Tue, 22 Jan 2013 15:34:16 -0800
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
Summary: Use SSE4.2 and AVX2 instructions for encodeArray intrinsic.
Reviewed-by: roland
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
49 # include "adfiles/ad_ppc.hpp"
50 #endif
52 // Optimization - Graph Style
54 //------------------------------implicit_null_check----------------------------
55 // Detect implicit-null-check opportunities. Basically, find NULL checks
56 // with suitable memory ops nearby. Use the memory op to do the NULL check.
57 // I can generate a memory op if there is not one nearby.
58 // The proj is the control projection for the not-null case.
59 // The val is the pointer being checked for nullness or
60 // decodeHeapOop_not_null node if it did not fold into address.
61 void Block::implicit_null_check(PhaseCFG *cfg, Node *proj, Node *val, int allowed_reasons) {
62 // Assume if null check need for 0 offset then always needed
63 // Intel solaris doesn't support any null checks yet and no
64 // mechanism exists (yet) to set the switches at an os_cpu level
65 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;
67 // Make sure the ptr-is-null path appears to be uncommon!
68 float f = end()->as_MachIf()->_prob;
69 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
70 if( f > PROB_UNLIKELY_MAG(4) ) return;
72 uint bidx = 0; // Capture index of value into memop
73 bool was_store; // Memory op is a store op
75 // Get the successor block for if the test ptr is non-null
76 Block* not_null_block; // this one goes with the proj
77 Block* null_block;
78 if (_nodes[_nodes.size()-1] == proj) {
79 null_block = _succs[0];
80 not_null_block = _succs[1];
81 } else {
82 assert(_nodes[_nodes.size()-2] == proj, "proj is one or the other");
83 not_null_block = _succs[0];
84 null_block = _succs[1];
85 }
86 while (null_block->is_Empty() == Block::empty_with_goto) {
87 null_block = null_block->_succs[0];
88 }
90 // Search the exception block for an uncommon trap.
91 // (See Parse::do_if and Parse::do_ifnull for the reason
92 // we need an uncommon trap. Briefly, we need a way to
93 // detect failure of this optimization, as in 6366351.)
94 {
95 bool found_trap = false;
96 for (uint i1 = 0; i1 < null_block->_nodes.size(); i1++) {
97 Node* nn = null_block->_nodes[i1];
98 if (nn->is_MachCall() &&
99 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) {
100 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();
101 if (trtype->isa_int() && trtype->is_int()->is_con()) {
102 jint tr_con = trtype->is_int()->get_con();
103 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
104 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
105 assert((int)reason < (int)BitsPerInt, "recode bit map");
106 if (is_set_nth_bit(allowed_reasons, (int) reason)
107 && action != Deoptimization::Action_none) {
108 // This uncommon trap is sure to recompile, eventually.
109 // When that happens, C->too_many_traps will prevent
110 // this transformation from happening again.
111 found_trap = true;
112 }
113 }
114 break;
115 }
116 }
117 if (!found_trap) {
118 // We did not find an uncommon trap.
119 return;
120 }
121 }
123 // Check for decodeHeapOop_not_null node which did not fold into address
124 bool is_decoden = ((intptr_t)val) & 1;
125 val = (Node*)(((intptr_t)val) & ~1);
127 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() &&
128 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity");
130 // Search the successor block for a load or store who's base value is also
131 // the tested value. There may be several.
132 Node_List *out = new Node_List(Thread::current()->resource_area());
133 MachNode *best = NULL; // Best found so far
134 for (DUIterator i = val->outs(); val->has_out(i); i++) {
135 Node *m = val->out(i);
136 if( !m->is_Mach() ) continue;
137 MachNode *mach = m->as_Mach();
138 was_store = false;
139 int iop = mach->ideal_Opcode();
140 switch( iop ) {
141 case Op_LoadB:
142 case Op_LoadUB:
143 case Op_LoadUS:
144 case Op_LoadD:
145 case Op_LoadF:
146 case Op_LoadI:
147 case Op_LoadL:
148 case Op_LoadP:
149 case Op_LoadN:
150 case Op_LoadS:
151 case Op_LoadKlass:
152 case Op_LoadNKlass:
153 case Op_LoadRange:
154 case Op_LoadD_unaligned:
155 case Op_LoadL_unaligned:
156 assert(mach->in(2) == val, "should be address");
157 break;
158 case Op_StoreB:
159 case Op_StoreC:
160 case Op_StoreCM:
161 case Op_StoreD:
162 case Op_StoreF:
163 case Op_StoreI:
164 case Op_StoreL:
165 case Op_StoreP:
166 case Op_StoreN:
167 case Op_StoreNKlass:
168 was_store = true; // Memory op is a store op
169 // Stores will have their address in slot 2 (memory in slot 1).
170 // If the value being nul-checked is in another slot, it means we
171 // are storing the checked value, which does NOT check the value!
172 if( mach->in(2) != val ) continue;
173 break; // Found a memory op?
