1464 |
// Java allows a 32-bit int to be incremented to a negative
|
|
1465 |
// value, which appears in a 64-bit register as a large
|
|
1466 |
// positive number. Using that large positive number as an
|
|
1467 |
// operand in pointer arithmetic has bad consequences.
|
|
1468 |
// On the other hand, 32-bit overflow is rare, and the possibility
|
|
1469 |
// can often be excluded, if we annotate the ConvI2L node with
|
|
1470 |
// a type assertion that its value is known to be a small positive
|
|
1471 |
// number. (The prior range check has ensured this.)
|
|
1472 |
// This assertion is used by ConvI2LNode::Ideal.
|
|
1473 |
int index_max = max_jint - 1; // array size is max_jint, index is one less
|
|
1474 |
if (sizetype != NULL) index_max = sizetype->_hi - 1;
|
|
1475 |
const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
|
|
1476 |
idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) );
|
|
1477 |
#endif
|
|
1478 |
Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) );
|
|
1479 |
return basic_plus_adr(ary, base, scale);
|
|
1480 |
}
|
|
1481 |
|
|
1482 |
//-------------------------load_array_element-------------------------
|
|
1483 |
Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
|
|
1484 |
const Type* elemtype = arytype->elem();
|
|
1485 |
BasicType elembt = elemtype->array_element_basic_type();
|
|
1486 |
Node* adr = array_element_address(ary, idx, elembt, arytype->size());
|
|
1487 |
Node* ld = make_load(ctl, adr, elemtype, elembt, arytype);
|
|
1488 |
return ld;
|
|
1489 |
}
|
|
1490 |
|
|
1491 |
//-------------------------set_arguments_for_java_call-------------------------
|
|
1492 |
// Arguments (pre-popped from the stack) are taken from the JVMS.
|
|
1493 |
void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
|
|
1494 |
// Add the call arguments:
|
|
1495 |
uint nargs = call->method()->arg_size();
|
|
1496 |
for (uint i = 0; i < nargs; i++) {
|
|
1497 |
Node* arg = argument(i);
|
|
1498 |
call->init_req(i + TypeFunc::Parms, arg);
|
|
1499 |
}
|
|
1500 |
}
|
|
1501 |
|
|
1502 |
//---------------------------set_edges_for_java_call---------------------------
|
|
1503 |
// Connect a newly created call into the current JVMS.
|
|
1504 |
// A return value node (if any) is returned from set_edges_for_java_call.
|
|
1505 |
void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw) {
|
|
1506 |
|
|
1507 |
// Add the predefined inputs:
|
|
1508 |
call->init_req( TypeFunc::Control, control() );
|
|
1509 |
call->init_req( TypeFunc::I_O , i_o() );
|
|
1510 |
call->init_req( TypeFunc::Memory , reset_memory() );
|
|
1511 |
call->init_req( TypeFunc::FramePtr, frameptr() );
|
|
1512 |
call->init_req( TypeFunc::ReturnAdr, top() );
|
|
1513 |
|
|
1514 |
add_safepoint_edges(call, must_throw);
|
|
1515 |
|
|
1516 |
Node* xcall = _gvn.transform(call);
|
|
1517 |
|
|
1518 |
if (xcall == top()) {
|
|
1519 |
set_control(top());
|
|
1520 |
return;
|
|
1521 |
}
|
|
1522 |
assert(xcall == call, "call identity is stable");
|
|
1523 |
|
|
1524 |
// Re-use the current map to produce the result.
|
|
1525 |
|
|
1526 |
set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control)));
|
|
1527 |
set_i_o( _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O )));
|
|
1528 |
set_all_memory_call(xcall);
|
|
1529 |
|
|
1530 |
//return xcall; // no need, caller already has it
|
|
1531 |
}
|
|
1532 |
|
|
1533 |
Node* GraphKit::set_results_for_java_call(CallJavaNode* call) {
|
|
1534 |
if (stopped()) return top(); // maybe the call folded up?
|
|
1535 |
|
|
1536 |
// Capture the return value, if any.
|
|
1537 |
Node* ret;
|
|
1538 |
if (call->method() == NULL ||
|
|
1539 |
call->method()->return_type()->basic_type() == T_VOID)
|
|
1540 |
ret = top();
|
|
1541 |
else ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
|
|
1542 |
|
|
1543 |
// Note: Since any out-of-line call can produce an exception,
|
|
1544 |
// we always insert an I_O projection from the call into the result.
|
|
1545 |
|
|
1546 |
make_slow_call_ex(call, env()->Throwable_klass(), false);
|
|
1547 |
|
|
1548 |
return ret;
|
|
1549 |
}
|
|
1550 |
|
|
1551 |
//--------------------set_predefined_input_for_runtime_call--------------------
|
|
1552 |
// Reading and setting the memory state is way conservative here.
|
|
1553 |
// The real problem is that I am not doing real Type analysis on memory,
|
|
1554 |
// so I cannot distinguish card mark stores from other stores. Across a GC
|
|
1555 |
// point the Store Barrier and the card mark memory has to agree. I cannot
|
|
1556 |
// have a card mark store and its barrier split across the GC point from
|
|
1557 |
// either above or below. Here I get that to happen by reading ALL of memory.
|
|
1558 |
// A better answer would be to separate out card marks from other memory.
|
|
1559 |
// For now, return the input memory state, so that it can be reused
|
|
1560 |
// after the call, if this call has restricted memory effects.
|
|
1561 |
Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
|
|
1562 |
// Set fixed predefined input arguments
|
|
1563 |
Node* memory = reset_memory();
|
|
1564 |
call->init_req( TypeFunc::Control, control() );
|
|
1565 |
call->init_req( TypeFunc::I_O, top() ); // does no i/o
|
|
1566 |
call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
|
|
1567 |
call->init_req( TypeFunc::FramePtr, frameptr() );
|
|
1568 |
call->init_req( TypeFunc::ReturnAdr, top() );
|
|
1569 |
return memory;
|
|
1570 |
}
|
|
1571 |
|
|
1572 |
//-------------------set_predefined_output_for_runtime_call--------------------
|
|
1573 |
// Set control and memory (not i_o) from the call.
|
|
1574 |
// If keep_mem is not NULL, use it for the output state,
|
|
1575 |
// except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
|
|
1576 |
// If hook_mem is NULL, this call produces no memory effects at all.
|
|
1577 |
// If hook_mem is a Java-visible memory slice (such as arraycopy operands),
|
|
1578 |
// then only that memory slice is taken from the call.
|
|
1579 |
// In the last case, we must put an appropriate memory barrier before
|
|
1580 |
// the call, so as to create the correct anti-dependencies on loads
|
|
1581 |
// preceding the call.
|
|
1582 |
void GraphKit::set_predefined_output_for_runtime_call(Node* call,
|
|
1583 |
Node* keep_mem,
|
|
1584 |
const TypePtr* hook_mem) {
|
|
1585 |
// no i/o
|
|
1586 |
set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) ));
|
|
1587 |
if (keep_mem) {
|
|
1588 |
// First clone the existing memory state
|
|
1589 |
set_all_memory(keep_mem);
|
|
1590 |
if (hook_mem != NULL) {
|
|
1591 |
// Make memory for the call
|
|
1592 |
Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) );
|
|
1593 |
// Set the RawPtr memory state only. This covers all the heap top/GC stuff
|
|
1594 |
// We also use hook_mem to extract specific effects from arraycopy stubs.
|
|
1595 |
set_memory(mem, hook_mem);
|
|
1596 |
}
|
|
1597 |
// ...else the call has NO memory effects.
|
|
1598 |
|
|
1599 |
// Make sure the call advertises its memory effects precisely.
|
|
1600 |
// This lets us build accurate anti-dependences in gcm.cpp.
|
|
1601 |
assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
|
|
1602 |
"call node must be constructed correctly");
|
|
1603 |
} else {
|
|
1604 |
assert(hook_mem == NULL, "");
|
|
1605 |
// This is not a "slow path" call; all memory comes from the call.
|
|
1606 |
set_all_memory_call(call);
|
|
1607 |
}
|
|
1608 |
}
|
|
1609 |
|
|
1610 |
//------------------------------increment_counter------------------------------
|
|
1611 |
// for statistics: increment a VM counter by 1
|
|
1612 |
|
|
1613 |
void GraphKit::increment_counter(address counter_addr) {
|
|
1614 |
Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
|
|
1615 |
increment_counter(adr1);
|
|
1616 |
}
|
|
1617 |
|
|
1618 |
void GraphKit::increment_counter(Node* counter_addr) {
|
|
1619 |
int adr_type = Compile::AliasIdxRaw;
|
|
1620 |
Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type);
|
|
1621 |
Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
|
|
1622 |
store_to_memory( NULL, counter_addr, incr, T_INT, adr_type );
|
|
1623 |
}
|
|
1624 |
|
|
1625 |
|
|
1626 |
//------------------------------uncommon_trap----------------------------------
|
|
1627 |
// Bail out to the interpreter in mid-method. Implemented by calling the
|
|
1628 |
// uncommon_trap blob. This helper function inserts a runtime call with the
|
|
1629 |
// right debug info.
|
|
1630 |
void GraphKit::uncommon_trap(int trap_request,
|
|
1631 |
ciKlass* klass, const char* comment,
|
|
1632 |
bool must_throw,
|
|
1633 |
bool keep_exact_action) {
|
|
1634 |
if (failing()) stop();
|
|
1635 |
if (stopped()) return; // trap reachable?
|
|
1636 |
|
|
1637 |
// Note: If ProfileTraps is true, and if a deopt. actually
|
|
1638 |
// occurs here, the runtime will make sure an MDO exists. There is
|
|
1639 |
// no need to call method()->build_method_data() at this point.
|
|
1640 |
|
|
1641 |
#ifdef ASSERT
|
|
1642 |
if (!must_throw) {
|
|
1643 |
// Make sure the stack has at least enough depth to execute
|
|
1644 |
// the current bytecode.
|
|
1645 |
int inputs, ignore;
|
|
1646 |
if (compute_stack_effects(inputs, ignore)) {
|
|
1647 |
assert(sp() >= inputs, "must have enough JVMS stack to execute");
|
|
1648 |
// It is a frequent error in library_call.cpp to issue an
|
|
1649 |
// uncommon trap with the _sp value already popped.
|
|
1650 |
}
|
|
1651 |
}
|
|
1652 |
#endif
|
|
1653 |
|
|
1654 |
Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
|
|
1655 |
Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
|
|
1656 |
|
|
1657 |
switch (action) {
|
|
1658 |
case Deoptimization::Action_maybe_recompile:
|
|
1659 |
case Deoptimization::Action_reinterpret:
|
|
1660 |
// Temporary fix for 6529811 to allow virtual calls to be sure they
|
|
1661 |
// get the chance to go from mono->bi->mega
|
|
1662 |
if (!keep_exact_action &&
|
|
1663 |
Deoptimization::trap_request_index(trap_request) < 0 &&
|
|
1664 |
too_many_recompiles(reason)) {
|
|
1665 |
// This BCI is causing too many recompilations.
