Wed, 18 Sep 2013 14:34:56 -0700
8024342: PPC64 (part 111): Support for C calling conventions that require 64-bit ints.
Summary: Some platforms, as ppc and s390x/zArch require that 32-bit ints are passed as 64-bit values to C functions. This change adds support to adapt the signature and to issue proper casts to c2-compiled stubs. The functions are used in generate_native_wrapper(). Adapt signature used by the compiler as in PhaseIdealLoop::intrinsify_fill().
Reviewed-by: kvn
1 /*
2 * Copyright (c) 1998, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "memory/allocation.inline.hpp"
27 #include "opto/block.hpp"
28 #include "opto/c2compiler.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/cfgnode.hpp"
31 #include "opto/machnode.hpp"
32 #include "opto/runtime.hpp"
33 #ifdef TARGET_ARCH_MODEL_x86_32
34 # include "adfiles/ad_x86_32.hpp"
35 #endif
36 #ifdef TARGET_ARCH_MODEL_x86_64
37 # include "adfiles/ad_x86_64.hpp"
38 #endif
39 #ifdef TARGET_ARCH_MODEL_sparc
40 # include "adfiles/ad_sparc.hpp"
41 #endif
42 #ifdef TARGET_ARCH_MODEL_zero
43 # include "adfiles/ad_zero.hpp"
44 #endif
45 #ifdef TARGET_ARCH_MODEL_arm
46 # include "adfiles/ad_arm.hpp"
47 #endif
48 #ifdef TARGET_ARCH_MODEL_ppc_32
49 # include "adfiles/ad_ppc_32.hpp"
50 #endif
51 #ifdef TARGET_ARCH_MODEL_ppc_64
52 # include "adfiles/ad_ppc_64.hpp"
53 #endif
55 // Optimization - Graph Style
57 //------------------------------implicit_null_check----------------------------
58 // Detect implicit-null-check opportunities. Basically, find NULL checks
59 // with suitable memory ops nearby. Use the memory op to do the NULL check.
60 // I can generate a memory op if there is not one nearby.
61 // The proj is the control projection for the not-null case.
62 // The val is the pointer being checked for nullness or
63 // decodeHeapOop_not_null node if it did not fold into address.
64 void Block::implicit_null_check(PhaseCFG *cfg, Node *proj, Node *val, int allowed_reasons) {
65 // Assume if null check need for 0 offset then always needed
66 // Intel solaris doesn't support any null checks yet and no
67 // mechanism exists (yet) to set the switches at an os_cpu level
68 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;
70 // Make sure the ptr-is-null path appears to be uncommon!
71 float f = end()->as_MachIf()->_prob;
72 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
73 if( f > PROB_UNLIKELY_MAG(4) ) return;
75 uint bidx = 0; // Capture index of value into memop
76 bool was_store; // Memory op is a store op
78 // Get the successor block for if the test ptr is non-null
79 Block* not_null_block; // this one goes with the proj
80 Block* null_block;
81 if (_nodes[_nodes.size()-1] == proj) {
82 null_block = _succs[0];
83 not_null_block = _succs[1];
84 } else {
85 assert(_nodes[_nodes.size()-2] == proj, "proj is one or the other");
86 not_null_block = _succs[0];
87 null_block = _succs[1];
88 }
89 while (null_block->is_Empty() == Block::empty_with_goto) {
90 null_block = null_block->_succs[0];
91 }
93 // Search the exception block for an uncommon trap.
94 // (See Parse::do_if and Parse::do_ifnull for the reason
95 // we need an uncommon trap. Briefly, we need a way to
96 // detect failure of this optimization, as in 6366351.)
97 {
98 bool found_trap = false;
99 for (uint i1 = 0; i1 < null_block->_nodes.size(); i1++) {
100 Node* nn = null_block->_nodes[i1];
101 if (nn->is_MachCall() &&
102 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) {
103 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();
104 if (trtype->isa_int() && trtype->is_int()->is_con()) {
105 jint tr_con = trtype->is_int()->get_con();
106 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
107 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
108 assert((int)reason < (int)BitsPerInt, "recode bit map");
109 if (is_set_nth_bit(allowed_reasons, (int) reason)
110 && action != Deoptimization::Action_none) {
111 // This uncommon trap is sure to recompile, eventually.
112 // When that happens, C->too_many_traps will prevent
113 // this transformation from happening again.
114 found_trap = true;
115 }
116 }
117 break;
118 }
119 }
120 if (!found_trap) {
121 // We did not find an uncommon trap.
122 return;
123 }
124 }
126 // Check for decodeHeapOop_not_null node which did not fold into address
127 bool is_decoden = ((intptr_t)val) & 1;
128 val = (Node*)(((intptr_t)val) & ~1);
130 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() &&
131 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity");
133 // Search the successor block for a load or store who's base value is also
134 // the tested value. There may be several.