174 case Op_StrComp:
175 case Op_StrEquals:
176 case Op_StrIndexOf:
177 case Op_AryEq:
178 case Op_EncodeISOArray:
179 // Not a legit memory op for implicit null check regardless of
180 // embedded loads
181 continue;
182 default: // Also check for embedded loads
183 if( !mach->needs_anti_dependence_check() )
184 continue; // Not an memory op; skip it
185 if( must_clone[iop] ) {
186 // Do not move nodes which produce flags because
187 // RA will try to clone it to place near branch and
188 // it will cause recompilation, see clone_node().
189 continue;
190 }
191 {
192 // Check that value is used in memory address in
193 // instructions with embedded load (CmpP val1,(val2+off)).
194 Node* base;
195 Node* index;
196 const MachOper* oper = mach->memory_inputs(base, index);
197 if (oper == NULL || oper == (MachOper*)-1) {
198 continue; // Not an memory op; skip it
199 }
200 if (val == base ||
201 val == index && val->bottom_type()->isa_narrowoop()) {
202 break; // Found it
203 } else {
204 continue; // Skip it
205 }
206 }
207 break;
208 }
209 // check if the offset is not too high for implicit exception
210 {
211 intptr_t offset = 0;
212 const TypePtr *adr_type = NULL; // Do not need this return value here
213 const Node* base = mach->get_base_and_disp(offset, adr_type);
214 if (base == NULL || base == NodeSentinel) {
215 // Narrow oop address doesn't have base, only index
216 if( val->bottom_type()->isa_narrowoop() &&
217 MacroAssembler::needs_explicit_null_check(offset) )
218 continue; // Give up if offset is beyond page size
219 // cannot reason about it; is probably not implicit null exception
220 } else {
221 const TypePtr* tptr;
222 if (UseCompressedOops && Universe::narrow_oop_shift() == 0) {
223 // 32-bits narrow oop can be the base of address expressions
224 tptr = base->bottom_type()->make_ptr();
225 } else {
226 // only regular oops are expected here
227 tptr = base->bottom_type()->is_ptr();
228 }
229 // Give up if offset is not a compile-time constant
230 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
231 continue;
232 offset += tptr->_offset; // correct if base is offseted
233 if( MacroAssembler::needs_explicit_null_check(offset) )
234 continue; // Give up is reference is beyond 4K page size
235 }
236 }
238 // Check ctrl input to see if the null-check dominates the memory op
239 Block *cb = cfg->_bbs[mach->_idx];
240 cb = cb->_idom; // Always hoist at least 1 block
241 if( !was_store ) { // Stores can be hoisted only one block
242 while( cb->_dom_depth > (_dom_depth + 1))
243 cb = cb->_idom; // Hoist loads as far as we want
244 // The non-null-block should dominate the memory op, too. Live
245 // range spilling will insert a spill in the non-null-block if it is
246 // needs to spill the memory op for an implicit null check.
247 if (cb->_dom_depth == (_dom_depth + 1)) {
248 if (cb != not_null_block) continue;
249 cb = cb->_idom;
250 }
251 }
252 if( cb != this ) continue;
254 // Found a memory user; see if it can be hoisted to check-block
255 uint vidx = 0; // Capture index of value into memop
256 uint j;
257 for( j = mach->req()-1; j > 0; j-- ) {
258 if( mach->in(j) == val ) {
259 vidx = j;
260 // Ignore DecodeN val which could be hoisted to where needed.
261 if( is_decoden ) continue;
262 }
263 // Block of memory-op input
264 Block *inb = cfg->_bbs[mach->in(j)->_idx];
265 Block *b = this; // Start from nul check
266 while( b != inb && b->_dom_depth > inb->_dom_depth )
267 b = b->_idom; // search upwards for input
268 // See if input dominates null check
269 if( b != inb )
270 break;
271 }
272 if( j > 0 )
273 continue;
274 Block *mb = cfg->_bbs[mach->_idx];
275 // Hoisting stores requires more checks for the anti-dependence case.
276 // Give up hoisting if we have to move the store past any load.
277 if( was_store ) {
278 Block *b = mb; // Start searching here for a local load
279 // mach use (faulting) trying to hoist
280 // n might be blocker to hoisting
281 while( b != this ) {
282 uint k;
283 for( k = 1; k < b->_nodes.size(); k++ ) {
284 Node *n = b->_nodes[k];
285 if( n->needs_anti_dependence_check() &&
286 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
287 break; // Found anti-dependent load
288 }
289 if( k < b->_nodes.size() )
290 break; // Found anti-dependent load
291 // Make sure control does not do a merge (would have to check allpaths)
292 if( b->num_preds() != 2 ) break;
293 b = cfg->_bbs[b->pred(1)->_idx]; // Move up to predecessor block
294 }
295 if( b != this ) continue;
296 }
298 // Make sure this memory op is not already being used for a NullCheck
299 Node *e = mb->end();
300 if( e->is_MachNullCheck() && e->in(1) == mach )
301 continue; // Already being used as a NULL check
303 // Found a candidate! Pick one with least dom depth - the highest
304 // in the dom tree should be closest to the null check.