|
|
1666 |
action = Deoptimization::Action_none;
|
|
1667 |
trap_request = Deoptimization::make_trap_request(reason, action);
|
|
1668 |
} else {
|
|
1669 |
C->set_trap_can_recompile(true);
|
|
1670 |
}
|
|
1671 |
break;
|
|
1672 |
case Deoptimization::Action_make_not_entrant:
|
|
1673 |
C->set_trap_can_recompile(true);
|
|
1674 |
break;
|
|
1675 |
#ifdef ASSERT
|
|
1676 |
case Deoptimization::Action_none:
|
|
1677 |
case Deoptimization::Action_make_not_compilable:
|
|
1678 |
break;
|
|
1679 |
default:
|
|
1680 |
assert(false, "bad action");
|
|
1681 |
#endif
|
|
1682 |
}
|
|
1683 |
|
|
1684 |
if (TraceOptoParse) {
|
|
1685 |
char buf[100];
|
|
1686 |
tty->print_cr("Uncommon trap %s at bci:%d",
|
|
1687 |
Deoptimization::format_trap_request(buf, sizeof(buf),
|
|
1688 |
trap_request), bci());
|
|
1689 |
}
|
|
1690 |
|
|
1691 |
CompileLog* log = C->log();
|
|
1692 |
if (log != NULL) {
|
|
1693 |
int kid = (klass == NULL)? -1: log->identify(klass);
|
|
1694 |
log->begin_elem("uncommon_trap bci='%d'", bci());
|
|
1695 |
char buf[100];
|
|
1696 |
log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
|
|
1697 |
trap_request));
|
|
1698 |
if (kid >= 0) log->print(" klass='%d'", kid);
|
|
1699 |
if (comment != NULL) log->print(" comment='%s'", comment);
|
|
1700 |
log->end_elem();
|
|
1701 |
}
|
|
1702 |
|
|
1703 |
// Make sure any guarding test views this path as very unlikely
|
|
1704 |
Node *i0 = control()->in(0);
|
|
1705 |
if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
|
|
1706 |
IfNode *iff = i0->as_If();
|
|
1707 |
float f = iff->_prob; // Get prob
|
|
1708 |
if (control()->Opcode() == Op_IfTrue) {
|
|
1709 |
if (f > PROB_UNLIKELY_MAG(4))
|
|
1710 |
iff->_prob = PROB_MIN;
|
|
1711 |
} else {
|
|
1712 |
if (f < PROB_LIKELY_MAG(4))
|
|
1713 |
iff->_prob = PROB_MAX;
|
|
1714 |
}
|
|
1715 |
}
|
|
1716 |
|
|
1717 |
// Clear out dead values from the debug info.
|
|
1718 |
kill_dead_locals();
|
|
1719 |
|
|
1720 |
// Now insert the uncommon trap subroutine call
|
|
1721 |
address call_addr = SharedRuntime::uncommon_trap_blob()->instructions_begin();
|
|
1722 |
const TypePtr* no_memory_effects = NULL;
|
|
1723 |
// Pass the index of the class to be loaded
|
|
1724 |
Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
|
|
1725 |
(must_throw ? RC_MUST_THROW : 0),
|
|
1726 |
OptoRuntime::uncommon_trap_Type(),
|
|
1727 |
call_addr, "uncommon_trap", no_memory_effects,
|
|
1728 |
intcon(trap_request));
|
|
1729 |
assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
|
|
1730 |
"must extract request correctly from the graph");
|
|
1731 |
assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
|
|
1732 |
|
|
1733 |
call->set_req(TypeFunc::ReturnAdr, returnadr());
|
|
1734 |
// The debug info is the only real input to this call.
|
|
1735 |
|
|
1736 |
// Halt-and-catch fire here. The above call should never return!
|
|
1737 |
HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr());
|
|
1738 |
_gvn.set_type_bottom(halt);
|
|
1739 |
root()->add_req(halt);
|
|
1740 |
|
|
1741 |
stop_and_kill_map();
|
|
1742 |
}
|
|
1743 |
|
|
1744 |
|
|
1745 |
//--------------------------just_allocated_object------------------------------
|
|
1746 |
// Report the object that was just allocated.
|
|
1747 |
// It must be the case that there are no intervening safepoints.
|
|
1748 |
// We use this to determine if an object is so "fresh" that
|
|
1749 |
// it does not require card marks.
|
|
1750 |
Node* GraphKit::just_allocated_object(Node* current_control) {
|
|
1751 |
if (C->recent_alloc_ctl() == current_control)
|
|
1752 |
return C->recent_alloc_obj();
|
|
1753 |
return NULL;
|
|
1754 |
}
|
|
1755 |
|
|
1756 |
|
|
1757 |
//------------------------------store_barrier----------------------------------
|
|
1758 |
// Insert a write-barrier store. This is to let generational GC work; we have
|
|
1759 |
// to flag all oop-stores before the next GC point.
|
|
1760 |
void GraphKit::write_barrier_post(Node* oop_store, Node* obj, Node* adr,
|
|
1761 |
Node* val, bool use_precise) {
|
|
1762 |
// No store check needed if we're storing a NULL or an old object
|
|
1763 |
// (latter case is probably a string constant). The concurrent
|
|
1764 |
// mark sweep garbage collector, however, needs to have all nonNull
|
|
1765 |
// oop updates flagged via card-marks.
|
|
1766 |
if (val != NULL && val->is_Con()) {
|
|
1767 |
// must be either an oop or NULL
|
|
1768 |
const Type* t = val->bottom_type();
|
|
1769 |
if (t == TypePtr::NULL_PTR || t == Type::TOP)
|
|
1770 |
// stores of null never (?) need barriers
|
|
1771 |
return;
|
|
1772 |
ciObject* con = t->is_oopptr()->const_oop();
|
|
1773 |
if (con != NULL
|
|
1774 |
&& con->is_perm()
|
|
1775 |
&& Universe::heap()->can_elide_permanent_oop_store_barriers())
|
|
1776 |
// no store barrier needed, because no old-to-new ref created
|
|
1777 |
return;
|
|
1778 |
}
|
|
1779 |
|
|
1780 |
if (use_ReduceInitialCardMarks()
|
|
1781 |
&& obj == just_allocated_object(control())) {
|
|
1782 |
// We can skip marks on a freshly-allocated object.
|
|
1783 |
// Keep this code in sync with do_eager_card_mark in runtime.cpp.
|
|
1784 |
// That routine eagerly marks the occasional object which is produced
|
|
1785 |
// by the slow path, so that we don't have to do it here.
|
|
1786 |
return;
|
|
1787 |
}
|
|
1788 |
|
|
1789 |
if (!use_precise) {
|
|
1790 |
// All card marks for a (non-array) instance are in one place:
|
|
1791 |
adr = obj;
|
|
1792 |
}
|
|
1793 |
// (Else it's an array (or unknown), and we want more precise card marks.)
|
|
1794 |
assert(adr != NULL, "");
|
|
1795 |
|
|
1796 |
// Get the alias_index for raw card-mark memory
|
|
1797 |
int adr_type = Compile::AliasIdxRaw;
|
|
1798 |
// Convert the pointer to an int prior to doing math on it
|
|
1799 |
Node* cast = _gvn.transform(new (C, 2) CastP2XNode(control(), adr));
|
|
1800 |
// Divide by card size
|
|
1801 |
assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
|
|
1802 |
"Only one we handle so far.");
|
|
1803 |
CardTableModRefBS* ct =
|
|
1804 |
(CardTableModRefBS*)(Universe::heap()->barrier_set());
|
|
1805 |
Node *b = _gvn.transform(new (C, 3) URShiftXNode( cast, _gvn.intcon(CardTableModRefBS::card_shift) ));
|
|
1806 |
// We store into a byte array, so do not bother to left-shift by zero
|
|
1807 |
// Get base of card map
|
|
1808 |
assert(sizeof(*ct->byte_map_base) == sizeof(jbyte),
|
|
1809 |
"adjust this code");
|
|
1810 |
Node *c = makecon(TypeRawPtr::make((address)ct->byte_map_base));
|
|
1811 |
// Combine
|
|
1812 |
Node *sb_ctl = control();
|
|
1813 |
Node *sb_adr = _gvn.transform(new (C, 4) AddPNode( top()/*no base ptr*/, c, b ));
|
|
1814 |
Node *sb_val = _gvn.intcon(0);
|
|
1815 |
// Smash zero into card
|
|
1816 |
if( !UseConcMarkSweepGC ) {
|
|
1817 |
BasicType bt = T_BYTE;
|
|
1818 |
store_to_memory(sb_ctl, sb_adr, sb_val, bt, adr_type);
|
|
1819 |
} else {
|
|
1820 |
// Specialized path for CM store barrier
|
|
1821 |
cms_card_mark( sb_ctl, sb_adr, sb_val, oop_store);
|
|
1822 |
}
|
|
1823 |
}
|
|
1824 |
|
|
1825 |
// Specialized path for CMS store barrier
|
|
1826 |
void GraphKit::cms_card_mark(Node* ctl, Node* adr, Node* val, Node *oop_store) {
|
|
1827 |
BasicType bt = T_BYTE;
|
|
1828 |
int adr_idx = Compile::AliasIdxRaw;
|
|
1829 |
Node* mem = memory(adr_idx);
|
|
1830 |
|
|
1831 |
// The type input is NULL in PRODUCT builds
|
|
1832 |
const TypePtr* type = NULL;
|
|
1833 |
debug_only(type = C->get_adr_type(adr_idx));
|
|
1834 |
|
|
1835 |
// Add required edge to oop_store, optimizer does not support precedence edges.
|
|
1836 |
// Convert required edge to precedence edge before allocation.
|
|
1837 |
Node *store = _gvn.transform( new (C, 5) StoreCMNode(ctl, mem, adr, type, val, oop_store) );
|
|
1838 |
set_memory(store, adr_idx);
|
|
1839 |
|
|
1840 |
// For CMS, back-to-back card-marks can only remove the first one
|
|
1841 |
// and this requires DU info. Push on worklist for optimizer.
|
|
1842 |
if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
|
|
1843 |
record_for_igvn(store);
|
|
1844 |
}
|
|
1845 |
|
|
1846 |
|
|
1847 |
void GraphKit::round_double_arguments(ciMethod* dest_method) {
|
|
1848 |
// (Note: TypeFunc::make has a cache that makes this fast.)