135 Node_List *out = new Node_List(Thread::current()->resource_area());
136 MachNode *best = NULL; // Best found so far
137 for (DUIterator i = val->outs(); val->has_out(i); i++) {
138 Node *m = val->out(i);
139 if( !m->is_Mach() ) continue;
140 MachNode *mach = m->as_Mach();
141 was_store = false;
142 int iop = mach->ideal_Opcode();
143 switch( iop ) {
144 case Op_LoadB:
145 case Op_LoadUB:
146 case Op_LoadUS:
147 case Op_LoadD:
148 case Op_LoadF:
149 case Op_LoadI:
150 case Op_LoadL:
151 case Op_LoadP:
152 case Op_LoadN:
153 case Op_LoadS:
154 case Op_LoadKlass:
155 case Op_LoadNKlass:
156 case Op_LoadRange:
157 case Op_LoadD_unaligned:
158 case Op_LoadL_unaligned:
159 assert(mach->in(2) == val, "should be address");
160 break;
161 case Op_StoreB:
162 case Op_StoreC:
163 case Op_StoreCM:
164 case Op_StoreD:
165 case Op_StoreF:
166 case Op_StoreI:
167 case Op_StoreL:
168 case Op_StoreP:
169 case Op_StoreN:
170 case Op_StoreNKlass:
171 was_store = true; // Memory op is a store op
172 // Stores will have their address in slot 2 (memory in slot 1).
173 // If the value being nul-checked is in another slot, it means we
174 // are storing the checked value, which does NOT check the value!
175 if( mach->in(2) != val ) continue;
176 break; // Found a memory op?
177 case Op_StrComp:
178 case Op_StrEquals:
179 case Op_StrIndexOf:
180 case Op_AryEq:
181 case Op_EncodeISOArray:
182 // Not a legit memory op for implicit null check regardless of
183 // embedded loads
184 continue;
185 default: // Also check for embedded loads
186 if( !mach->needs_anti_dependence_check() )
187 continue; // Not an memory op; skip it
188 if( must_clone[iop] ) {
189 // Do not move nodes which produce flags because
190 // RA will try to clone it to place near branch and
191 // it will cause recompilation, see clone_node().
192 continue;
193 }
194 {
195 // Check that value is used in memory address in
196 // instructions with embedded load (CmpP val1,(val2+off)).
197 Node* base;
198 Node* index;
199 const MachOper* oper = mach->memory_inputs(base, index);
200 if (oper == NULL || oper == (MachOper*)-1) {
201 continue; // Not an memory op; skip it
202 }
203 if (val == base ||
204 val == index && val->bottom_type()->isa_narrowoop()) {
205 break; // Found it
206 } else {
207 continue; // Skip it
208 }
209 }
210 break;
211 }
212 // check if the offset is not too high for implicit exception
213 {
214 intptr_t offset = 0;
215 const TypePtr *adr_type = NULL; // Do not need this return value here
216 const Node* base = mach->get_base_and_disp(offset, adr_type);
217 if (base == NULL || base == NodeSentinel) {
218 // Narrow oop address doesn't have base, only index
219 if( val->bottom_type()->isa_narrowoop() &&
220 MacroAssembler::needs_explicit_null_check(offset) )
221 continue; // Give up if offset is beyond page size
222 // cannot reason about it; is probably not implicit null exception
223 } else {
224 const TypePtr* tptr;
225 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 ||
226 Universe::narrow_klass_shift() == 0)) {
227 // 32-bits narrow oop can be the base of address expressions
228 tptr = base->get_ptr_type();
229 } else {
230 // only regular oops are expected here
231 tptr = base->bottom_type()->is_ptr();
232 }
233 // Give up if offset is not a compile-time constant
234 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
235 continue;
236 offset += tptr->_offset; // correct if base is offseted
237 if( MacroAssembler::needs_explicit_null_check(offset) )
238 continue; // Give up is reference is beyond 4K page size
239 }
240 }
242 // Check ctrl input to see if the null-check dominates the memory op
243 Block *cb = cfg->get_block_for_node(mach);
244 cb = cb->_idom; // Always hoist at least 1 block
245 if( !was_store ) { // Stores can be hoisted only one block
246 while( cb->_dom_depth > (_dom_depth + 1))
247 cb = cb->_idom; // Hoist loads as far as we want
248 // The non-null-block should dominate the memory op, too. Live
249 // range spilling will insert a spill in the non-null-block if it is
250 // needs to spill the memory op for an implicit null check.
251 if (cb->_dom_depth == (_dom_depth + 1)) {
252 if (cb != not_null_block) continue;
253 cb = cb->_idom;
254 }
255 }
256 if( cb != this ) continue;
258 // Found a memory user; see if it can be hoisted to check-block
259 uint vidx = 0; // Capture index of value into memop
260 uint j;
261 for( j = mach->req()-1; j > 0; j-- ) {
262 if( mach->in(j) == val ) {
263 vidx = j;
264 // Ignore DecodeN val which could be hoisted to where needed.
265 if( is_decoden ) continue;
266 }
267 // Block of memory-op input
268 Block *inb = cfg->get_block_for_node(mach->in(j));
269 Block *b = this; // Start from nul check
270 while( b != inb && b->_dom_depth > inb->_dom_depth )
271 b = b->_idom; // search upwards for input
272 // See if input dominates null check
273 if( b != inb )
274 break;
275 }
276 if( j > 0 )
277 continue;
278 Block *mb = cfg->get_block_for_node(mach);
279 // Hoisting stores requires more checks for the anti-dependence case.