305 if( !best ||
306 cfg->_bbs[mach->_idx]->_dom_depth < cfg->_bbs[best->_idx]->_dom_depth ) {
307 best = mach;
308 bidx = vidx;
310 }
311 }
312 // No candidate!
313 if( !best ) return;
315 // ---- Found an implicit null check
316 extern int implicit_null_checks;
317 implicit_null_checks++;
319 if( is_decoden ) {
320 // Check if we need to hoist decodeHeapOop_not_null first.
321 Block *valb = cfg->_bbs[val->_idx];
322 if( this != valb && this->_dom_depth < valb->_dom_depth ) {
323 // Hoist it up to the end of the test block.
324 valb->find_remove(val);
325 this->add_inst(val);
326 cfg->_bbs.map(val->_idx,this);
327 // DecodeN on x86 may kill flags. Check for flag-killing projections
328 // that also need to be hoisted.
329 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
330 Node* n = val->fast_out(j);
331 if( n->is_MachProj() ) {
332 cfg->_bbs[n->_idx]->find_remove(n);
333 this->add_inst(n);
334 cfg->_bbs.map(n->_idx,this);
335 }
336 }
337 }
338 }
339 // Hoist the memory candidate up to the end of the test block.
340 Block *old_block = cfg->_bbs[best->_idx];
341 old_block->find_remove(best);
342 add_inst(best);
343 cfg->_bbs.map(best->_idx,this);
345 // Move the control dependence
346 if (best->in(0) && best->in(0) == old_block->_nodes[0])
347 best->set_req(0, _nodes[0]);
349 // Check for flag-killing projections that also need to be hoisted
350 // Should be DU safe because no edge updates.
351 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
352 Node* n = best->fast_out(j);
353 if( n->is_MachProj() ) {
354 cfg->_bbs[n->_idx]->find_remove(n);
355 add_inst(n);
356 cfg->_bbs.map(n->_idx,this);
357 }
358 }
360 Compile *C = cfg->C;
361 // proj==Op_True --> ne test; proj==Op_False --> eq test.
362 // One of two graph shapes got matched:
363 // (IfTrue (If (Bool NE (CmpP ptr NULL))))
364 // (IfFalse (If (Bool EQ (CmpP ptr NULL))))
365 // NULL checks are always branch-if-eq. If we see a IfTrue projection
366 // then we are replacing a 'ne' test with a 'eq' NULL check test.
367 // We need to flip the projections to keep the same semantics.
368 if( proj->Opcode() == Op_IfTrue ) {
369 // Swap order of projections in basic block to swap branch targets
370 Node *tmp1 = _nodes[end_idx()+1];
371 Node *tmp2 = _nodes[end_idx()+2];
372 _nodes.map(end_idx()+1, tmp2);
373 _nodes.map(end_idx()+2, tmp1);
374 Node *tmp = new (C) Node(C->top()); // Use not NULL input
375 tmp1->replace_by(tmp);
376 tmp2->replace_by(tmp1);
377 tmp->replace_by(tmp2);
378 tmp->destruct();
379 }
381 // Remove the existing null check; use a new implicit null check instead.
382 // Since schedule-local needs precise def-use info, we need to correct
383 // it as well.
384 Node *old_tst = proj->in(0);
385 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);
386 _nodes.map(end_idx(),nul_chk);
387 cfg->_bbs.map(nul_chk->_idx,this);
388 // Redirect users of old_test to nul_chk
389 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
390 old_tst->last_out(i2)->set_req(0, nul_chk);
391 // Clean-up any dead code
392 for (uint i3 = 0; i3 < old_tst->req(); i3++)
393 old_tst->set_req(i3, NULL);
395 cfg->latency_from_uses(nul_chk);
396 cfg->latency_from_uses(best);
397 }
400 //------------------------------select-----------------------------------------
401 // Select a nice fellow from the worklist to schedule next. If there is only
402 // one choice, then use it. Projections take top priority for correctness
403 // reasons - if I see a projection, then it is next. There are a number of
404 // other special cases, for instructions that consume condition codes, et al.
405 // These are chosen immediately. Some instructions are required to immediately
406 // precede the last instruction in the block, and these are taken last. Of the
407 // remaining cases (most), choose the instruction with the greatest latency
408 // (that is, the most number of pseudo-cycles required to the end of the
409 // routine). If there is a tie, choose the instruction with the most inputs.
410 Node *Block::select(PhaseCFG *cfg, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) {
412 // If only a single entry on the stack, use it
413 uint cnt = worklist.size();
414 if (cnt == 1) {
415 Node *n = worklist[0];
416 worklist.map(0,worklist.pop());
417 return n;
418 }
420 uint choice = 0; // Bigger is most important
421 uint latency = 0; // Bigger is scheduled first
422 uint score = 0; // Bigger is better
423 int idx = -1; // Index in worklist
425 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
426 // Order in worklist is used to break ties.