|
|
1849 |
const TypeFunc* tf = TypeFunc::make(dest_method);
|
|
1850 |
int nargs = tf->_domain->_cnt - TypeFunc::Parms;
|
|
1851 |
for (int j = 0; j < nargs; j++) {
|
|
1852 |
const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
|
|
1853 |
if( targ->basic_type() == T_DOUBLE ) {
|
|
1854 |
// If any parameters are doubles, they must be rounded before
|
|
1855 |
// the call, dstore_rounding does gvn.transform
|
|
1856 |
Node *arg = argument(j);
|
|
1857 |
arg = dstore_rounding(arg);
|
|
1858 |
set_argument(j, arg);
|
|
1859 |
}
|
|
1860 |
}
|
|
1861 |
}
|
|
1862 |
|
|
1863 |
void GraphKit::round_double_result(ciMethod* dest_method) {
|
|
1864 |
// A non-strict method may return a double value which has an extended
|
|
1865 |
// exponent, but this must not be visible in a caller which is 'strict'
|
|
1866 |
// If a strict caller invokes a non-strict callee, round a double result
|
|
1867 |
|
|
1868 |
BasicType result_type = dest_method->return_type()->basic_type();
|
|
1869 |
assert( method() != NULL, "must have caller context");
|
|
1870 |
if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
|
|
1871 |
// Destination method's return value is on top of stack
|
|
1872 |
// dstore_rounding() does gvn.transform
|
|
1873 |
Node *result = pop_pair();
|
|
1874 |
result = dstore_rounding(result);
|
|
1875 |
push_pair(result);
|
|
1876 |
}
|
|
1877 |
}
|
|
1878 |
|
|
1879 |
// rounding for strict float precision conformance
|
|
1880 |
Node* GraphKit::precision_rounding(Node* n) {
|
|
1881 |
return UseStrictFP && _method->flags().is_strict()
|
|
1882 |
&& UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
|
|
1883 |
? _gvn.transform( new (C, 2) RoundFloatNode(0, n) )
|
|
1884 |
: n;
|
|
1885 |
}
|
|
1886 |
|
|
1887 |
// rounding for strict double precision conformance
|
|
1888 |
Node* GraphKit::dprecision_rounding(Node *n) {
|
|
1889 |
return UseStrictFP && _method->flags().is_strict()
|
|
1890 |
&& UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
|
|
1891 |
? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
|
|
1892 |
: n;
|
|
1893 |
}
|
|
1894 |
|
|
1895 |
// rounding for non-strict double stores
|
|
1896 |
Node* GraphKit::dstore_rounding(Node* n) {
|
|
1897 |
return Matcher::strict_fp_requires_explicit_rounding
|
|
1898 |
&& UseSSE <= 1
|
|
1899 |
? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
|
|
1900 |
: n;
|
|
1901 |
}
|
|
1902 |
|
|
1903 |
//=============================================================================
|
|
1904 |
// Generate a fast path/slow path idiom. Graph looks like:
|
|
1905 |
// [foo] indicates that 'foo' is a parameter
|
|
1906 |
//
|
|
1907 |
// [in] NULL
|
|
1908 |
// \ /
|
|
1909 |
// CmpP
|
|
1910 |
// Bool ne
|
|
1911 |
// If
|
|
1912 |
// / \
|
|
1913 |
// True False-<2>
|
|
1914 |
// / |
|
|
1915 |
// / cast_not_null
|
|
1916 |
// Load | | ^
|
|
1917 |
// [fast_test] | |
|
|
1918 |
// gvn to opt_test | |
|
|
1919 |
// / \ | <1>
|
|
1920 |
// True False |
|
|
1921 |
// | \\ |
|
|
1922 |
// [slow_call] \[fast_result]
|
|
1923 |
// Ctl Val \ \
|
|
1924 |
// | \ \
|
|
1925 |
// Catch <1> \ \
|
|
1926 |
// / \ ^ \ \
|
|
1927 |
// Ex No_Ex | \ \
|
|
1928 |
// | \ \ | \ <2> \
|
|
1929 |
// ... \ [slow_res] | | \ [null_result]
|
|
1930 |
// \ \--+--+--- | |
|
|
1931 |
// \ | / \ | /
|
|
1932 |
// --------Region Phi
|
|
1933 |
//
|
|
1934 |
//=============================================================================
|
|
1935 |
// Code is structured as a series of driver functions all called 'do_XXX' that
|
|
1936 |
// call a set of helper functions. Helper functions first, then drivers.
|
|
1937 |
|
|
1938 |
//------------------------------null_check_oop---------------------------------
|
|
1939 |
// Null check oop. Set null-path control into Region in slot 3.
|
|
1940 |
// Make a cast-not-nullness use the other not-null control. Return cast.
|
|
1941 |
Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
|
|
1942 |
bool never_see_null) {
|
|
1943 |
// Initial NULL check taken path
|
|
1944 |
(*null_control) = top();
|
|
1945 |
Node* cast = null_check_common(value, T_OBJECT, false, null_control);
|
|
1946 |
|
|
1947 |
// Generate uncommon_trap:
|
|
1948 |
if (never_see_null && (*null_control) != top()) {
|
|
1949 |
// If we see an unexpected null at a check-cast we record it and force a
|
|
1950 |
// recompile; the offending check-cast will be compiled to handle NULLs.
|
|
1951 |
// If we see more than one offending BCI, then all checkcasts in the
|
|
1952 |
// method will be compiled to handle NULLs.
|
|
1953 |
PreserveJVMState pjvms(this);
|
|
1954 |
set_control(*null_control);
|
|
1955 |
uncommon_trap(Deoptimization::Reason_null_check,
|
|
1956 |
Deoptimization::Action_make_not_entrant);
|
|
1957 |
(*null_control) = top(); // NULL path is dead
|
|
1958 |
}
|
|
1959 |
|
|
1960 |
// Cast away null-ness on the result
|
|
1961 |
return cast;
|
|
1962 |
}
|
|
1963 |
|
|
1964 |
//------------------------------opt_iff----------------------------------------
|
|
1965 |
// Optimize the fast-check IfNode. Set the fast-path region slot 2.
|
|
1966 |
// Return slow-path control.
|
|
1967 |
Node* GraphKit::opt_iff(Node* region, Node* iff) {
|
|
1968 |
IfNode *opt_iff = _gvn.transform(iff)->as_If();
|
|
1969 |
|
|
1970 |
// Fast path taken; set region slot 2
|
|
1971 |
Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) );
|
|
1972 |
region->init_req(2,fast_taken); // Capture fast-control
|
|
1973 |
|
|
1974 |
// Fast path not-taken, i.e. slow path
|
|
1975 |
Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) );
|
|
1976 |
return slow_taken;
|
|
1977 |
}
|
|
1978 |
|
|
1979 |
//-----------------------------make_runtime_call-------------------------------
|
|
1980 |
Node* GraphKit::make_runtime_call(int flags,
|
|
1981 |
const TypeFunc* call_type, address call_addr,
|
|
1982 |
const char* call_name,
|
|
1983 |
const TypePtr* adr_type,
|
|
1984 |
// The following parms are all optional.
|
|
1985 |
// The first NULL ends the list.
|
|
1986 |
Node* parm0, Node* parm1,
|
|
1987 |
Node* parm2, Node* parm3,
|
|
1988 |
Node* parm4, Node* parm5,
|
|
1989 |
Node* parm6, Node* parm7) {
|
|
1990 |
// Slow-path call
|
|
1991 |
int size = call_type->domain()->cnt();
|
|
1992 |
bool is_leaf = !(flags & RC_NO_LEAF);
|
|
1993 |
bool has_io = (!is_leaf && !(flags & RC_NO_IO));
|
|
1994 |
if (call_name == NULL) {
|
|
1995 |
assert(!is_leaf, "must supply name for leaf");
|
|
1996 |
call_name = OptoRuntime::stub_name(call_addr);
|
|
1997 |
}
|
|
1998 |
CallNode* call;
|
|
1999 |
if (!is_leaf) {
|
|
2000 |
call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name,
|
|
2001 |
bci(), adr_type);
|
|
2002 |
} else if (flags & RC_NO_FP) {
|
|
2003 |
call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
|
|
2004 |
} else {
|
|
2005 |
call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type);
|
|
2006 |
}
|
|
2007 |
|
|
2008 |
// The following is similar to set_edges_for_java_call,
|
|
2009 |
// except that the memory effects of the call are restricted to AliasIdxRaw.
|
|
2010 |
|
|
2011 |
// Slow path call has no side-effects, uses few values
|
|
2012 |
bool wide_in = !(flags & RC_NARROW_MEM);
|
|
2013 |
bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
|
|
2014 |
|
|
2015 |
Node* prev_mem = NULL;
|
|
2016 |
if (wide_in) {
|
|
2017 |
prev_mem = set_predefined_input_for_runtime_call(call);
|
|
2018 |
} else {
|
|
2019 |
assert(!wide_out, "narrow in => narrow out");
|
|
2020 |
Node* narrow_mem = memory(adr_type);
|
|
2021 |
prev_mem = reset_memory();
|
|
2022 |
map()->set_memory(narrow_mem);
|
|
2023 |
set_predefined_input_for_runtime_call(call);
|
|
2024 |
}
|
|
2025 |
|
|
2026 |
// Hook each parm in order. Stop looking at the first NULL.
|
|
2027 |
if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
|
|
2028 |
if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
|
|
2029 |
if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
|
|
2030 |
if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
|
|
2031 |
if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
|
|
2032 |
if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
|
|
2033 |
if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
|
|
2034 |
if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
|
|
2035 |
/* close each nested if ===> */ } } } } } } } }
|
|
2036 |
assert(call->in(call->req()-1) != NULL, "must initialize all parms");
|
|
2037 |
|
|
2038 |
if (!is_leaf) {
|
|
2039 |
// Non-leaves can block and take safepoints:
|
|
2040 |
add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
|
|
2041 |
}
|
|
2042 |
// Non-leaves can throw exceptions:
|
|
2043 |
if (has_io) {
|
|
2044 |
call->set_req(TypeFunc::I_O, i_o());
|
|
2045 |
}
|
|
2046 |
|
|
2047 |
if (flags & RC_UNCOMMON) {
|
|
2048 |
// Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
|
|
2049 |
// (An "if" probability corresponds roughly to an unconditional count.
|
|
2050 |
// Sort of.)
|
|
2051 |
call->set_cnt(PROB_UNLIKELY_MAG(4));
|
|
2052 |
}
|
|
2053 |
|
|
2054 |
Node* c = _gvn.transform(call);
|
|
2055 |
assert(c == call, "cannot disappear");
|
|
2056 |
|
|
2057 |
if (wide_out) {
|
|
2058 |
// Slow path call has full side-effects.
|
|
2059 |
set_predefined_output_for_runtime_call(call);
|
|
2060 |
} else {
|
|
2061 |
// Slow path call has few side-effects, and/or sets few values.
|
|
2062 |
set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
|
|
2063 |
}
|
|
2064 |
|
|
2065 |
if (has_io) {
|
|
2066 |
set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O)));
|
|
2067 |
}
|
|
2068 |
return call;
|
|
2069 |
|
|
2070 |
}
|
|
2071 |
|
|
2072 |
//------------------------------merge_memory-----------------------------------
|
|
2073 |
// Merge memory from one path into the current memory state.
|
|
2074 |
void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
|
|
2075 |
for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
|
|
2076 |
Node* old_slice = mms.force_memory();
|
|
2077 |
Node* new_slice = mms.memory2();
|
|
2078 |
if (old_slice != new_slice) {
|
|
2079 |
PhiNode* phi;
|
|
2080 |
if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
|
|
2081 |
phi = new_slice->as_Phi();
|
|
2082 |
#ifdef ASSERT
|
|
2083 |
if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
|
|
2084 |
old_slice = old_slice->in(new_path);
|
|
2085 |
// Caller is responsible for ensuring that any pre-existing
|
|
2086 |
// phis are already aware of old memory.
|
|
2087 |
int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path
|
|
2088 |
assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
|
|
2089 |
#endif
|
|
2090 |
mms.set_memory(phi);
|
|
2091 |
} else {
|
|
2092 |
phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
|
|
2093 |
_gvn.set_type(phi, Type::MEMORY);
|
|
2094 |
phi->set_req(new_path, new_slice);
|
|
2095 |
mms.set_memory(_gvn.transform(phi)); // assume it is complete
|
|
2096 |
}
|
|
2097 |
}
|
|
2098 |
}
|
|
2099 |
}
|
|
2100 |
|
|
2101 |
//------------------------------make_slow_call_ex------------------------------
|
|
2102 |
// Make the exception handler hookups for the slow call
|
|
2103 |
void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
|
|
2104 |
if (stopped()) return;
|
|
2105 |
|
|
2106 |
// Make a catch node with just two handlers: fall-through and catch-all
|
|
2107 |
Node* i_o = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
|
|
2108 |
Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) );
|
|
2109 |
Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
|
|
2110 |
Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
|
|
2111 |
|
|
2112 |
{ PreserveJVMState pjvms(this);
|
|
2113 |
set_control(excp);
|
|
2114 |
set_i_o(i_o);
|
|
2115 |
|
|
2116 |
if (excp != top()) {
|
|
2117 |
// Create an exception state also.