280 // Give up hoisting if we have to move the store past any load.
281 if( was_store ) {
282 Block *b = mb; // Start searching here for a local load
283 // mach use (faulting) trying to hoist
284 // n might be blocker to hoisting
285 while( b != this ) {
286 uint k;
287 for( k = 1; k < b->_nodes.size(); k++ ) {
288 Node *n = b->_nodes[k];
289 if( n->needs_anti_dependence_check() &&
290 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
291 break; // Found anti-dependent load
292 }
293 if( k < b->_nodes.size() )
294 break; // Found anti-dependent load
295 // Make sure control does not do a merge (would have to check allpaths)
296 if( b->num_preds() != 2 ) break;
297 b = cfg->get_block_for_node(b->pred(1)); // Move up to predecessor block
298 }
299 if( b != this ) continue;
300 }
302 // Make sure this memory op is not already being used for a NullCheck
303 Node *e = mb->end();
304 if( e->is_MachNullCheck() && e->in(1) == mach )
305 continue; // Already being used as a NULL check
307 // Found a candidate! Pick one with least dom depth - the highest
308 // in the dom tree should be closest to the null check.
309 if (best == NULL || cfg->get_block_for_node(mach)->_dom_depth < cfg->get_block_for_node(best)->_dom_depth) {
310 best = mach;
311 bidx = vidx;
312 }
313 }
314 // No candidate!
315 if (best == NULL) {
316 return;
317 }
319 // ---- Found an implicit null check
320 extern int implicit_null_checks;
321 implicit_null_checks++;
323 if( is_decoden ) {
324 // Check if we need to hoist decodeHeapOop_not_null first.
325 Block *valb = cfg->get_block_for_node(val);
326 if( this != valb && this->_dom_depth < valb->_dom_depth ) {
327 // Hoist it up to the end of the test block.
328 valb->find_remove(val);
329 this->add_inst(val);
330 cfg->map_node_to_block(val, this);
331 // DecodeN on x86 may kill flags. Check for flag-killing projections
332 // that also need to be hoisted.
333 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
334 Node* n = val->fast_out(j);
335 if( n->is_MachProj() ) {
336 cfg->get_block_for_node(n)->find_remove(n);
337 this->add_inst(n);
338 cfg->map_node_to_block(n, this);
339 }
340 }
341 }
342 }
343 // Hoist the memory candidate up to the end of the test block.
344 Block *old_block = cfg->get_block_for_node(best);
345 old_block->find_remove(best);
346 add_inst(best);
347 cfg->map_node_to_block(best, this);
349 // Move the control dependence
350 if (best->in(0) && best->in(0) == old_block->_nodes[0])
351 best->set_req(0, _nodes[0]);
353 // Check for flag-killing projections that also need to be hoisted
354 // Should be DU safe because no edge updates.
355 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
356 Node* n = best->fast_out(j);
357 if( n->is_MachProj() ) {
358 cfg->get_block_for_node(n)->find_remove(n);
359 add_inst(n);
360 cfg->map_node_to_block(n, this);
361 }
362 }
364 Compile *C = cfg->C;
365 // proj==Op_True --> ne test; proj==Op_False --> eq test.
366 // One of two graph shapes got matched:
367 // (IfTrue (If (Bool NE (CmpP ptr NULL))))
368 // (IfFalse (If (Bool EQ (CmpP ptr NULL))))
369 // NULL checks are always branch-if-eq. If we see a IfTrue projection
370 // then we are replacing a 'ne' test with a 'eq' NULL check test.
371 // We need to flip the projections to keep the same semantics.
372 if( proj->Opcode() == Op_IfTrue ) {
373 // Swap order of projections in basic block to swap branch targets
374 Node *tmp1 = _nodes[end_idx()+1];
375 Node *tmp2 = _nodes[end_idx()+2];
376 _nodes.map(end_idx()+1, tmp2);
377 _nodes.map(end_idx()+2, tmp1);
378 Node *tmp = new (C) Node(C->top()); // Use not NULL input
379 tmp1->replace_by(tmp);
380 tmp2->replace_by(tmp1);
381 tmp->replace_by(tmp2);
382 tmp->destruct();
383 }
385 // Remove the existing null check; use a new implicit null check instead.
386 // Since schedule-local needs precise def-use info, we need to correct
387 // it as well.
388 Node *old_tst = proj->in(0);
389 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);
390 _nodes.map(end_idx(),nul_chk);
391 cfg->map_node_to_block(nul_chk, this);
392 // Redirect users of old_test to nul_chk
393 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
394 old_tst->last_out(i2)->set_req(0, nul_chk);
395 // Clean-up any dead code
396 for (uint i3 = 0; i3 < old_tst->req(); i3++)
397 old_tst->set_req(i3, NULL);
399 cfg->latency_from_uses(nul_chk);
400 cfg->latency_from_uses(best);
401 }
404 //------------------------------select-----------------------------------------
405 // Select a nice fellow from the worklist to schedule next. If there is only
406 // one choice, then use it. Projections take top priority for correctness
407 // reasons - if I see a projection, then it is next. There are a number of
408 // other special cases, for instructions that consume condition codes, et al.