427 // See caller for how this is used to delay scheduling
428 // of induction variable increments to after the other
429 // uses of the phi are scheduled.
430 Node *n = worklist[i]; // Get Node on worklist
432 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
433 if( n->is_Proj() || // Projections always win
434 n->Opcode()== Op_Con || // So does constant 'Top'
435 iop == Op_CreateEx || // Create-exception must start block
436 iop == Op_CheckCastPP
437 ) {
438 worklist.map(i,worklist.pop());
439 return n;
440 }
442 // Final call in a block must be adjacent to 'catch'
443 Node *e = end();
444 if( e->is_Catch() && e->in(0)->in(0) == n )
445 continue;
447 // Memory op for an implicit null check has to be at the end of the block
448 if( e->is_MachNullCheck() && e->in(1) == n )
449 continue;
451 // Schedule IV increment last.
452 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd &&
453 e->in(1)->in(1) == n && n->is_iteratively_computed())
454 continue;
456 uint n_choice = 2;
458 // See if this instruction is consumed by a branch. If so, then (as the
459 // branch is the last instruction in the basic block) force it to the
460 // end of the basic block
461 if ( must_clone[iop] ) {
462 // See if any use is a branch
463 bool found_machif = false;
465 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
466 Node* use = n->fast_out(j);
468 // The use is a conditional branch, make them adjacent
469 if (use->is_MachIf() && cfg->_bbs[use->_idx]==this ) {
470 found_machif = true;
471 break;
472 }
474 // More than this instruction pending for successor to be ready,
475 // don't choose this if other opportunities are ready
476 if (ready_cnt.at(use->_idx) > 1)
477 n_choice = 1;
478 }
480 // loop terminated, prefer not to use this instruction
481 if (found_machif)
482 continue;
483 }
485 // See if this has a predecessor that is "must_clone", i.e. sets the
486 // condition code. If so, choose this first
487 for (uint j = 0; j < n->req() ; j++) {
488 Node *inn = n->in(j);
489 if (inn) {
490 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
491 n_choice = 3;
492 break;
493 }
494 }
495 }
497 // MachTemps should be scheduled last so they are near their uses
498 if (n->is_MachTemp()) {
499 n_choice = 1;
500 }
502 uint n_latency = cfg->_node_latency->at_grow(n->_idx);
503 uint n_score = n->req(); // Many inputs get high score to break ties
505 // Keep best latency found
506 if( choice < n_choice ||
507 ( choice == n_choice &&
508 ( latency < n_latency ||
509 ( latency == n_latency &&
510 ( score < n_score ))))) {
511 choice = n_choice;
512 latency = n_latency;
513 score = n_score;
514 idx = i; // Also keep index in worklist
515 }
516 } // End of for all ready nodes in worklist
518 assert(idx >= 0, "index should be set");
519 Node *n = worklist[(uint)idx]; // Get the winner
521 worklist.map((uint)idx, worklist.pop()); // Compress worklist
522 return n;
523 }
526 //------------------------------set_next_call----------------------------------
527 void Block::set_next_call( Node *n, VectorSet &next_call, Block_Array &bbs ) {
528 if( next_call.test_set(n->_idx) ) return;
529 for( uint i=0; i<n->len(); i++ ) {
530 Node *m = n->in(i);
531 if( !m ) continue; // must see all nodes in block that precede call
532 if( bbs[m->_idx] == this )
533 set_next_call( m, next_call, bbs );
534 }
535 }
537 //------------------------------needed_for_next_call---------------------------
538 // Set the flag 'next_call' for each Node that is needed for the next call to
539 // be scheduled. This flag lets me bias scheduling so Nodes needed for the
540 // next subroutine call get priority - basically it moves things NOT needed
541 // for the next call till after the call. This prevents me from trying to
542 // carry lots of stuff live across a call.
543 void Block::needed_for_next_call(Node *this_call, VectorSet &next_call, Block_Array &bbs) {
544 // Find the next control-defining Node in this block
545 Node* call = NULL;
546 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
547 Node* m = this_call->fast_out(i);
548 if( bbs[m->_idx] == this && // Local-block user
549 m != this_call && // Not self-start node
550 m->is_MachCall() )
551 call = m;
552 break;
553 }
554 if (call == NULL) return; // No next call (e.g., block end is near)
555 // Set next-call for all inputs to this call
556 set_next_call(call, next_call, bbs);
557 }
559 //------------------------------add_call_kills-------------------------------------
560 void Block::add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) {
561 // Fill in the kill mask for the call
562 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
563 if( !regs.Member(r) ) { // Not already defined by the call
564 // Save-on-call register?