|
|
2118 |
// Use an exact type if the caller has specified a specific exception.
|
|
2119 |
const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
|
|
2120 |
Node* ex_oop = new (C, 2) CreateExNode(ex_type, control(), i_o);
|
|
2121 |
add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
|
|
2122 |
}
|
|
2123 |
}
|
|
2124 |
|
|
2125 |
// Get the no-exception control from the CatchNode.
|
|
2126 |
set_control(norm);
|
|
2127 |
}
|
|
2128 |
|
|
2129 |
|
|
2130 |
//-------------------------------gen_subtype_check-----------------------------
|
|
2131 |
// Generate a subtyping check. Takes as input the subtype and supertype.
|
|
2132 |
// Returns 2 values: sets the default control() to the true path and returns
|
|
2133 |
// the false path. Only reads invariant memory; sets no (visible) memory.
|
|
2134 |
// The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
|
|
2135 |
// but that's not exposed to the optimizer. This call also doesn't take in an
|
|
2136 |
// Object; if you wish to check an Object you need to load the Object's class
|
|
2137 |
// prior to coming here.
|
|
2138 |
Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
|
|
2139 |
// Fast check for identical types, perhaps identical constants.
|
|
2140 |
// The types can even be identical non-constants, in cases
|
|
2141 |
// involving Array.newInstance, Object.clone, etc.
|
|
2142 |
if (subklass == superklass)
|
|
2143 |
return top(); // false path is dead; no test needed.
|
|
2144 |
|
|
2145 |
if (_gvn.type(superklass)->singleton()) {
|
|
2146 |
ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
|
|
2147 |
ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
|
|
2148 |
|
|
2149 |
// In the common case of an exact superklass, try to fold up the
|
|
2150 |
// test before generating code. You may ask, why not just generate
|
|
2151 |
// the code and then let it fold up? The answer is that the generated
|
|
2152 |
// code will necessarily include null checks, which do not always
|
|
2153 |
// completely fold away. If they are also needless, then they turn
|
|
2154 |
// into a performance loss. Example:
|
|
2155 |
// Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
|
|
2156 |
// Here, the type of 'fa' is often exact, so the store check
|
|
2157 |
// of fa[1]=x will fold up, without testing the nullness of x.
|
|
2158 |
switch (static_subtype_check(superk, subk)) {
|
|
2159 |
case SSC_always_false:
|
|
2160 |
{
|
|
2161 |
Node* always_fail = control();
|
|
2162 |
set_control(top());
|
|
2163 |
return always_fail;
|
|
2164 |
}
|
|
2165 |
case SSC_always_true:
|
|
2166 |
return top();
|
|
2167 |
case SSC_easy_test:
|
|
2168 |
{
|
|
2169 |
// Just do a direct pointer compare and be done.
|
|
2170 |
Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) );
|
|
2171 |
Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
|
|
2172 |
IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
|
|
2173 |
set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) );
|
|
2174 |
return _gvn.transform( new(C, 1) IfFalseNode(iff) );
|
|
2175 |
}
|
|
2176 |
case SSC_full_test:
|
|
2177 |
break;
|
|
2178 |
default:
|
|
2179 |
ShouldNotReachHere();
|
|
2180 |
}
|
|
2181 |
}
|
|
2182 |
|
|
2183 |
// %%% Possible further optimization: Even if the superklass is not exact,
|
|
2184 |
// if the subklass is the unique subtype of the superklass, the check
|
|
2185 |
// will always succeed. We could leave a dependency behind to ensure this.
|
|
2186 |
|
|
2187 |
// First load the super-klass's check-offset
|
|
2188 |
Node *p1 = basic_plus_adr( superklass, superklass, sizeof(oopDesc) + Klass::super_check_offset_offset_in_bytes() );
|
|
2189 |
Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) );
|
|
2190 |
int cacheoff_con = sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes();
|
|
2191 |
bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
|
|
2192 |
|
|
2193 |
// Load from the sub-klass's super-class display list, or a 1-word cache of
|
|
2194 |
// the secondary superclass list, or a failing value with a sentinel offset
|
|
2195 |
// if the super-klass is an interface or exceptionally deep in the Java
|
|
2196 |
// hierarchy and we have to scan the secondary superclass list the hard way.
|
|
2197 |
// Worst-case type is a little odd: NULL is allowed as a result (usually
|
|
2198 |
// klass loads can never produce a NULL).
|
|
2199 |
Node *chk_off_X = ConvI2X(chk_off);
|
|
2200 |
Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) );
|
|
2201 |
// For some types like interfaces the following loadKlass is from a 1-word
|
|
2202 |
// cache which is mutable so can't use immutable memory. Other
|
|
2203 |
// types load from the super-class display table which is immutable.
|
|
2204 |
Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
|
|
2205 |
Node *nkls = _gvn.transform( new (C, 3) LoadKlassNode( NULL, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
|
|
2206 |
|
|
2207 |
// Compile speed common case: ARE a subtype and we canNOT fail
|
|
2208 |
if( superklass == nkls )
|
|
2209 |
return top(); // false path is dead; no test needed.
|
|
2210 |
|
|
2211 |
// See if we get an immediate positive hit. Happens roughly 83% of the
|
|
2212 |
// time. Test to see if the value loaded just previously from the subklass
|
|
2213 |
// is exactly the superklass.
|
|
2214 |
Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) );
|
|
2215 |
Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) );
|
|
2216 |
IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
|
|
2217 |
Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) );
|
|
2218 |
set_control( _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) );
|
|
2219 |
|
|
2220 |
// Compile speed common case: Check for being deterministic right now. If
|
|
2221 |
// chk_off is a constant and not equal to cacheoff then we are NOT a
|
|
2222 |
// subklass. In this case we need exactly the 1 test above and we can
|
|
2223 |
// return those results immediately.
|
|
2224 |
if (!might_be_cache) {
|
|
2225 |
Node* not_subtype_ctrl = control();
|
|
2226 |
set_control(iftrue1); // We need exactly the 1 test above
|
|
2227 |
return not_subtype_ctrl;
|
|
2228 |
}
|
|
2229 |
|
|
2230 |
// Gather the various success & failures here
|
|
2231 |
RegionNode *r_ok_subtype = new (C, 4) RegionNode(4);
|
|
2232 |
record_for_igvn(r_ok_subtype);
|
|
2233 |
RegionNode *r_not_subtype = new (C, 3) RegionNode(3);
|
|
2234 |
record_for_igvn(r_not_subtype);
|
|
2235 |
|
|
2236 |
r_ok_subtype->init_req(1, iftrue1);
|
|
2237 |
|
|
2238 |
// Check for immediate negative hit. Happens roughly 11% of the time (which
|
|
2239 |
// is roughly 63% of the remaining cases). Test to see if the loaded
|
|
2240 |
// check-offset points into the subklass display list or the 1-element
|
|
2241 |
// cache. If it points to the display (and NOT the cache) and the display
|
|
2242 |
// missed then it's not a subtype.
|
|
2243 |
Node *cacheoff = _gvn.intcon(cacheoff_con);
|
|
2244 |
Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) );
|
|
2245 |
Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) );
|
|
2246 |
IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
|
|
2247 |
r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) );
|
|
2248 |
set_control( _gvn.transform( new (C, 1) IfFalseNode(iff2) ) );
|
|
2249 |
|
|
2250 |
// Check for self. Very rare to get here, but its taken 1/3 the time.
|
|
2251 |
// No performance impact (too rare) but allows sharing of secondary arrays
|
|
2252 |
// which has some footprint reduction.
|
|
2253 |
Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) );
|
|
2254 |
Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) );
|
|
2255 |
IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
|
|
2256 |
r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) );
|
|
2257 |
set_control( _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) );
|
|
2258 |
|
|
2259 |
// Now do a linear scan of the secondary super-klass array. Again, no real
|
|
2260 |
// performance impact (too rare) but it's gotta be done.
|
|
2261 |
// (The stub also contains the self-check of subklass == superklass.
|
|
2262 |
// Since the code is rarely used, there is no penalty for moving it
|
|
2263 |
// out of line, and it can only improve I-cache density.)
|
|
2264 |
Node* psc = _gvn.transform(
|
|
2265 |
new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) );
|
|
2266 |
|
|
2267 |
Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) );
|
|
2268 |
Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) );
|
|
2269 |
IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
|
|
2270 |
r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) );
|
|
2271 |
r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) );
|
|
2272 |
|
|
2273 |
// Return false path; set default control to true path.
|
|
2274 |
set_control( _gvn.transform(r_ok_subtype) );
|
|
2275 |
return _gvn.transform(r_not_subtype);
|
|
2276 |
}
|
|
2277 |
|
|
2278 |
//----------------------------static_subtype_check-----------------------------
|
|
2279 |
// Shortcut important common cases when superklass is exact:
|
|
2280 |
// (0) superklass is java.lang.Object (can occur in reflective code)
|
|
2281 |
// (1) subklass is already limited to a subtype of superklass => always ok
|
|
2282 |
// (2) subklass does not overlap with superklass => always fail
|
|
2283 |
// (3) superklass has NO subtypes and we can check with a simple compare.
|
|
2284 |
int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
|
|
2285 |
if (StressReflectiveCode) {
|
|
2286 |
return SSC_full_test; // Let caller generate the general case.
|
|
2287 |
}
|
|
2288 |
|
|
2289 |
if (superk == env()->Object_klass()) {
|
|
2290 |
return SSC_always_true; // (0) this test cannot fail
|
|
2291 |
}
|
|
2292 |
|
|
2293 |
ciType* superelem = superk;
|
|
2294 |
if (superelem->is_array_klass())
|
|
2295 |
superelem = superelem->as_array_klass()->base_element_type();
|
|
2296 |
|
|
2297 |
if (!subk->is_interface()) { // cannot trust static interface types yet
|
|
2298 |
if (subk->is_subtype_of(superk)) {
|
|
2299 |
return SSC_always_true; // (1) false path dead; no dynamic test needed
|
|
2300 |
}
|
|
2301 |
if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
|
|
2302 |
!superk->is_subtype_of(subk)) {
|
|
2303 |
return SSC_always_false;
|
|
2304 |
}
|
|
2305 |
}
|
|
2306 |
|
|
2307 |
// If casting to an instance klass, it must have no subtypes
|
|
2308 |
if (superk->is_interface()) {
|
|
2309 |
// Cannot trust interfaces yet.
|
|
2310 |
// %%% S.B. superk->nof_implementors() == 1
|
|
2311 |
} else if (superelem->is_instance_klass()) {
|
|
2312 |
ciInstanceKlass* ik = superelem->as_instance_klass();
|
|
2313 |
if (!ik->has_subklass() && !ik->is_interface()) {
|
|
2314 |
if (!ik->is_final()) {
|
|
2315 |
// Add a dependency if there is a chance of a later subclass.
|
|
2316 |
C->dependencies()->assert_leaf_type(ik);
|
|
2317 |
}
|
|
2318 |
return SSC_easy_test; // (3) caller can do a simple ptr comparison
|
|
2319 |
}
|
|
2320 |
} else {
|
|
2321 |
// A primitive array type has no subtypes.