409 // These are chosen immediately. Some instructions are required to immediately
410 // precede the last instruction in the block, and these are taken last. Of the
411 // remaining cases (most), choose the instruction with the greatest latency
412 // (that is, the most number of pseudo-cycles required to the end of the
413 // routine). If there is a tie, choose the instruction with the most inputs.
414 Node *Block::select(PhaseCFG *cfg, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) {
416 // If only a single entry on the stack, use it
417 uint cnt = worklist.size();
418 if (cnt == 1) {
419 Node *n = worklist[0];
420 worklist.map(0,worklist.pop());
421 return n;
422 }
424 uint choice = 0; // Bigger is most important
425 uint latency = 0; // Bigger is scheduled first
426 uint score = 0; // Bigger is better
427 int idx = -1; // Index in worklist
428 int cand_cnt = 0; // Candidate count
430 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
431 // Order in worklist is used to break ties.
432 // See caller for how this is used to delay scheduling
433 // of induction variable increments to after the other
434 // uses of the phi are scheduled.
435 Node *n = worklist[i]; // Get Node on worklist
437 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
438 if( n->is_Proj() || // Projections always win
439 n->Opcode()== Op_Con || // So does constant 'Top'
440 iop == Op_CreateEx || // Create-exception must start block
441 iop == Op_CheckCastPP
442 ) {
443 worklist.map(i,worklist.pop());
444 return n;
445 }
447 // Final call in a block must be adjacent to 'catch'
448 Node *e = end();
449 if( e->is_Catch() && e->in(0)->in(0) == n )
450 continue;
452 // Memory op for an implicit null check has to be at the end of the block
453 if( e->is_MachNullCheck() && e->in(1) == n )
454 continue;
456 // Schedule IV increment last.
457 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd &&
458 e->in(1)->in(1) == n && n->is_iteratively_computed())
459 continue;
461 uint n_choice = 2;
463 // See if this instruction is consumed by a branch. If so, then (as the
464 // branch is the last instruction in the basic block) force it to the
465 // end of the basic block
466 if ( must_clone[iop] ) {
467 // See if any use is a branch
468 bool found_machif = false;
470 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
471 Node* use = n->fast_out(j);
473 // The use is a conditional branch, make them adjacent
474 if (use->is_MachIf() && cfg->get_block_for_node(use) == this) {
475 found_machif = true;
476 break;
477 }
479 // More than this instruction pending for successor to be ready,
480 // don't choose this if other opportunities are ready
481 if (ready_cnt.at(use->_idx) > 1)
482 n_choice = 1;
483 }
485 // loop terminated, prefer not to use this instruction
486 if (found_machif)
487 continue;
488 }
490 // See if this has a predecessor that is "must_clone", i.e. sets the
491 // condition code. If so, choose this first
492 for (uint j = 0; j < n->req() ; j++) {
493 Node *inn = n->in(j);
494 if (inn) {
495 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
496 n_choice = 3;
497 break;
498 }
499 }
500 }
502 // MachTemps should be scheduled last so they are near their uses
503 if (n->is_MachTemp()) {
504 n_choice = 1;
505 }
507 uint n_latency = cfg->get_latency_for_node(n);
508 uint n_score = n->req(); // Many inputs get high score to break ties
510 // Keep best latency found
511 cand_cnt++;
512 if (choice < n_choice ||
513 (choice == n_choice &&
514 ((StressLCM && Compile::randomized_select(cand_cnt)) ||
515 (!StressLCM &&
516 (latency < n_latency ||
517 (latency == n_latency &&
518 (score < n_score))))))) {
519 choice = n_choice;
520 latency = n_latency;
521 score = n_score;
522 idx = i; // Also keep index in worklist
523 }
524 } // End of for all ready nodes in worklist
526 assert(idx >= 0, "index should be set");
527 Node *n = worklist[(uint)idx]; // Get the winner
529 worklist.map((uint)idx, worklist.pop()); // Compress worklist
530 return n;
531 }
534 //------------------------------set_next_call----------------------------------
535 void Block::set_next_call( Node *n, VectorSet &next_call, PhaseCFG* cfg) {
536 if( next_call.test_set(n->_idx) ) return;
537 for( uint i=0; i<n->len(); i++ ) {
538 Node *m = n->in(i);
539 if( !m ) continue; // must see all nodes in block that precede call
540 if (cfg->get_block_for_node(m) == this) {
541 set_next_call(m, next_call, cfg);
542 }
543 }
544 }
546 //------------------------------needed_for_next_call---------------------------
547 // Set the flag 'next_call' for each Node that is needed for the next call to
548 // be scheduled. This flag lets me bias scheduling so Nodes needed for the
549 // next subroutine call get priority - basically it moves things NOT needed
550 // for the next call till after the call. This prevents me from trying to
551 // carry lots of stuff live across a call.