565 if ((save_policy[r] == 'C') ||
566 (save_policy[r] == 'A') ||
567 ((save_policy[r] == 'E') && exclude_soe)) {
568 proj->_rout.Insert(r);
569 }
570 }
571 }
572 }
575 //------------------------------sched_call-------------------------------------
576 uint Block::sched_call( Matcher &matcher, Block_Array &bbs, uint node_cnt, Node_List &worklist, GrowableArray<int> &ready_cnt, MachCallNode *mcall, VectorSet &next_call ) {
577 RegMask regs;
579 // Schedule all the users of the call right now. All the users are
580 // projection Nodes, so they must be scheduled next to the call.
581 // Collect all the defined registers.
582 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
583 Node* n = mcall->fast_out(i);
584 assert( n->is_MachProj(), "" );
585 int n_cnt = ready_cnt.at(n->_idx)-1;
586 ready_cnt.at_put(n->_idx, n_cnt);
587 assert( n_cnt == 0, "" );
588 // Schedule next to call
589 _nodes.map(node_cnt++, n);
590 // Collect defined registers
591 regs.OR(n->out_RegMask());
592 // Check for scheduling the next control-definer
593 if( n->bottom_type() == Type::CONTROL )
594 // Warm up next pile of heuristic bits
595 needed_for_next_call(n, next_call, bbs);
597 // Children of projections are now all ready
598 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
599 Node* m = n->fast_out(j); // Get user
600 if( bbs[m->_idx] != this ) continue;
601 if( m->is_Phi() ) continue;
602 int m_cnt = ready_cnt.at(m->_idx)-1;
603 ready_cnt.at_put(m->_idx, m_cnt);
604 if( m_cnt == 0 )
605 worklist.push(m);
606 }
608 }
610 // Act as if the call defines the Frame Pointer.
611 // Certainly the FP is alive and well after the call.
612 regs.Insert(matcher.c_frame_pointer());
614 // Set all registers killed and not already defined by the call.
615 uint r_cnt = mcall->tf()->range()->cnt();
616 int op = mcall->ideal_Opcode();
617 MachProjNode *proj = new (matcher.C) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
618 bbs.map(proj->_idx,this);
619 _nodes.insert(node_cnt++, proj);
621 // Select the right register save policy.
622 const char * save_policy;
623 switch (op) {
624 case Op_CallRuntime:
625 case Op_CallLeaf:
626 case Op_CallLeafNoFP:
627 // Calling C code so use C calling convention
628 save_policy = matcher._c_reg_save_policy;
629 break;
631 case Op_CallStaticJava:
632 case Op_CallDynamicJava:
633 // Calling Java code so use Java calling convention
634 save_policy = matcher._register_save_policy;
635 break;
637 default:
638 ShouldNotReachHere();
639 }
641 // When using CallRuntime mark SOE registers as killed by the call
642 // so values that could show up in the RegisterMap aren't live in a
643 // callee saved register since the register wouldn't know where to
644 // find them. CallLeaf and CallLeafNoFP are ok because they can't
645 // have debug info on them. Strictly speaking this only needs to be
646 // done for oops since idealreg2debugmask takes care of debug info
647 // references but there no way to handle oops differently than other
648 // pointers as far as the kill mask goes.
649 bool exclude_soe = op == Op_CallRuntime;
651 // If the call is a MethodHandle invoke, we need to exclude the
652 // register which is used to save the SP value over MH invokes from
653 // the mask. Otherwise this register could be used for
654 // deoptimization information.
655 if (op == Op_CallStaticJava) {
656 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall;
657 if (mcallstaticjava->_method_handle_invoke)
658 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
659 }
661 add_call_kills(proj, regs, save_policy, exclude_soe);
663 return node_cnt;
664 }
667 //------------------------------schedule_local---------------------------------
668 // Topological sort within a block. Someday become a real scheduler.
669 bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, GrowableArray<int> &ready_cnt, VectorSet &next_call) {
670 // Already "sorted" are the block start Node (as the first entry), and
671 // the block-ending Node and any trailing control projections. We leave
672 // these alone. PhiNodes and ParmNodes are made to follow the block start
673 // Node. Everything else gets topo-sorted.