|
|
2322 |
return SSC_easy_test; // (3) caller can do a simple ptr comparison
|
|
2323 |
}
|
|
2324 |
|
|
2325 |
return SSC_full_test;
|
|
2326 |
}
|
|
2327 |
|
|
2328 |
// Profile-driven exact type check:
|
|
2329 |
Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
|
|
2330 |
float prob,
|
|
2331 |
Node* *casted_receiver) {
|
|
2332 |
const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
|
|
2333 |
Node* recv_klass = load_object_klass(receiver);
|
|
2334 |
Node* want_klass = makecon(tklass);
|
|
2335 |
Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) );
|
|
2336 |
Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
|
|
2337 |
IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
|
|
2338 |
set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ));
|
|
2339 |
Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) );
|
|
2340 |
|
|
2341 |
const TypeOopPtr* recv_xtype = tklass->as_instance_type();
|
|
2342 |
assert(recv_xtype->klass_is_exact(), "");
|
|
2343 |
|
|
2344 |
// Subsume downstream occurrences of receiver with a cast to
|
|
2345 |
// recv_xtype, since now we know what the type will be.
|
|
2346 |
Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype);
|
|
2347 |
(*casted_receiver) = _gvn.transform(cast);
|
|
2348 |
// (User must make the replace_in_map call.)
|
|
2349 |
|
|
2350 |
return fail;
|
|
2351 |
}
|
|
2352 |
|
|
2353 |
|
|
2354 |
//-------------------------------gen_instanceof--------------------------------
|
|
2355 |
// Generate an instance-of idiom. Used by both the instance-of bytecode
|
|
2356 |
// and the reflective instance-of call.
|
|
2357 |
Node* GraphKit::gen_instanceof( Node *subobj, Node* superklass ) {
|
|
2358 |
C->set_has_split_ifs(true); // Has chance for split-if optimization
|
|
2359 |
assert( !stopped(), "dead parse path should be checked in callers" );
|
|
2360 |
assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
|
|
2361 |
"must check for not-null not-dead klass in callers");
|
|
2362 |
|
|
2363 |
// Make the merge point
|
|
2364 |
enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
|
|
2365 |
RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
|
|
2366 |
Node* phi = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
|
|
2367 |
C->set_has_split_ifs(true); // Has chance for split-if optimization
|
|
2368 |
|
|
2369 |
// Null check; get casted pointer; set region slot 3
|
|
2370 |
Node* null_ctl = top();
|
|
2371 |
Node* not_null_obj = null_check_oop(subobj, &null_ctl);
|
|
2372 |
|
|
2373 |
// If not_null_obj is dead, only null-path is taken
|
|
2374 |
if (stopped()) { // Doing instance-of on a NULL?
|
|
2375 |
set_control(null_ctl);
|
|
2376 |
return intcon(0);
|
|
2377 |
}
|
|
2378 |
region->init_req(_null_path, null_ctl);
|
|
2379 |
phi ->init_req(_null_path, intcon(0)); // Set null path value
|
|
2380 |
|
|
2381 |
// Load the object's klass
|
|
2382 |
Node* obj_klass = load_object_klass(not_null_obj);
|
|
2383 |
|
|
2384 |
// Generate the subtype check
|
|
2385 |
Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
|
|
2386 |
|
|
2387 |
// Plug in the success path to the general merge in slot 1.
|
|
2388 |
region->init_req(_obj_path, control());
|
|
2389 |
phi ->init_req(_obj_path, intcon(1));
|
|
2390 |
|
|
2391 |
// Plug in the failing path to the general merge in slot 2.
|
|
2392 |
region->init_req(_fail_path, not_subtype_ctrl);
|
|
2393 |
phi ->init_req(_fail_path, intcon(0));
|
|
2394 |
|
|
2395 |
// Return final merged results
|
|
2396 |
set_control( _gvn.transform(region) );
|
|
2397 |
record_for_igvn(region);
|
|
2398 |
return _gvn.transform(phi);
|
|
2399 |
}
|
|
2400 |
|
|
2401 |
//-------------------------------gen_checkcast---------------------------------
|
|
2402 |
// Generate a checkcast idiom. Used by both the checkcast bytecode and the
|
|
2403 |
// array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
|
|
2404 |
// uncommon-trap paths work. Adjust stack after this call.
|
|
2405 |
// If failure_control is supplied and not null, it is filled in with
|
|
2406 |
// the control edge for the cast failure. Otherwise, an appropriate
|
|
2407 |
// uncommon trap or exception is thrown.
|
|
2408 |
Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
|
|
2409 |
Node* *failure_control) {
|
|
2410 |
kill_dead_locals(); // Benefit all the uncommon traps
|
|
2411 |
const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
|
|
2412 |
const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
|
|
2413 |
|
|
2414 |
// Fast cutout: Check the case that the cast is vacuously true.
|
|
2415 |
// This detects the common cases where the test will short-circuit
|
|
2416 |
// away completely. We do this before we perform the null check,
|
|
2417 |
// because if the test is going to turn into zero code, we don't
|
|
2418 |
// want a residual null check left around. (Causes a slowdown,
|
|
2419 |
// for example, in some objArray manipulations, such as a[i]=a[j].)
|
|
2420 |
if (tk->singleton()) {
|
|
2421 |
const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
|
|
2422 |
if (objtp != NULL && objtp->klass() != NULL) {
|
|
2423 |
switch (static_subtype_check(tk->klass(), objtp->klass())) {
|
|
2424 |
case SSC_always_true:
|
|
2425 |
return obj;
|
|
2426 |
case SSC_always_false:
|
|
2427 |
// It needs a null check because a null will *pass* the cast check.
|
|
2428 |
// A non-null value will always produce an exception.
|
|
2429 |
return do_null_assert(obj, T_OBJECT);
|
|
2430 |
}
|
|
2431 |
}
|
|
2432 |
}
|
|
2433 |
|
|
2434 |
ciProfileData* data = NULL;
|
|
2435 |
if (failure_control == NULL) { // use MDO in regular case only
|
|
2436 |
assert(java_bc() == Bytecodes::_aastore ||
|
|
2437 |
java_bc() == Bytecodes::_checkcast,
|
|
2438 |
"interpreter profiles type checks only for these BCs");
|
|
2439 |
data = method()->method_data()->bci_to_data(bci());
|
|
2440 |
}
|
|
2441 |
|
|
2442 |
// Make the merge point
|
|
2443 |
enum { _obj_path = 1, _null_path, PATH_LIMIT };
|
|
2444 |
RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
|
|
2445 |
Node* phi = new (C, PATH_LIMIT) PhiNode(region, toop);
|
|
2446 |
C->set_has_split_ifs(true); // Has chance for split-if optimization
|
|
2447 |
|
|
2448 |
// Use null-cast information if it is available
|
|
2449 |
bool never_see_null = false;
|
|
2450 |
// If we see an unexpected null at a check-cast we record it and force a
|
|
2451 |
// recompile; the offending check-cast will be compiled to handle NULLs.
|
|
2452 |
// If we see several offending BCIs, then all checkcasts in the
|
|
2453 |
// method will be compiled to handle NULLs.
|
|
2454 |
if (UncommonNullCast // Cutout for this technique
|
|
2455 |
&& failure_control == NULL // regular case
|
|
2456 |
&& obj != null() // And not the -Xcomp stupid case?
|
|
2457 |
&& !too_many_traps(Deoptimization::Reason_null_check)) {
|
|
2458 |
// Finally, check the "null_seen" bit from the interpreter.
|
|
2459 |
if (data == NULL || !data->as_BitData()->null_seen()) {
|
|
2460 |
never_see_null = true;
|
|
2461 |
}
|
|
2462 |
}
|
|
2463 |
|
|
2464 |
// Null check; get casted pointer; set region slot 3
|
|
2465 |
Node* null_ctl = top();
|
|
2466 |
Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
|
|
2467 |
|
|
2468 |
// If not_null_obj is dead, only null-path is taken
|
|
2469 |
if (stopped()) { // Doing instance-of on a NULL?
|
|
2470 |
set_control(null_ctl);
|
|
2471 |
return null();
|
|
2472 |
}
|
|
2473 |
region->init_req(_null_path, null_ctl);
|
|
2474 |
phi ->init_req(_null_path, null()); // Set null path value
|
|
2475 |
|
|
2476 |
Node* cast_obj = NULL; // the casted version of the object
|
|
2477 |
|
|
2478 |
// If the profile has seen exactly one type, narrow to that type.
|
|
2479 |
// (The subsequent subtype check will always fold up.)
|
|
2480 |
if (UseTypeProfile && TypeProfileCasts && data != NULL &&
|
|
2481 |
// Counter has never been decremented (due to cast failure).
|
|
2482 |
// ...This is a reasonable thing to expect. It is true of
|
|
2483 |
// all casts inserted by javac to implement generic types.
|
|
2484 |
data->as_CounterData()->count() >= 0 &&
|
|
2485 |
!too_many_traps(Deoptimization::Reason_class_check)) {
|
|
2486 |
// (No, this isn't a call, but it's enough like a virtual call
|
|
2487 |
// to use the same ciMethod accessor to get the profile info...)
|
|
2488 |
ciCallProfile profile = method()->call_profile_at_bci(bci());
|
|
2489 |
if (profile.count() >= 0 && // no cast failures here
|
|
2490 |
profile.has_receiver(0) &&
|
|
2491 |
profile.morphism() == 1) {
|
|
2492 |
ciKlass* exact_kls = profile.receiver(0);
|
|
2493 |
int ssc = static_subtype_check(tk->klass(), exact_kls);
|
|
2494 |
if (ssc == SSC_always_true) {
|
|
2495 |
// If we narrow the type to match what the type profile sees,
|
|
2496 |
// we can then remove the rest of the cast.
|
|
2497 |
// This is a win, even if the exact_kls is very specific,
|
|
2498 |
// because downstream operations, such as method calls,
|
|
2499 |
// will often benefit from the sharper type.
|
|
2500 |
Node* exact_obj = not_null_obj; // will get updated in place...
|
|
2501 |
Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
|
|
2502 |
&exact_obj);
|
|
2503 |
{ PreserveJVMState pjvms(this);
|
|
2504 |
set_control(slow_ctl);
|
|
2505 |
uncommon_trap(Deoptimization::Reason_class_check,
|
|
2506 |
Deoptimization::Action_maybe_recompile);
|
|
2507 |
}
|
|
2508 |
if (failure_control != NULL) // failure is now impossible
|
|
2509 |
(*failure_control) = top();
|
|
2510 |
replace_in_map(not_null_obj, exact_obj);
|
|
2511 |
// adjust the type of the phi to the exact klass:
|
|
2512 |
phi->raise_bottom_type(_gvn.type(exact_obj)->meet(TypePtr::NULL_PTR));
|
|
2513 |
cast_obj = exact_obj;
|
|
2514 |
}
|
|
2515 |
// assert(cast_obj != NULL)... except maybe the profile lied to us.