552 void Block::needed_for_next_call(Node *this_call, VectorSet &next_call, PhaseCFG* cfg) {
553 // Find the next control-defining Node in this block
554 Node* call = NULL;
555 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
556 Node* m = this_call->fast_out(i);
557 if(cfg->get_block_for_node(m) == this && // Local-block user
558 m != this_call && // Not self-start node
559 m->is_MachCall() )
560 call = m;
561 break;
562 }
563 if (call == NULL) return; // No next call (e.g., block end is near)
564 // Set next-call for all inputs to this call
565 set_next_call(call, next_call, cfg);
566 }
568 //------------------------------add_call_kills-------------------------------------
569 void Block::add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) {
570 // Fill in the kill mask for the call
571 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
572 if( !regs.Member(r) ) { // Not already defined by the call
573 // Save-on-call register?
574 if ((save_policy[r] == 'C') ||
575 (save_policy[r] == 'A') ||
576 ((save_policy[r] == 'E') && exclude_soe)) {
577 proj->_rout.Insert(r);
578 }
579 }
580 }
581 }
584 //------------------------------sched_call-------------------------------------
585 uint Block::sched_call( Matcher &matcher, PhaseCFG* cfg, uint node_cnt, Node_List &worklist, GrowableArray<int> &ready_cnt, MachCallNode *mcall, VectorSet &next_call ) {
586 RegMask regs;
588 // Schedule all the users of the call right now. All the users are
589 // projection Nodes, so they must be scheduled next to the call.
590 // Collect all the defined registers.
591 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
592 Node* n = mcall->fast_out(i);
593 assert( n->is_MachProj(), "" );
594 int n_cnt = ready_cnt.at(n->_idx)-1;
595 ready_cnt.at_put(n->_idx, n_cnt);
596 assert( n_cnt == 0, "" );
597 // Schedule next to call
598 _nodes.map(node_cnt++, n);
599 // Collect defined registers
600 regs.OR(n->out_RegMask());
601 // Check for scheduling the next control-definer
602 if( n->bottom_type() == Type::CONTROL )
603 // Warm up next pile of heuristic bits
604 needed_for_next_call(n, next_call, cfg);
606 // Children of projections are now all ready
607 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
608 Node* m = n->fast_out(j); // Get user
609 if(cfg->get_block_for_node(m) != this) {
610 continue;
611 }
612 if( m->is_Phi() ) continue;
613 int m_cnt = ready_cnt.at(m->_idx)-1;
614 ready_cnt.at_put(m->_idx, m_cnt);
615 if( m_cnt == 0 )
616 worklist.push(m);
617 }
619 }
621 // Act as if the call defines the Frame Pointer.
622 // Certainly the FP is alive and well after the call.
623 regs.Insert(matcher.c_frame_pointer());
625 // Set all registers killed and not already defined by the call.
626 uint r_cnt = mcall->tf()->range()->cnt();
627 int op = mcall->ideal_Opcode();
628 MachProjNode *proj = new (matcher.C) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
629 cfg->map_node_to_block(proj, this);
630 _nodes.insert(node_cnt++, proj);
632 // Select the right register save policy.
633 const char * save_policy;
634 switch (op) {
635 case Op_CallRuntime:
636 case Op_CallLeaf:
637 case Op_CallLeafNoFP:
638 // Calling C code so use C calling convention
639 save_policy = matcher._c_reg_save_policy;
640 break;
642 case Op_CallStaticJava:
643 case Op_CallDynamicJava:
644 // Calling Java code so use Java calling convention
645 save_policy = matcher._register_save_policy;
646 break;
648 default:
649 ShouldNotReachHere();
650 }
652 // When using CallRuntime mark SOE registers as killed by the call
653 // so values that could show up in the RegisterMap aren't live in a
654 // callee saved register since the register wouldn't know where to
655 // find them. CallLeaf and CallLeafNoFP are ok because they can't
656 // have debug info on them. Strictly speaking this only needs to be
657 // done for oops since idealreg2debugmask takes care of debug info
658 // references but there no way to handle oops differently than other
659 // pointers as far as the kill mask goes.
660 bool exclude_soe = op == Op_CallRuntime;
662 // If the call is a MethodHandle invoke, we need to exclude the
663 // register which is used to save the SP value over MH invokes from
664 // the mask. Otherwise this register could be used for
665 // deoptimization information.
666 if (op == Op_CallStaticJava) {
667 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall;
668 if (mcallstaticjava->_method_handle_invoke)
669 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
670 }
672 add_call_kills(proj, regs, save_policy, exclude_soe);
674 return node_cnt;
675 }
678 //------------------------------schedule_local---------------------------------
679 // Topological sort within a block. Someday become a real scheduler.
680 bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, GrowableArray<int> &ready_cnt, VectorSet &next_call) {
681 // Already "sorted" are the block start Node (as the first entry), and
682 // the block-ending Node and any trailing control projections. We leave
683 // these alone. PhiNodes and ParmNodes are made to follow the block start
684 // Node. Everything else gets topo-sorted.