675 #ifndef PRODUCT
676 if (cfg->trace_opto_pipelining()) {
677 tty->print_cr("# --- schedule_local B%d, before: ---", _pre_order);
678 for (uint i = 0;i < _nodes.size();i++) {
679 tty->print("# ");
680 _nodes[i]->fast_dump();
681 }
682 tty->print_cr("#");
683 }
684 #endif
686 // RootNode is already sorted
687 if( _nodes.size() == 1 ) return true;
689 // Move PhiNodes and ParmNodes from 1 to cnt up to the start
690 uint node_cnt = end_idx();
691 uint phi_cnt = 1;
692 uint i;
693 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
694 Node *n = _nodes[i];
695 if( n->is_Phi() || // Found a PhiNode or ParmNode
696 (n->is_Proj() && n->in(0) == head()) ) {
697 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
698 _nodes.map(i,_nodes[phi_cnt]);
699 _nodes.map(phi_cnt++,n); // swap Phi/Parm up front
700 } else { // All others
701 // Count block-local inputs to 'n'
702 uint cnt = n->len(); // Input count
703 uint local = 0;
704 for( uint j=0; j<cnt; j++ ) {
705 Node *m = n->in(j);
706 if( m && cfg->_bbs[m->_idx] == this && !m->is_top() )
707 local++; // One more block-local input
708 }
709 ready_cnt.at_put(n->_idx, local); // Count em up
711 #ifdef ASSERT
712 if( UseConcMarkSweepGC || UseG1GC ) {
713 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
714 // Check the precedence edges
715 for (uint prec = n->req(); prec < n->len(); prec++) {
716 Node* oop_store = n->in(prec);
717 if (oop_store != NULL) {
718 assert(cfg->_bbs[oop_store->_idx]->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark");
719 }
720 }
721 }
722 }
723 #endif
725 // A few node types require changing a required edge to a precedence edge
726 // before allocation.
727 if( n->is_Mach() && n->req() > TypeFunc::Parms &&
728 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
729 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
730 // MemBarAcquire could be created without Precedent edge.
731 // del_req() replaces the specified edge with the last input edge
732 // and then removes the last edge. If the specified edge > number of
733 // edges the last edge will be moved outside of the input edges array
734 // and the edge will be lost. This is why this code should be
735 // executed only when Precedent (== TypeFunc::Parms) edge is present.
736 Node *x = n->in(TypeFunc::Parms);
737 n->del_req(TypeFunc::Parms);
738 n->add_prec(x);
739 }
740 }
741 }
742 for(uint i2=i; i2<_nodes.size(); i2++ ) // Trailing guys get zapped count
743 ready_cnt.at_put(_nodes[i2]->_idx, 0);
745 // All the prescheduled guys do not hold back internal nodes
746 uint i3;
747 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled
748 Node *n = _nodes[i3]; // Get pre-scheduled
749 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
750 Node* m = n->fast_out(j);
751 if( cfg->_bbs[m->_idx] ==this ) { // Local-block user
752 int m_cnt = ready_cnt.at(m->_idx)-1;
753 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count
754 }
755 }
756 }
758 Node_List delay;
759 // Make a worklist
760 Node_List worklist;
761 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
762 Node *m = _nodes[i4];
763 if( !ready_cnt.at(m->_idx) ) { // Zero ready count?
764 if (m->is_iteratively_computed()) {
765 // Push induction variable increments last to allow other uses
766 // of the phi to be scheduled first. The select() method breaks
767 // ties in scheduling by worklist order.
768 delay.push(m);
769 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
770 // Force the CreateEx to the top of the list so it's processed
771 // first and ends up at the start of the block.
772 worklist.insert(0, m);
773 } else {
774 worklist.push(m); // Then on to worklist!
775 }
776 }
777 }
778 while (delay.size()) {
779 Node* d = delay.pop();
780 worklist.push(d);
781 }
783 // Warm up the 'next_call' heuristic bits
784 needed_for_next_call(_nodes[0], next_call, cfg->_bbs);
786 #ifndef PRODUCT
787 if (cfg->trace_opto_pipelining()) {
788 for (uint j=0; j<_nodes.size(); j++) {
789 Node *n = _nodes[j];
790 int idx = n->_idx;
791 tty->print("# ready cnt:%3d ", ready_cnt.at(idx));
792 tty->print("latency:%3d ", cfg->_node_latency->at_grow(idx));
793 tty->print("%4d: %s\n", idx, n->Name());
794 }
795 }
796 #endif
798 uint max_idx = (uint)ready_cnt.length();
799 // Pull from worklist and schedule
800 while( worklist.size() ) { // Worklist is not ready
802 #ifndef PRODUCT
803 if (cfg->trace_opto_pipelining()) {
804 tty->print("# ready list:");
805 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
806 Node *n = worklist[i]; // Get Node on worklist
807 tty->print(" %d", n->_idx);
808 }
809 tty->cr();
810 }
811 #endif
813 // Select and pop a ready guy from worklist
814 Node* n = select(cfg, worklist, ready_cnt, next_call, phi_cnt);
815 _nodes.map(phi_cnt++,n); // Schedule him next
817 #ifndef PRODUCT
818 if (cfg->trace_opto_pipelining()) {
819 tty->print("# select %d: %s", n->_idx, n->Name());
820 tty->print(", latency:%d", cfg->_node_latency->at_grow(n->_idx));