|
|
2516 |
}
|
|
2517 |
}
|
|
2518 |
|
|
2519 |
if (cast_obj == NULL) {
|
|
2520 |
// Load the object's klass
|
|
2521 |
Node* obj_klass = load_object_klass(not_null_obj);
|
|
2522 |
|
|
2523 |
// Generate the subtype check
|
|
2524 |
Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
|
|
2525 |
|
|
2526 |
// Plug in success path into the merge
|
|
2527 |
cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(),
|
|
2528 |
not_null_obj, toop));
|
|
2529 |
// Failure path ends in uncommon trap (or may be dead - failure impossible)
|
|
2530 |
if (failure_control == NULL) {
|
|
2531 |
if (not_subtype_ctrl != top()) { // If failure is possible
|
|
2532 |
PreserveJVMState pjvms(this);
|
|
2533 |
set_control(not_subtype_ctrl);
|
|
2534 |
builtin_throw(Deoptimization::Reason_class_check, obj_klass);
|
|
2535 |
}
|
|
2536 |
} else {
|
|
2537 |
(*failure_control) = not_subtype_ctrl;
|
|
2538 |
}
|
|
2539 |
}
|
|
2540 |
|
|
2541 |
region->init_req(_obj_path, control());
|
|
2542 |
phi ->init_req(_obj_path, cast_obj);
|
|
2543 |
|
|
2544 |
// A merge of NULL or Casted-NotNull obj
|
|
2545 |
Node* res = _gvn.transform(phi);
|
|
2546 |
|
|
2547 |
// Note I do NOT always 'replace_in_map(obj,result)' here.
|
|
2548 |
// if( tk->klass()->can_be_primary_super() )
|
|
2549 |
// This means that if I successfully store an Object into an array-of-String
|
|
2550 |
// I 'forget' that the Object is really now known to be a String. I have to
|
|
2551 |
// do this because we don't have true union types for interfaces - if I store
|
|
2552 |
// a Baz into an array-of-Interface and then tell the optimizer it's an
|
|
2553 |
// Interface, I forget that it's also a Baz and cannot do Baz-like field
|
|
2554 |
// references to it. FIX THIS WHEN UNION TYPES APPEAR!
|
|
2555 |
// replace_in_map( obj, res );
|
|
2556 |
|
|
2557 |
// Return final merged results
|
|
2558 |
set_control( _gvn.transform(region) );
|
|
2559 |
record_for_igvn(region);
|
|
2560 |
return res;
|
|
2561 |
}
|
|
2562 |
|
|
2563 |
//------------------------------next_monitor-----------------------------------
|
|
2564 |
// What number should be given to the next monitor?
|
|
2565 |
int GraphKit::next_monitor() {
|
|
2566 |
int current = jvms()->monitor_depth()* C->sync_stack_slots();
|
|
2567 |
int next = current + C->sync_stack_slots();
|
|
2568 |
// Keep the toplevel high water mark current:
|
|
2569 |
if (C->fixed_slots() < next) C->set_fixed_slots(next);
|
|
2570 |
return current;
|
|
2571 |
}
|
|
2572 |
|
|
2573 |
//------------------------------insert_mem_bar---------------------------------
|
|
2574 |
// Memory barrier to avoid floating things around
|
|
2575 |
// The membar serves as a pinch point between both control and all memory slices.
|
|
2576 |
Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
|
|
2577 |
MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
|
|
2578 |
mb->init_req(TypeFunc::Control, control());
|
|
2579 |
mb->init_req(TypeFunc::Memory, reset_memory());
|
|
2580 |
Node* membar = _gvn.transform(mb);
|
|
2581 |
set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) ));
|
|
2582 |
set_all_memory_call(membar);
|
|
2583 |
return membar;
|
|
2584 |
}
|
|
2585 |
|
|
2586 |
//-------------------------insert_mem_bar_volatile----------------------------
|
|
2587 |
// Memory barrier to avoid floating things around
|
|
2588 |
// The membar serves as a pinch point between both control and memory(alias_idx).
|
|
2589 |
// If you want to make a pinch point on all memory slices, do not use this
|
|
2590 |
// function (even with AliasIdxBot); use insert_mem_bar() instead.
|
|
2591 |
Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
|
|
2592 |
// When Parse::do_put_xxx updates a volatile field, it appends a series
|
|
2593 |
// of MemBarVolatile nodes, one for *each* volatile field alias category.
|
|
2594 |
// The first membar is on the same memory slice as the field store opcode.
|
|
2595 |
// This forces the membar to follow the store. (Bug 6500685 broke this.)
|
|
2596 |
// All the other membars (for other volatile slices, including AliasIdxBot,
|
|
2597 |
// which stands for all unknown volatile slices) are control-dependent
|
|
2598 |
// on the first membar. This prevents later volatile loads or stores
|
|
2599 |
// from sliding up past the just-emitted store.
|
|
2600 |
|
|
2601 |
MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
|
|
2602 |
mb->set_req(TypeFunc::Control,control());
|
|
2603 |
if (alias_idx == Compile::AliasIdxBot) {
|
|
2604 |
mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
|
|
2605 |
} else {
|
|
2606 |
assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
|
|
2607 |
mb->set_req(TypeFunc::Memory, memory(alias_idx));
|
|
2608 |
}
|
|
2609 |
Node* membar = _gvn.transform(mb);
|
|
2610 |
set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control)));
|
|
2611 |
if (alias_idx == Compile::AliasIdxBot) {
|
|
2612 |
merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)));
|
|
2613 |
} else {
|
|
2614 |
set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx);
|
|
2615 |
}
|
|
2616 |
return membar;
|
|
2617 |
}
|
|
2618 |
|
|
2619 |
//------------------------------shared_lock------------------------------------
|
|
2620 |
// Emit locking code.
|
|
2621 |
FastLockNode* GraphKit::shared_lock(Node* obj) {
|
|
2622 |
// bci is either a monitorenter bc or InvocationEntryBci
|
|
2623 |
// %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
|
|
2624 |
assert(SynchronizationEntryBCI == InvocationEntryBci, "");
|
|
2625 |
|
|
2626 |
if( !GenerateSynchronizationCode )
|
|
2627 |
return NULL; // Not locking things?
|
|
2628 |
if (stopped()) // Dead monitor?
|
|
2629 |
return NULL;
|
|
2630 |
|
|
2631 |
assert(dead_locals_are_killed(), "should kill locals before sync. point");
|
|
2632 |
|
|
2633 |
// Box the stack location
|
|
2634 |
Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor()));
|
|
2635 |
Node* mem = reset_memory();
|
|
2636 |
|
|
2637 |
FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock();
|
|
2638 |
if (PrintPreciseBiasedLockingStatistics) {
|
|
2639 |
// Create the counters for this fast lock.
|
|
2640 |
flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
|
|
2641 |
}
|
|
2642 |
// Add monitor to debug info for the slow path. If we block inside the
|
|
2643 |
// slow path and de-opt, we need the monitor hanging around
|
|
2644 |
map()->push_monitor( flock );
|
|
2645 |
|
|
2646 |
const TypeFunc *tf = LockNode::lock_type();
|
|
2647 |
LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf);
|
|
2648 |
|
|
2649 |
lock->init_req( TypeFunc::Control, control() );
|
|
2650 |
lock->init_req( TypeFunc::Memory , mem );
|
|
2651 |
lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
|
|
2652 |
lock->init_req( TypeFunc::FramePtr, frameptr() );
|
|
2653 |
lock->init_req( TypeFunc::ReturnAdr, top() );
|
|
2654 |
|
|
2655 |
lock->init_req(TypeFunc::Parms + 0, obj);
|
|
2656 |
lock->init_req(TypeFunc::Parms + 1, box);
|
|
2657 |
lock->init_req(TypeFunc::Parms + 2, flock);
|
|
2658 |
add_safepoint_edges(lock);
|
|
2659 |
|
|
2660 |
lock = _gvn.transform( lock )->as_Lock();
|
|
2661 |
|
|
2662 |
// lock has no side-effects, sets few values
|
|
2663 |
set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
|
|
2664 |
|
|
2665 |
insert_mem_bar(Op_MemBarAcquire);
|
|
2666 |
|
|
2667 |
// Add this to the worklist so that the lock can be eliminated
|
|
2668 |
record_for_igvn(lock);
|
|
2669 |
|
|
2670 |
#ifndef PRODUCT
|
|
2671 |
if (PrintLockStatistics) {
|
|
2672 |
// Update the counter for this lock. Don't bother using an atomic
|
|
2673 |
// operation since we don't require absolute accuracy.
|
|
2674 |
lock->create_lock_counter(map()->jvms());
|
|
2675 |
int adr_type = Compile::AliasIdxRaw;
|
|
2676 |
Node* counter_addr = makecon(TypeRawPtr::make(lock->counter()->addr()));
|
|
2677 |
Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type);
|
|
2678 |
Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
|
|
2679 |
store_to_memory(control(), counter_addr, incr, T_INT, adr_type);
|
|
2680 |
}
|
|
2681 |
#endif
|
|
2682 |
|
|
2683 |
return flock;
|
|
2684 |
}
|
|
2685 |
|
|
2686 |
|
|
2687 |
//------------------------------shared_unlock----------------------------------
|
|
2688 |
// Emit unlocking code.
|
|
2689 |
void GraphKit::shared_unlock(Node* box, Node* obj) {
|
|
2690 |
// bci is either a monitorenter bc or InvocationEntryBci
|
|
2691 |
// %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
|
|
2692 |
assert(SynchronizationEntryBCI == InvocationEntryBci, "");
|
|
2693 |
|
|
2694 |
if( !GenerateSynchronizationCode )
|
|
2695 |
return;
|
|
2696 |
if (stopped()) { // Dead monitor?
|
|
2697 |
map()->pop_monitor(); // Kill monitor from debug info
|
|
2698 |
return;
|
|
2699 |
}
|
|
2700 |
|
|
2701 |
// Memory barrier to avoid floating things down past the locked region
|
|
2702 |
insert_mem_bar(Op_MemBarRelease);
|
|
2703 |
|
|
2704 |
const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
|
|
2705 |
UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf);
|
|
2706 |
uint raw_idx = Compile::AliasIdxRaw;
|
|
2707 |
unlock->init_req( TypeFunc::Control, control() );
|
|
2708 |
unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
|
|
2709 |
unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
|
|
2710 |
unlock->init_req( TypeFunc::FramePtr, frameptr() );
|
|
2711 |
unlock->init_req( TypeFunc::ReturnAdr, top() );
|
|
2712 |
|
|
2713 |
unlock->init_req(TypeFunc::Parms + 0, obj);
|
|
2714 |
unlock->init_req(TypeFunc::Parms + 1, box);
|
|
2715 |
unlock = _gvn.transform(unlock)->as_Unlock();
|
|
2716 |
|
|
2717 |
Node* mem = reset_memory();
|
|
2718 |
|
|
2719 |
// unlock has no side-effects, sets few values
|
|
2720 |
set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
|
|
2721 |
|
|
2722 |
// Kill monitor from debug info
|
|
2723 |
map()->pop_monitor( );
|
|
2724 |
}
|
|
2725 |
|
|
2726 |
//-------------------------------get_layout_helper-----------------------------
|
|
2727 |
// If the given klass is a constant or known to be an array,
|
|
2728 |
// fetch the constant layout helper value into constant_value
|
|
2729 |
// and return (Node*)NULL. Otherwise, load the non-constant
|
|
2730 |
// layout helper value, and return the node which represents it.
|
|
2731 |
// This two-faced routine is useful because allocation sites
|
|
2732 |
// almost always feature constant types.