686 #ifndef PRODUCT
687 if (cfg->trace_opto_pipelining()) {
688 tty->print_cr("# --- schedule_local B%d, before: ---", _pre_order);
689 for (uint i = 0;i < _nodes.size();i++) {
690 tty->print("# ");
691 _nodes[i]->fast_dump();
692 }
693 tty->print_cr("#");
694 }
695 #endif
697 // RootNode is already sorted
698 if( _nodes.size() == 1 ) return true;
700 // Move PhiNodes and ParmNodes from 1 to cnt up to the start
701 uint node_cnt = end_idx();
702 uint phi_cnt = 1;
703 uint i;
704 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
705 Node *n = _nodes[i];
706 if( n->is_Phi() || // Found a PhiNode or ParmNode
707 (n->is_Proj() && n->in(0) == head()) ) {
708 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
709 _nodes.map(i,_nodes[phi_cnt]);
710 _nodes.map(phi_cnt++,n); // swap Phi/Parm up front
711 } else { // All others
712 // Count block-local inputs to 'n'
713 uint cnt = n->len(); // Input count
714 uint local = 0;
715 for( uint j=0; j<cnt; j++ ) {
716 Node *m = n->in(j);
717 if( m && cfg->get_block_for_node(m) == this && !m->is_top() )
718 local++; // One more block-local input
719 }
720 ready_cnt.at_put(n->_idx, local); // Count em up
722 #ifdef ASSERT
723 if( UseConcMarkSweepGC || UseG1GC ) {
724 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
725 // Check the precedence edges
726 for (uint prec = n->req(); prec < n->len(); prec++) {
727 Node* oop_store = n->in(prec);
728 if (oop_store != NULL) {
729 assert(cfg->get_block_for_node(oop_store)->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark");
730 }
731 }
732 }
733 }
734 #endif
736 // A few node types require changing a required edge to a precedence edge
737 // before allocation.
738 if( n->is_Mach() && n->req() > TypeFunc::Parms &&
739 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
740 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
741 // MemBarAcquire could be created without Precedent edge.
742 // del_req() replaces the specified edge with the last input edge
743 // and then removes the last edge. If the specified edge > number of
744 // edges the last edge will be moved outside of the input edges array
745 // and the edge will be lost. This is why this code should be
746 // executed only when Precedent (== TypeFunc::Parms) edge is present.
747 Node *x = n->in(TypeFunc::Parms);
748 n->del_req(TypeFunc::Parms);
749 n->add_prec(x);
750 }
751 }
752 }
753 for(uint i2=i; i2<_nodes.size(); i2++ ) // Trailing guys get zapped count
754 ready_cnt.at_put(_nodes[i2]->_idx, 0);
756 // All the prescheduled guys do not hold back internal nodes
757 uint i3;
758 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled
759 Node *n = _nodes[i3]; // Get pre-scheduled
760 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
761 Node* m = n->fast_out(j);
762 if (cfg->get_block_for_node(m) == this) { // Local-block user
763 int m_cnt = ready_cnt.at(m->_idx)-1;
764 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count
765 }
766 }
767 }
769 Node_List delay;
770 // Make a worklist
771 Node_List worklist;
772 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
773 Node *m = _nodes[i4];
774 if( !ready_cnt.at(m->_idx) ) { // Zero ready count?
775 if (m->is_iteratively_computed()) {
776 // Push induction variable increments last to allow other uses
777 // of the phi to be scheduled first. The select() method breaks
778 // ties in scheduling by worklist order.
779 delay.push(m);
780 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
781 // Force the CreateEx to the top of the list so it's processed
782 // first and ends up at the start of the block.
783 worklist.insert(0, m);
784 } else {
785 worklist.push(m); // Then on to worklist!
786 }
787 }
788 }
789 while (delay.size()) {
790 Node* d = delay.pop();
791 worklist.push(d);
792 }
794 // Warm up the 'next_call' heuristic bits
795 needed_for_next_call(_nodes[0], next_call, cfg);
797 #ifndef PRODUCT
798 if (cfg->trace_opto_pipelining()) {
799 for (uint j=0; j<_nodes.size(); j++) {
800 Node *n = _nodes[j];
801 int idx = n->_idx;
802 tty->print("# ready cnt:%3d ", ready_cnt.at(idx));
803 tty->print("latency:%3d ", cfg->get_latency_for_node(n));
804 tty->print("%4d: %s\n", idx, n->Name());
805 }
806 }
807 #endif
809 uint max_idx = (uint)ready_cnt.length();
810 // Pull from worklist and schedule
811 while( worklist.size() ) { // Worklist is not ready
813 #ifndef PRODUCT
814 if (cfg->trace_opto_pipelining()) {
815 tty->print("# ready list:");
816 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
817 Node *n = worklist[i]; // Get Node on worklist
818 tty->print(" %d", n->_idx);
819 }
820 tty->cr();
821 }
822 #endif
824 // Select and pop a ready guy from worklist
825 Node* n = select(cfg, worklist, ready_cnt, next_call, phi_cnt);
826 _nodes.map(phi_cnt++,n); // Schedule him next
828 #ifndef PRODUCT
829 if (cfg->trace_opto_pipelining()) {
830 tty->print("# select %d: %s", n->_idx, n->Name());
831 tty->print(", latency:%d", cfg->get_latency_for_node(n));
832 n->dump();
833 if (Verbose) {