821 n->dump();
822 if (Verbose) {
823 tty->print("# ready list:");
824 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
825 Node *n = worklist[i]; // Get Node on worklist
826 tty->print(" %d", n->_idx);
827 }
828 tty->cr();
829 }
830 }
832 #endif
833 if( n->is_MachCall() ) {
834 MachCallNode *mcall = n->as_MachCall();
835 phi_cnt = sched_call(matcher, cfg->_bbs, phi_cnt, worklist, ready_cnt, mcall, next_call);
836 continue;
837 }
839 if (n->is_Mach() && n->as_Mach()->has_call()) {
840 RegMask regs;
841 regs.Insert(matcher.c_frame_pointer());
842 regs.OR(n->out_RegMask());
844 MachProjNode *proj = new (matcher.C) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj );
845 cfg->_bbs.map(proj->_idx,this);
846 _nodes.insert(phi_cnt++, proj);
848 add_call_kills(proj, regs, matcher._c_reg_save_policy, false);
849 }
851 // Children are now all ready
852 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
853 Node* m = n->fast_out(i5); // Get user
854 if( cfg->_bbs[m->_idx] != this ) continue;
855 if( m->is_Phi() ) continue;
856 if (m->_idx >= max_idx) { // new node, skip it
857 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types");
858 continue;
859 }
860 int m_cnt = ready_cnt.at(m->_idx)-1;
861 ready_cnt.at_put(m->_idx, m_cnt);
862 if( m_cnt == 0 )
863 worklist.push(m);
864 }
865 }
867 if( phi_cnt != end_idx() ) {
868 // did not schedule all. Retry, Bailout, or Die
869 Compile* C = matcher.C;
870 if (C->subsume_loads() == true && !C->failing()) {
871 // Retry with subsume_loads == false
872 // If this is the first failure, the sentinel string will "stick"
873 // to the Compile object, and the C2Compiler will see it and retry.
874 C->record_failure(C2Compiler::retry_no_subsuming_loads());
875 }
876 // assert( phi_cnt == end_idx(), "did not schedule all" );
877 return false;
878 }
880 #ifndef PRODUCT
881 if (cfg->trace_opto_pipelining()) {
882 tty->print_cr("#");
883 tty->print_cr("# after schedule_local");
884 for (uint i = 0;i < _nodes.size();i++) {
885 tty->print("# ");
886 _nodes[i]->fast_dump();
887 }
888 tty->cr();
889 }
890 #endif
893 return true;
894 }
896 //--------------------------catch_cleanup_fix_all_inputs-----------------------
897 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
898 for (uint l = 0; l < use->len(); l++) {
899 if (use->in(l) == old_def) {
900 if (l < use->req()) {
901 use->set_req(l, new_def);
902 } else {
903 use->rm_prec(l);
904 use->add_prec(new_def);
905 l--;
906 }
907 }
908 }
909 }
911 //------------------------------catch_cleanup_find_cloned_def------------------
912 static Node *catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {
913 assert( use_blk != def_blk, "Inter-block cleanup only");
915 // The use is some block below the Catch. Find and return the clone of the def
916 // that dominates the use. If there is no clone in a dominating block, then
917 // create a phi for the def in a dominating block.
919 // Find which successor block dominates this use. The successor
920 // blocks must all be single-entry (from the Catch only; I will have
921 // split blocks to make this so), hence they all dominate.
922 while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
923 use_blk = use_blk->_idom;
925 // Find the successor
926 Node *fixup = NULL;
928 uint j;
929 for( j = 0; j < def_blk->_num_succs; j++ )
930 if( use_blk == def_blk->_succs[j] )
931 break;
933 if( j == def_blk->_num_succs ) {
934 // Block at same level in dom-tree is not a successor. It needs a
935 // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
936 Node_Array inputs = new Node_List(Thread::current()->resource_area());
937 for(uint k = 1; k < use_blk->num_preds(); k++) {
938 inputs.map(k, catch_cleanup_find_cloned_def(bbs[use_blk->pred(k)->_idx], def, def_blk, bbs, n_clone_idx));
939 }
941 // Check to see if the use_blk already has an identical phi inserted.
942 // If it exists, it will be at the first position since all uses of a
943 // def are processed together.
944 Node *phi = use_blk->_nodes[1];
945 if( phi->is_Phi() ) {
946 fixup = phi;
947 for (uint k = 1; k < use_blk->num_preds(); k++) {
948 if (phi->in(k) != inputs[k]) {
949 // Not a match
950 fixup = NULL;
951 break;
952 }
953 }
954 }
956 // If an existing PhiNode was not found, make a new one.
957 if (fixup == NULL) {
958 Node *new_phi = PhiNode::make(use_blk->head(), def);
959 use_blk->_nodes.insert(1, new_phi);
960 bbs.map(new_phi->_idx, use_blk);
961 for (uint k = 1; k < use_blk->num_preds(); k++) {
962 new_phi->set_req(k, inputs[k]);
963 }
964 fixup = new_phi;
965 }
967 } else {
968 // Found the use just below the Catch. Make it use the clone.
969 fixup = use_blk->_nodes[n_clone_idx];
970 }
972 return fixup;
973 }
975 //--------------------------catch_cleanup_intra_block--------------------------
976 // Fix all input edges in use that reference "def". The use is in the same
977 // block as the def and both have been cloned in each successor block.