|
|
2733 |
Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
|
|
2734 |
const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
|
|
2735 |
if (!StressReflectiveCode && inst_klass != NULL) {
|
|
2736 |
ciKlass* klass = inst_klass->klass();
|
|
2737 |
bool xklass = inst_klass->klass_is_exact();
|
|
2738 |
if (xklass || klass->is_array_klass()) {
|
|
2739 |
jint lhelper = klass->layout_helper();
|
|
2740 |
if (lhelper != Klass::_lh_neutral_value) {
|
|
2741 |
constant_value = lhelper;
|
|
2742 |
return (Node*) NULL;
|
|
2743 |
}
|
|
2744 |
}
|
|
2745 |
}
|
|
2746 |
constant_value = Klass::_lh_neutral_value; // put in a known value
|
|
2747 |
Node* lhp = basic_plus_adr(klass_node, klass_node, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc));
|
|
2748 |
return make_load(NULL, lhp, TypeInt::INT, T_INT);
|
|
2749 |
}
|
|
2750 |
|
|
2751 |
// We just put in an allocate/initialize with a big raw-memory effect.
|
|
2752 |
// Hook selected additional alias categories on the initialization.
|
|
2753 |
static void hook_memory_on_init(GraphKit& kit, int alias_idx,
|
|
2754 |
MergeMemNode* init_in_merge,
|
|
2755 |
Node* init_out_raw) {
|
|
2756 |
DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
|
|
2757 |
assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
|
|
2758 |
|
|
2759 |
Node* prevmem = kit.memory(alias_idx);
|
|
2760 |
init_in_merge->set_memory_at(alias_idx, prevmem);
|
|
2761 |
kit.set_memory(init_out_raw, alias_idx);
|
|
2762 |
}
|
|
2763 |
|
|
2764 |
//---------------------------set_output_for_allocation-------------------------
|
|
2765 |
Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
|
|
2766 |
const TypeOopPtr* oop_type,
|
|
2767 |
bool raw_mem_only) {
|
|
2768 |
int rawidx = Compile::AliasIdxRaw;
|
|
2769 |
alloc->set_req( TypeFunc::FramePtr, frameptr() );
|
|
2770 |
add_safepoint_edges(alloc);
|
|
2771 |
Node* allocx = _gvn.transform(alloc);
|
|
2772 |
set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) );
|
|
2773 |
// create memory projection for i_o
|
|
2774 |
set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
|
|
2775 |
make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true);
|
|
2776 |
|
|
2777 |
// create a memory projection as for the normal control path
|
|
2778 |
Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory));
|
|
2779 |
set_memory(malloc, rawidx);
|
|
2780 |
|
|
2781 |
// a normal slow-call doesn't change i_o, but an allocation does
|
|
2782 |
// we create a separate i_o projection for the normal control path
|
|
2783 |
set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) );
|
|
2784 |
Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) );
|
|
2785 |
|
|
2786 |
// put in an initialization barrier
|
|
2787 |
InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
|
|
2788 |
rawoop)->as_Initialize();
|
|
2789 |
assert(alloc->initialization() == init, "2-way macro link must work");
|
|
2790 |
assert(init ->allocation() == alloc, "2-way macro link must work");
|
|
2791 |
if (ReduceFieldZeroing && !raw_mem_only) {
|
|
2792 |
// Extract memory strands which may participate in the new object's
|
|
2793 |
// initialization, and source them from the new InitializeNode.
|
|
2794 |
// This will allow us to observe initializations when they occur,
|
|
2795 |
// and link them properly (as a group) to the InitializeNode.
|
|
2796 |
Node* klass_node = alloc->in(AllocateNode::KlassNode);
|
|
2797 |
assert(init->in(InitializeNode::Memory) == malloc, "");
|
|
2798 |
MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
|
|
2799 |
init->set_req(InitializeNode::Memory, minit_in);
|
|
2800 |
record_for_igvn(minit_in); // fold it up later, if possible
|
|
2801 |
Node* minit_out = memory(rawidx);
|
|
2802 |
assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
|
|
2803 |
if (oop_type->isa_aryptr()) {
|
|
2804 |
const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
|
|
2805 |
int elemidx = C->get_alias_index(telemref);
|
|
2806 |
hook_memory_on_init(*this, elemidx, minit_in, minit_out);
|
|
2807 |
} else if (oop_type->isa_instptr()) {
|
|
2808 |
ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
|
|
2809 |
for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
|
|
2810 |
ciField* field = ik->nonstatic_field_at(i);
|
|
2811 |
if (field->offset() >= TrackedInitializationLimit)
|
|
2812 |
continue; // do not bother to track really large numbers of fields
|
|
2813 |
// Find (or create) the alias category for this field:
|
|
2814 |
int fieldidx = C->alias_type(field)->index();
|
|
2815 |
hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
|
|
2816 |
}
|
|
2817 |
}
|
|
2818 |
}
|
|
2819 |
|
|
2820 |
// Cast raw oop to the real thing...
|
|
2821 |
Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type);
|
|
2822 |
javaoop = _gvn.transform(javaoop);
|
|
2823 |
C->set_recent_alloc(control(), javaoop);
|
|
2824 |
assert(just_allocated_object(control()) == javaoop, "just allocated");
|
|
2825 |
|
|
2826 |
#ifdef ASSERT
|
|
2827 |
{ // Verify that the AllocateNode::Ideal_foo recognizers work:
|
|
2828 |
Node* kn = alloc->in(AllocateNode::KlassNode);
|
|
2829 |
Node* ln = alloc->in(AllocateNode::ALength);
|
|
2830 |
assert(AllocateNode::Ideal_klass(rawoop, &_gvn) == kn,
|
|
2831 |
"Ideal_klass works");
|
|
2832 |
assert(AllocateNode::Ideal_klass(javaoop, &_gvn) == kn,
|
|
2833 |
"Ideal_klass works");
|
|
2834 |
if (alloc->is_AllocateArray()) {
|
|
2835 |
assert(AllocateArrayNode::Ideal_length(rawoop, &_gvn) == ln,
|
|
2836 |
"Ideal_length works");
|
|
2837 |
assert(AllocateArrayNode::Ideal_length(javaoop, &_gvn) == ln,
|
|
2838 |
"Ideal_length works");
|
|
2839 |
} else {
|
|
2840 |
assert(ln->is_top(), "no length, please");
|
|
2841 |
}
|
|
2842 |
}
|
|
2843 |
#endif //ASSERT
|
|
2844 |
|
|
2845 |
return javaoop;
|
|
2846 |
}
|
|
2847 |
|
|
2848 |
//---------------------------new_instance--------------------------------------
|
|
2849 |
// This routine takes a klass_node which may be constant (for a static type)
|
|
2850 |
// or may be non-constant (for reflective code). It will work equally well
|
|
2851 |
// for either, and the graph will fold nicely if the optimizer later reduces
|
|
2852 |
// the type to a constant.
|
|
2853 |
// The optional arguments are for specialized use by intrinsics:
|
|
2854 |
// - If 'extra_slow_test' if not null is an extra condition for the slow-path.
|
|
2855 |
// - If 'raw_mem_only', do not cast the result to an oop.
|
|
2856 |
// - If 'return_size_val', report the the total object size to the caller.
|
|
2857 |
Node* GraphKit::new_instance(Node* klass_node,
|
|
2858 |
Node* extra_slow_test,
|
|
2859 |
bool raw_mem_only, // affect only raw memory
|
|
2860 |
Node* *return_size_val) {
|
|
2861 |
// Compute size in doublewords
|
|
2862 |
// The size is always an integral number of doublewords, represented
|
|
2863 |
// as a positive bytewise size stored in the klass's layout_helper.
|
|
2864 |
// The layout_helper also encodes (in a low bit) the need for a slow path.
|
|
2865 |
jint layout_con = Klass::_lh_neutral_value;
|
|
2866 |
Node* layout_val = get_layout_helper(klass_node, layout_con);
|
|
2867 |
int layout_is_con = (layout_val == NULL);
|
|
2868 |
|
|
2869 |
if (extra_slow_test == NULL) extra_slow_test = intcon(0);
|
|
2870 |
// Generate the initial go-slow test. It's either ALWAYS (return a
|
|
2871 |
// Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
|
|
2872 |
// case) a computed value derived from the layout_helper.
|
|
2873 |
Node* initial_slow_test = NULL;
|
|
2874 |
if (layout_is_con) {
|
|
2875 |
assert(!StressReflectiveCode, "stress mode does not use these paths");
|
|
2876 |
bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
|
|
2877 |
initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
|
|
2878 |
|
|
2879 |
} else { // reflective case
|
|
2880 |
// This reflective path is used by Unsafe.allocateInstance.
|
|
2881 |
// (It may be stress-tested by specifying StressReflectiveCode.)
|
|
2882 |
// Basically, we want to get into the VM is there's an illegal argument.
|
|
2883 |
Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
|
|
2884 |
initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) );
|
|
2885 |
if (extra_slow_test != intcon(0)) {
|
|
2886 |
initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) );
|
|
2887 |
}
|
|
2888 |
// (Macro-expander will further convert this to a Bool, if necessary.)
|
|
2889 |
}
|
|
2890 |
|
|
2891 |
// Find the size in bytes. This is easy; it's the layout_helper.
|
|
2892 |
// The size value must be valid even if the slow path is taken.
|
|
2893 |
Node* size = NULL;
|
|
2894 |
if (layout_is_con) {
|
|
2895 |
size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
|
|
2896 |
} else { // reflective case
|
|
2897 |
// This reflective path is used by clone and Unsafe.allocateInstance.
|
|
2898 |
size = ConvI2X(layout_val);
|
|
2899 |
|
|
2900 |
// Clear the low bits to extract layout_helper_size_in_bytes:
|
|
2901 |
assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
|
|
2902 |
Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
|
|
2903 |
size = _gvn.transform( new (C, 3) AndXNode(size, mask) );
|
|
2904 |
}
|
|
2905 |
if (return_size_val != NULL) {
|
|
2906 |
(*return_size_val) = size;
|
|
2907 |
}
|
|
2908 |
|
|
2909 |
// This is a precise notnull oop of the klass.
|
|
2910 |
// (Actually, it need not be precise if this is a reflective allocation.)
|
|
2911 |
// It's what we cast the result to.
|
|
2912 |
const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
|
|
2913 |
if (!tklass) tklass = TypeKlassPtr::OBJECT;
|
|
2914 |
const TypeOopPtr* oop_type = tklass->as_instance_type();
|
|
2915 |
|
|
2916 |
// Now generate allocation code
|
|
2917 |
AllocateNode* alloc
|
|
2918 |
= new (C, AllocateNode::ParmLimit)
|
|
2919 |
AllocateNode(C, AllocateNode::alloc_type(),
|
|
2920 |
control(), memory(Compile::AliasIdxRaw), i_o(),
|
|
2921 |
size, klass_node,
|
|
2922 |
initial_slow_test);
|
|
2923 |
|
|
2924 |
return set_output_for_allocation(alloc, oop_type, raw_mem_only);
|
|
2925 |
}
|
|
2926 |
|
|
2927 |
//-------------------------------new_array-------------------------------------
|
|
2928 |
// helper for both newarray and anewarray
|
|
2929 |
// The 'length' parameter is (obviously) the length of the array.
|
|
2930 |
// See comments on new_instance for the meaning of the other arguments.
|
|
2931 |
Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
|
|
2932 |
Node* length, // number of array elements
|
|
2933 |
bool raw_mem_only, // affect only raw memory
|
|
2934 |
Node* *return_size_val) {
|
|
2935 |
jint layout_con = Klass::_lh_neutral_value;
|
|
2936 |
Node* layout_val = get_layout_helper(klass_node, layout_con);
|
|
2937 |
int layout_is_con = (layout_val == NULL);
|
|
2938 |
|
|
2939 |
if (!layout_is_con && !StressReflectiveCode &&
|
|
2940 |
!too_many_traps(Deoptimization::Reason_class_check)) {
|
|
2941 |
// This is a reflective array creation site.
