Fri, 04 Mar 2016 16:15:48 +0300
8130150: Implement BigInteger.montgomeryMultiply intrinsic
Reviewed-by: kvn, mdoerr
1 /*
2 * Copyright (c) 1998, 2015, 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 #if defined AD_MD_HPP
34 # include AD_MD_HPP
35 #elif defined TARGET_ARCH_MODEL_x86_32
36 # include "adfiles/ad_x86_32.hpp"
37 #elif defined TARGET_ARCH_MODEL_x86_64
38 # include "adfiles/ad_x86_64.hpp"
39 #elif defined TARGET_ARCH_MODEL_sparc
40 # include "adfiles/ad_sparc.hpp"
41 #elif defined TARGET_ARCH_MODEL_zero
42 # include "adfiles/ad_zero.hpp"
43 #elif defined TARGET_ARCH_MODEL_ppc_64
44 # include "adfiles/ad_ppc_64.hpp"
45 #endif
47 // Optimization - Graph Style
49 // Check whether val is not-null-decoded compressed oop,
50 // i.e. will grab into the base of the heap if it represents NULL.
51 static bool accesses_heap_base_zone(Node *val) {
52 if (Universe::narrow_oop_base() > 0) { // Implies UseCompressedOops.
53 if (val && val->is_Mach()) {
54 if (val->as_Mach()->ideal_Opcode() == Op_DecodeN) {
55 // This assumes all Decodes with TypePtr::NotNull are matched to nodes that
56 // decode NULL to point to the heap base (Decode_NN).
57 if (val->bottom_type()->is_oopptr()->ptr() == TypePtr::NotNull) {
58 return true;
59 }
60 }
61 // Must recognize load operation with Decode matched in memory operand.
62 // We should not reach here exept for PPC/AIX, as os::zero_page_read_protected()
63 // returns true everywhere else. On PPC, no such memory operands
64 // exist, therefore we did not yet implement a check for such operands.
65 NOT_AIX(Unimplemented());
66 }
67 }
68 return false;
69 }
71 static bool needs_explicit_null_check_for_read(Node *val) {
72 // On some OSes (AIX) the page at address 0 is only write protected.
73 // If so, only Store operations will trap.
74 if (os::zero_page_read_protected()) {
75 return false; // Implicit null check will work.
76 }
77 // Also a read accessing the base of a heap-based compressed heap will trap.
78 if (accesses_heap_base_zone(val) && // Hits the base zone page.
79 Universe::narrow_oop_use_implicit_null_checks()) { // Base zone page is protected.
80 return false;
81 }
83 return true;
84 }
86 //------------------------------implicit_null_check----------------------------
87 // Detect implicit-null-check opportunities. Basically, find NULL checks
88 // with suitable memory ops nearby. Use the memory op to do the NULL check.
89 // I can generate a memory op if there is not one nearby.
90 // The proj is the control projection for the not-null case.
91 // The val is the pointer being checked for nullness or
92 // decodeHeapOop_not_null node if it did not fold into address.
93 void PhaseCFG::implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons) {
94 // Assume if null check need for 0 offset then always needed
95 // Intel solaris doesn't support any null checks yet and no
96 // mechanism exists (yet) to set the switches at an os_cpu level
97 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;
99 // Make sure the ptr-is-null path appears to be uncommon!
100 float f = block->end()->as_MachIf()->_prob;
101 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
102 if( f > PROB_UNLIKELY_MAG(4) ) return;
104 uint bidx = 0; // Capture index of value into memop
105 bool was_store; // Memory op is a store op
107 // Get the successor block for if the test ptr is non-null
108 Block* not_null_block; // this one goes with the proj
109 Block* null_block;
110 if (block->get_node(block->number_of_nodes()-1) == proj) {
111 null_block = block->_succs[0];
112 not_null_block = block->_succs[1];
113 } else {
114 assert(block->get_node(block->number_of_nodes()-2) == proj, "proj is one or the other");
115 not_null_block = block->_succs[0];
116 null_block = block->_succs[1];
117 }
118 while (null_block->is_Empty() == Block::empty_with_goto) {
119 null_block = null_block->_succs[0];
120 }
122 // Search the exception block for an uncommon trap.
123 // (See Parse::do_if and Parse::do_ifnull for the reason
124 // we need an uncommon trap. Briefly, we need a way to
125 // detect failure of this optimization, as in 6366351.)
126 {
127 bool found_trap = false;
128 for (uint i1 = 0; i1 < null_block->number_of_nodes(); i1++) {
129 Node* nn = null_block->get_node(i1);
130 if (nn->is_MachCall() &&
131 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) {
132 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();
133 if (trtype->isa_int() && trtype->is_int()->is_con()) {
134 jint tr_con = trtype->is_int()->get_con();
135 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
136 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
137 assert((int)reason < (int)BitsPerInt, "recode bit map");
138 if (is_set_nth_bit(allowed_reasons, (int) reason)
139 && action != Deoptimization::Action_none) {
140 // This uncommon trap is sure to recompile, eventually.