978 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
980 // Both the use and def have been cloned. For each successor block,
981 // get the clone of the use, and make its input the clone of the def
982 // found in that block.
984 uint use_idx = blk->find_node(use);
985 uint offset_idx = use_idx - beg;
986 for( uint k = 0; k < blk->_num_succs; k++ ) {
987 // Get clone in each successor block
988 Block *sb = blk->_succs[k];
989 Node *clone = sb->_nodes[offset_idx+1];
990 assert( clone->Opcode() == use->Opcode(), "" );
992 // Make use-clone reference the def-clone
993 catch_cleanup_fix_all_inputs(clone, def, sb->_nodes[n_clone_idx]);
994 }
995 }
997 //------------------------------catch_cleanup_inter_block---------------------
998 // Fix all input edges in use that reference "def". The use is in a different
999 // block than the def.
1000 static void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {
1001 if( !use_blk ) return; // Can happen if the use is a precedence edge
1003 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, bbs, n_clone_idx);
1004 catch_cleanup_fix_all_inputs(use, def, new_def);
1005 }
1007 //------------------------------call_catch_cleanup-----------------------------
1008 // If we inserted any instructions between a Call and his CatchNode,
1009 // clone the instructions on all paths below the Catch.
1010 void Block::call_catch_cleanup(Block_Array &bbs, Compile* C) {
1012 // End of region to clone
1013 uint end = end_idx();
1014 if( !_nodes[end]->is_Catch() ) return;
1015 // Start of region to clone
1016 uint beg = end;
1017 while(!_nodes[beg-1]->is_MachProj() ||
1018 !_nodes[beg-1]->in(0)->is_MachCall() ) {
1019 beg--;
1020 assert(beg > 0,"Catch cleanup walking beyond block boundary");
1021 }
1022 // Range of inserted instructions is [beg, end)
1023 if( beg == end ) return;
1025 // Clone along all Catch output paths. Clone area between the 'beg' and
1026 // 'end' indices.
1027 for( uint i = 0; i < _num_succs; i++ ) {
1028 Block *sb = _succs[i];
1029 // Clone the entire area; ignoring the edge fixup for now.
1030 for( uint j = end; j > beg; j-- ) {
1031 // It is safe here to clone a node with anti_dependence
1032 // since clones dominate on each path.
1033 Node *clone = _nodes[j-1]->clone();
1034 sb->_nodes.insert( 1, clone );
1035 bbs.map(clone->_idx,sb);
1036 }
1037 }
1040 // Fixup edges. Check the def-use info per cloned Node
1041 for(uint i2 = beg; i2 < end; i2++ ) {
1042 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
1043 Node *n = _nodes[i2]; // Node that got cloned
1044 // Need DU safe iterator because of edge manipulation in calls.
1045 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
1046 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
1047 out->push(n->fast_out(j1));
1048 }
1049 uint max = out->size();
1050 for (uint j = 0; j < max; j++) {// For all users
1051 Node *use = out->pop();
1052 Block *buse = bbs[use->_idx];
1053 if( use->is_Phi() ) {
1054 for( uint k = 1; k < use->req(); k++ )
1055 if( use->in(k) == n ) {
1056 Node *fixup = catch_cleanup_find_cloned_def(bbs[buse->pred(k)->_idx], n, this, bbs, n_clone_idx);
1057 use->set_req(k, fixup);
1058 }
1059 } else {
1060 if (this == buse) {
1061 catch_cleanup_intra_block(use, n, this, beg, n_clone_idx);
1062 } else {
1063 catch_cleanup_inter_block(use, buse, n, this, bbs, n_clone_idx);
1064 }
1065 }
1066 } // End for all users
1068 } // End of for all Nodes in cloned area
1070 // Remove the now-dead cloned ops
1071 for(uint i3 = beg; i3 < end; i3++ ) {
1072 _nodes[beg]->disconnect_inputs(NULL, C);
1073 _nodes.remove(beg);
1074 }
1076 // If the successor blocks have a CreateEx node, move it back to the top
1077 for(uint i4 = 0; i4 < _num_succs; i4++ ) {
1078 Block *sb = _succs[i4];
1079 uint new_cnt = end - beg;
1080 // Remove any newly created, but dead, nodes.
1081 for( uint j = new_cnt; j > 0; j-- ) {
1082 Node *n = sb->_nodes[j];
1083 if (n->outcnt() == 0 &&
1084 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
1085 n->disconnect_inputs(NULL, C);
1086 sb->_nodes.remove(j);
1087 new_cnt--;
1088 }
1089 }
1090 // If any newly created nodes remain, move the CreateEx node to the top
1091 if (new_cnt > 0) {
1092 Node *cex = sb->_nodes[1+new_cnt];
1093 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
1094 sb->_nodes.remove(1+new_cnt);
1095 sb->_nodes.insert(1,cex);
1096 }
1097 }
1098 }
1099 }