|
|
2942 |
// Optimistically assume that it is a subtype of Object[],
|
|
2943 |
// so that we can fold up all the address arithmetic.
|
|
2944 |
layout_con = Klass::array_layout_helper(T_OBJECT);
|
|
2945 |
Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) );
|
|
2946 |
Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) );
|
|
2947 |
{ BuildCutout unless(this, bol_lh, PROB_MAX);
|
|
2948 |
uncommon_trap(Deoptimization::Reason_class_check,
|
|
2949 |
Deoptimization::Action_maybe_recompile);
|
|
2950 |
}
|
|
2951 |
layout_val = NULL;
|
|
2952 |
layout_is_con = true;
|
|
2953 |
}
|
|
2954 |
|
|
2955 |
// Generate the initial go-slow test. Make sure we do not overflow
|
|
2956 |
// if length is huge (near 2Gig) or negative! We do not need
|
|
2957 |
// exact double-words here, just a close approximation of needed
|
|
2958 |
// double-words. We can't add any offset or rounding bits, lest we
|
|
2959 |
// take a size -1 of bytes and make it positive. Use an unsigned
|
|
2960 |
// compare, so negative sizes look hugely positive.
|
|
2961 |
int fast_size_limit = FastAllocateSizeLimit;
|
|
2962 |
if (layout_is_con) {
|
|
2963 |
assert(!StressReflectiveCode, "stress mode does not use these paths");
|
|
2964 |
// Increase the size limit if we have exact knowledge of array type.
|
|
2965 |
int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
|
|
2966 |
fast_size_limit <<= (LogBytesPerLong - log2_esize);
|
|
2967 |
}
|
|
2968 |
|
|
2969 |
Node* initial_slow_cmp = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) );
|
|
2970 |
Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) );
|
|
2971 |
if (initial_slow_test->is_Bool()) {
|
|
2972 |
// Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
|
|
2973 |
initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
|
|
2974 |
}
|
|
2975 |
|
|
2976 |
// --- Size Computation ---
|
|
2977 |
// array_size = round_to_heap(array_header + (length << elem_shift));
|
|
2978 |
// where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
|
|
2979 |
// and round_to(x, y) == ((x + y-1) & ~(y-1))
|
|
2980 |
// The rounding mask is strength-reduced, if possible.
|
|
2981 |
int round_mask = MinObjAlignmentInBytes - 1;
|
|
2982 |
Node* header_size = NULL;
|
|
2983 |
int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
|
|
2984 |
// (T_BYTE has the weakest alignment and size restrictions...)
|
|
2985 |
if (layout_is_con) {
|
|
2986 |
int hsize = Klass::layout_helper_header_size(layout_con);
|
|
2987 |
int eshift = Klass::layout_helper_log2_element_size(layout_con);
|
|
2988 |
BasicType etype = Klass::layout_helper_element_type(layout_con);
|
|
2989 |
if ((round_mask & ~right_n_bits(eshift)) == 0)
|
|
2990 |
round_mask = 0; // strength-reduce it if it goes away completely
|
|
2991 |
assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
|
|
2992 |
assert(header_size_min <= hsize, "generic minimum is smallest");
|
|
2993 |
header_size_min = hsize;
|
|
2994 |
header_size = intcon(hsize + round_mask);
|
|
2995 |
} else {
|
|
2996 |
Node* hss = intcon(Klass::_lh_header_size_shift);
|
|
2997 |
Node* hsm = intcon(Klass::_lh_header_size_mask);
|
|
2998 |
Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) );
|
|
2999 |
hsize = _gvn.transform( new(C, 3) AndINode(hsize, hsm) );
|
|
3000 |
Node* mask = intcon(round_mask);
|
|
3001 |
header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) );
|
|
3002 |
}
|
|
3003 |
|
|
3004 |
Node* elem_shift = NULL;
|
|
3005 |
if (layout_is_con) {
|
|
3006 |
int eshift = Klass::layout_helper_log2_element_size(layout_con);
|
|
3007 |
if (eshift != 0)
|
|
3008 |
elem_shift = intcon(eshift);
|
|
3009 |
} else {
|
|
3010 |
// There is no need to mask or shift this value.
|
|
3011 |
// The semantics of LShiftINode include an implicit mask to 0x1F.
|
|
3012 |
assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
|
|
3013 |
elem_shift = layout_val;
|
|
3014 |
}
|
|
3015 |
|
|
3016 |
// Transition to native address size for all offset calculations:
|
|
3017 |
Node* lengthx = ConvI2X(length);
|
|
3018 |
Node* headerx = ConvI2X(header_size);
|
|
3019 |
#ifdef _LP64
|
|
3020 |
{ const TypeLong* tllen = _gvn.find_long_type(lengthx);
|
|
3021 |
if (tllen != NULL && tllen->_lo < 0) {
|
|
3022 |
// Add a manual constraint to a positive range. Cf. array_element_address.
|
|
3023 |
jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
|
|
3024 |
if (size_max > tllen->_hi) size_max = tllen->_hi;
|
|
3025 |
const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
|
|
3026 |
lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon));
|
|
3027 |
}
|
|
3028 |
}
|
|
3029 |
#endif
|
|
3030 |
|
|
3031 |
// Combine header size (plus rounding) and body size. Then round down.
|
|
3032 |
// This computation cannot overflow, because it is used only in two
|
|
3033 |
// places, one where the length is sharply limited, and the other
|
|
3034 |
// after a successful allocation.
|
|
3035 |
Node* abody = lengthx;
|
|
3036 |
if (elem_shift != NULL)
|
|
3037 |
abody = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) );
|
|
3038 |
Node* size = _gvn.transform( new(C, 3) AddXNode(headerx, abody) );
|
|
3039 |
if (round_mask != 0) {
|
|
3040 |
Node* mask = MakeConX(~round_mask);
|
|
3041 |
size = _gvn.transform( new(C, 3) AndXNode(size, mask) );
|
|
3042 |
}
|
|
3043 |
// else if round_mask == 0, the size computation is self-rounding
|
|
3044 |
|
|
3045 |
if (return_size_val != NULL) {
|
|
3046 |
// This is the size
|
|
3047 |
(*return_size_val) = size;
|
|
3048 |
}
|
|
3049 |
|
|
3050 |
// Now generate allocation code
|
|
3051 |
// Create the AllocateArrayNode and its result projections
|
|
3052 |
AllocateArrayNode* alloc
|
|
3053 |
= new (C, AllocateArrayNode::ParmLimit)
|
|
3054 |
AllocateArrayNode(C, AllocateArrayNode::alloc_type(),
|
|
3055 |
control(), memory(Compile::AliasIdxRaw), i_o(),
|
|
3056 |
size, klass_node,
|
|
3057 |
initial_slow_test,
|
|
3058 |
length);
|
|
3059 |
|
|
3060 |
// Cast to correct type. Note that the klass_node may be constant or not,
|
|
3061 |
// and in the latter case the actual array type will be inexact also.
|
|
3062 |
// (This happens via a non-constant argument to inline_native_newArray.)
|
|
3063 |
// In any case, the value of klass_node provides the desired array type.
|
|
3064 |
const TypeInt* length_type = _gvn.find_int_type(length);
|
|
3065 |
const TypeInt* narrow_length_type = NULL;
|
|
3066 |
const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
|
|
3067 |
if (ary_type->isa_aryptr() && length_type != NULL) {
|
|
3068 |
// Try to get a better type than POS for the size
|
|
3069 |
ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
|
|
3070 |
narrow_length_type = ary_type->is_aryptr()->size();
|
|
3071 |
if (narrow_length_type == length_type)
|
|
3072 |
narrow_length_type = NULL;
|
|
3073 |
}
|
|
3074 |
|
|
3075 |
Node* javaoop = set_output_for_allocation(alloc, ary_type, raw_mem_only);
|
|
3076 |
|
|
3077 |
// Cast length on remaining path to be positive:
|
|
3078 |
if (narrow_length_type != NULL) {
|
|
3079 |
Node* ccast = new (C, 2) CastIINode(length, narrow_length_type);
|
|
3080 |
ccast->set_req(0, control());
|
|
3081 |
_gvn.set_type_bottom(ccast);
|
|
3082 |
record_for_igvn(ccast);
|
|
3083 |
if (map()->find_edge(length) >= 0) {
|
|
3084 |
replace_in_map(length, ccast);
|
|
3085 |
}
|
|
3086 |
}
|
|
3087 |
|
|
3088 |
return javaoop;
|
|
3089 |
}
|
|
3090 |
|
|
3091 |
// The following "Ideal_foo" functions are placed here because they recognize
|
|
3092 |
// the graph shapes created by the functions immediately above.
|
|
3093 |
|
|
3094 |
//---------------------------Ideal_allocation----------------------------------
|
|
3095 |
// Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
|
|
3096 |
AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
|
|
3097 |
if (ptr == NULL) { // reduce dumb test in callers
|
|
3098 |
return NULL;
|
|
3099 |
}
|
|
3100 |
if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast
|
|
3101 |
ptr = ptr->in(1);
|
|
3102 |
if (ptr == NULL) return NULL;
|
|
3103 |
}
|
|
3104 |
if (ptr->is_Proj()) {
|
|
3105 |
Node* allo = ptr->in(0);
|
|
3106 |
if (allo != NULL && allo->is_Allocate()) {
|
|
3107 |
return allo->as_Allocate();
|
|
3108 |
}
|
|
3109 |
}
|
|
3110 |
// Report failure to match.
|
|
3111 |
return NULL;
|
|
3112 |
}
|
|
3113 |
|
|
3114 |
// Fancy version which also strips off an offset (and reports it to caller).
|
|
3115 |
AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
|
|
3116 |
intptr_t& offset) {
|
|
3117 |
Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
|
|
3118 |
if (base == NULL) return NULL;
|
|
3119 |
return Ideal_allocation(base, phase);
|
|
3120 |
}
|
|
3121 |
|
|
3122 |
// Trace Initialize <- Proj[Parm] <- Allocate
|
|
3123 |
AllocateNode* InitializeNode::allocation() {
|
|
3124 |
Node* rawoop = in(InitializeNode::RawAddress);
|
|
3125 |
if (rawoop->is_Proj()) {
|
|
3126 |
Node* alloc = rawoop->in(0);
|
|
3127 |
if (alloc->is_Allocate()) {
|
|
3128 |
return alloc->as_Allocate();
|
|
3129 |
}
|
|
3130 |
}
|
|
3131 |
return NULL;
|
|
3132 |
}
|
|
3133 |
|
|
3134 |
// Trace Allocate -> Proj[Parm] -> Initialize
|
|
3135 |
InitializeNode* AllocateNode::initialization() {
|
|
3136 |
ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
|
|
3137 |
if (rawoop == NULL) return NULL;
|
|
3138 |
for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
|
|
3139 |
Node* init = rawoop->fast_out(i);
|
|
3140 |
if (init->is_Initialize()) {
|
|
3141 |
assert(init->as_Initialize()->allocation() == this, "2-way link");
|
|
3142 |
return init->as_Initialize();
|
|
3143 |
}
|
|
3144 |
}
|
|
3145 |
return NULL;
|
|
3146 |
}
|