834 tty->print("# ready list:");
835 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
836 Node *n = worklist[i]; // Get Node on worklist
837 tty->print(" %d", n->_idx);
838 }
839 tty->cr();
840 }
841 }
843 #endif
844 if( n->is_MachCall() ) {
845 MachCallNode *mcall = n->as_MachCall();
846 phi_cnt = sched_call(matcher, cfg, phi_cnt, worklist, ready_cnt, mcall, next_call);
847 continue;
848 }
850 if (n->is_Mach() && n->as_Mach()->has_call()) {
851 RegMask regs;
852 regs.Insert(matcher.c_frame_pointer());
853 regs.OR(n->out_RegMask());
855 MachProjNode *proj = new (matcher.C) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj );
856 cfg->map_node_to_block(proj, this);
857 _nodes.insert(phi_cnt++, proj);
859 add_call_kills(proj, regs, matcher._c_reg_save_policy, false);
860 }
862 // Children are now all ready
863 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
864 Node* m = n->fast_out(i5); // Get user
865 if (cfg->get_block_for_node(m) != this) {
866 continue;
867 }
868 if( m->is_Phi() ) continue;
869 if (m->_idx >= max_idx) { // new node, skip it
870 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types");
871 continue;
872 }
873 int m_cnt = ready_cnt.at(m->_idx)-1;
874 ready_cnt.at_put(m->_idx, m_cnt);
875 if( m_cnt == 0 )
876 worklist.push(m);
877 }
878 }
880 if( phi_cnt != end_idx() ) {
881 // did not schedule all. Retry, Bailout, or Die
882 Compile* C = matcher.C;
883 if (C->subsume_loads() == true && !C->failing()) {
884 // Retry with subsume_loads == false
885 // If this is the first failure, the sentinel string will "stick"
886 // to the Compile object, and the C2Compiler will see it and retry.
887 C->record_failure(C2Compiler::retry_no_subsuming_loads());
888 }
889 // assert( phi_cnt == end_idx(), "did not schedule all" );
890 return false;
891 }
893 #ifndef PRODUCT
894 if (cfg->trace_opto_pipelining()) {
895 tty->print_cr("#");
896 tty->print_cr("# after schedule_local");
897 for (uint i = 0;i < _nodes.size();i++) {
898 tty->print("# ");
899 _nodes[i]->fast_dump();
900 }
901 tty->cr();
902 }
903 #endif
906 return true;
907 }
909 //--------------------------catch_cleanup_fix_all_inputs-----------------------
910 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
911 for (uint l = 0; l < use->len(); l++) {
912 if (use->in(l) == old_def) {
913 if (l < use->req()) {
914 use->set_req(l, new_def);
915 } else {
916 use->rm_prec(l);
917 use->add_prec(new_def);
918 l--;
919 }
920 }
921 }
922 }
924 //------------------------------catch_cleanup_find_cloned_def------------------
925 static Node *catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, PhaseCFG* cfg, int n_clone_idx) {
926 assert( use_blk != def_blk, "Inter-block cleanup only");
928 // The use is some block below the Catch. Find and return the clone of the def
929 // that dominates the use. If there is no clone in a dominating block, then
930 // create a phi for the def in a dominating block.
932 // Find which successor block dominates this use. The successor
933 // blocks must all be single-entry (from the Catch only; I will have
934 // split blocks to make this so), hence they all dominate.
935 while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
936 use_blk = use_blk->_idom;
938 // Find the successor
939 Node *fixup = NULL;
941 uint j;
942 for( j = 0; j < def_blk->_num_succs; j++ )
943 if( use_blk == def_blk->_succs[j] )
944 break;
946 if( j == def_blk->_num_succs ) {
947 // Block at same level in dom-tree is not a successor. It needs a
948 // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
949 Node_Array inputs = new Node_List(Thread::current()->resource_area());
950 for(uint k = 1; k < use_blk->num_preds(); k++) {
951 Block* block = cfg->get_block_for_node(use_blk->pred(k));
952 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, cfg, n_clone_idx));
953 }
955 // Check to see if the use_blk already has an identical phi inserted.
956 // If it exists, it will be at the first position since all uses of a
957 // def are processed together.
958 Node *phi = use_blk->_nodes[1];
959 if( phi->is_Phi() ) {
960 fixup = phi;
961 for (uint k = 1; k < use_blk->num_preds(); k++) {
962 if (phi->in(k) != inputs[k]) {
963 // Not a match
964 fixup = NULL;
965 break;
966 }
967 }
968 }
970 // If an existing PhiNode was not found, make a new one.
971 if (fixup == NULL) {
972 Node *new_phi = PhiNode::make(use_blk->head(), def);
973 use_blk->_nodes.insert(1, new_phi);
974 cfg->map_node_to_block(new_phi, use_blk);
975 for (uint k = 1; k < use_blk->num_preds(); k++) {
976 new_phi->set_req(k, inputs[k]);
977 }
978 fixup = new_phi;
979 }
981 } else {
982 // Found the use just below the Catch. Make it use the clone.
983 fixup = use_blk->_nodes[n_clone_idx];
984 }
986 return fixup;
987 }
989 //--------------------------catch_cleanup_intra_block--------------------------
990 // Fix all input edges in use that reference "def". The use is in the same
991 // block as the def and both have been cloned in each successor block.
992 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
994 // Both the use and def have been cloned. For each successor block,
995 // get the clone of the use, and make its input the clone of the def
996 // found in that block.
998 uint use_idx = blk->find_node(use);
999 uint offset_idx = use_idx - beg;
1000 for( uint k = 0; k < blk->_num_succs; k++ ) {
1001 // Get clone in each successor block
1002 Block *sb = blk->_succs[k];
1003 Node *clone = sb->_nodes[offset_idx+1];
1004 assert( clone->Opcode() == use->Opcode(), "" );
1006 // Make use-clone reference the def-clone
1007 catch_cleanup_fix_all_inputs(clone, def, sb->_nodes[n_clone_idx]);
1008 }
1009 }
1011 //------------------------------catch_cleanup_inter_block---------------------
1012 // Fix all input edges in use that reference "def". The use is in a different
1013 // block than the def.
1014 static void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, PhaseCFG* cfg, int n_clone_idx) {
1015 if( !use_blk ) return; // Can happen if the use is a precedence edge
1017 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, cfg, n_clone_idx);
1018 catch_cleanup_fix_all_inputs(use, def, new_def);
1019 }
1021 //------------------------------call_catch_cleanup-----------------------------
1022 // If we inserted any instructions between a Call and his CatchNode,
1023 // clone the instructions on all paths below the Catch.
1024 void Block::call_catch_cleanup(PhaseCFG* cfg, Compile* C) {
1026 // End of region to clone
1027 uint end = end_idx();
1028 if( !_nodes[end]->is_Catch() ) return;
1029 // Start of region to clone
1030 uint beg = end;
1031 while(!_nodes[beg-1]->is_MachProj() ||
1032 !_nodes[beg-1]->in(0)->is_MachCall() ) {
1033 beg--;
1034 assert(beg > 0,"Catch cleanup walking beyond block boundary");
1035 }
1036 // Range of inserted instructions is [beg, end)
1037 if( beg == end ) return;
1039 // Clone along all Catch output paths. Clone area between the 'beg' and
1040 // 'end' indices.
1041 for( uint i = 0; i < _num_succs; i++ ) {
1042 Block *sb = _succs[i];
1043 // Clone the entire area; ignoring the edge fixup for now.
1044 for( uint j = end; j > beg; j-- ) {
1045 // It is safe here to clone a node with anti_dependence
1046 // since clones dominate on each path.
1047 Node *clone = _nodes[j-1]->clone();
1048 sb->_nodes.insert( 1, clone );
1049 cfg->map_node_to_block(clone, sb);
1050 }
1051 }
1054 // Fixup edges. Check the def-use info per cloned Node
1055 for(uint i2 = beg; i2 < end; i2++ ) {
1056 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
1057 Node *n = _nodes[i2]; // Node that got cloned
1058 // Need DU safe iterator because of edge manipulation in calls.
1059 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
1060 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
1061 out->push(n->fast_out(j1));
1062 }
1063 uint max = out->size();
1064 for (uint j = 0; j < max; j++) {// For all users
1065 Node *use = out->pop();
1066 Block *buse = cfg->get_block_for_node(use);
1067 if( use->is_Phi() ) {
1068 for( uint k = 1; k < use->req(); k++ )
1069 if( use->in(k) == n ) {
1070 Block* block = cfg->get_block_for_node(buse->pred(k));
1071 Node *fixup = catch_cleanup_find_cloned_def(block, n, this, cfg, n_clone_idx);
1072 use->set_req(k, fixup);
1073 }
1074 } else {
1075 if (this == buse) {
1076 catch_cleanup_intra_block(use, n, this, beg, n_clone_idx);
1077 } else {
1078 catch_cleanup_inter_block(use, buse, n, this, cfg, n_clone_idx);
1079 }
1080 }
1081 } // End for all users
1083 } // End of for all Nodes in cloned area
1085 // Remove the now-dead cloned ops
1086 for(uint i3 = beg; i3 < end; i3++ ) {
1087 _nodes[beg]->disconnect_inputs(NULL, C);
1088 _nodes.remove(beg);
1089 }
1091 // If the successor blocks have a CreateEx node, move it back to the top
1092 for(uint i4 = 0; i4 < _num_succs; i4++ ) {
1093 Block *sb = _succs[i4];
1094 uint new_cnt = end - beg;
1095 // Remove any newly created, but dead, nodes.
1096 for( uint j = new_cnt; j > 0; j-- ) {
1097 Node *n = sb->_nodes[j];
1098 if (n->outcnt() == 0 &&
1099 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
1100 n->disconnect_inputs(NULL, C);
1101 sb->_nodes.remove(j);
1102 new_cnt--;
1103 }
1104 }
1105 // If any newly created nodes remain, move the CreateEx node to the top
1106 if (new_cnt > 0) {
1107 Node *cex = sb->_nodes[1+new_cnt];
1108 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
1109 sb->_nodes.remove(1+new_cnt);
1110 sb->_nodes.insert(1,cex);
1111 }
1112 }
1113 }
1114 }