141 // When that happens, C->too_many_traps will prevent
142 // this transformation from happening again.
143 found_trap = true;
144 }
145 }
146 break;
147 }
148 }
149 if (!found_trap) {
150 // We did not find an uncommon trap.
151 return;
152 }
153 }
155 // Check for decodeHeapOop_not_null node which did not fold into address
156 bool is_decoden = ((intptr_t)val) & 1;
157 val = (Node*)(((intptr_t)val) & ~1);
159 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() &&
160 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity");
162 // Search the successor block for a load or store who's base value is also
163 // the tested value. There may be several.
164 Node_List *out = new Node_List(Thread::current()->resource_area());
165 MachNode *best = NULL; // Best found so far
166 for (DUIterator i = val->outs(); val->has_out(i); i++) {
167 Node *m = val->out(i);
168 if( !m->is_Mach() ) continue;
169 MachNode *mach = m->as_Mach();
170 was_store = false;
171 int iop = mach->ideal_Opcode();
172 switch( iop ) {
173 case Op_LoadB:
174 case Op_LoadUB:
175 case Op_LoadUS:
176 case Op_LoadD:
177 case Op_LoadF:
178 case Op_LoadI:
179 case Op_LoadL:
180 case Op_LoadP:
181 case Op_LoadN:
182 case Op_LoadS:
183 case Op_LoadKlass:
184 case Op_LoadNKlass:
185 case Op_LoadRange:
186 case Op_LoadD_unaligned:
187 case Op_LoadL_unaligned:
188 assert(mach->in(2) == val, "should be address");
189 break;
190 case Op_StoreB:
191 case Op_StoreC:
192 case Op_StoreCM:
193 case Op_StoreD:
194 case Op_StoreF:
195 case Op_StoreI:
196 case Op_StoreL:
197 case Op_StoreP:
198 case Op_StoreN:
199 case Op_StoreNKlass:
200 was_store = true; // Memory op is a store op
201 // Stores will have their address in slot 2 (memory in slot 1).
202 // If the value being nul-checked is in another slot, it means we
203 // are storing the checked value, which does NOT check the value!
204 if( mach->in(2) != val ) continue;
205 break; // Found a memory op?
206 case Op_StrComp:
207 case Op_StrEquals:
208 case Op_StrIndexOf:
209 case Op_AryEq:
210 case Op_EncodeISOArray:
211 // Not a legit memory op for implicit null check regardless of
212 // embedded loads
213 continue;
214 default: // Also check for embedded loads
215 if( !mach->needs_anti_dependence_check() )
216 continue; // Not an memory op; skip it
217 if( must_clone[iop] ) {
218 // Do not move nodes which produce flags because
219 // RA will try to clone it to place near branch and
220 // it will cause recompilation, see clone_node().
221 continue;
222 }
223 {
224 // Check that value is used in memory address in
225 // instructions with embedded load (CmpP val1,(val2+off)).
226 Node* base;
227 Node* index;
228 const MachOper* oper = mach->memory_inputs(base, index);
229 if (oper == NULL || oper == (MachOper*)-1) {
230 continue; // Not an memory op; skip it
231 }
232 if (val == base ||
233 val == index && val->bottom_type()->isa_narrowoop()) {
234 break; // Found it
235 } else {
236 continue; // Skip it
237 }
238 }
239 break;
240 }
242 // On some OSes (AIX) the page at address 0 is only write protected.
243 // If so, only Store operations will trap.
244 // But a read accessing the base of a heap-based compressed heap will trap.
245 if (!was_store && needs_explicit_null_check_for_read(val)) {
246 continue;
247 }
249 // check if the offset is not too high for implicit exception
250 {
251 intptr_t offset = 0;
252 const TypePtr *adr_type = NULL; // Do not need this return value here
253 const Node* base = mach->get_base_and_disp(offset, adr_type);
254 if (base == NULL || base == NodeSentinel) {
255 // Narrow oop address doesn't have base, only index
256 if( val->bottom_type()->isa_narrowoop() &&
257 MacroAssembler::needs_explicit_null_check(offset) )
258 continue; // Give up if offset is beyond page size
259 // cannot reason about it; is probably not implicit null exception
260 } else {
261 const TypePtr* tptr;
262 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 ||
263 Universe::narrow_klass_shift() == 0)) {
264 // 32-bits narrow oop can be the base of address expressions
265 tptr = base->get_ptr_type();
266 } else {
267 // only regular oops are expected here
268 tptr = base->bottom_type()->is_ptr();
269 }
270 // Give up if offset is not a compile-time constant
271 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
272 continue;
273 offset += tptr->_offset; // correct if base is offseted
274 if( MacroAssembler::needs_explicit_null_check(offset) )
275 continue; // Give up is reference is beyond 4K page size
276 }
277 }
279 // Check ctrl input to see if the null-check dominates the memory op
280 Block *cb = get_block_for_node(mach);
281 cb = cb->_idom; // Always hoist at least 1 block
282 if( !was_store ) { // Stores can be hoisted only one block
283 while( cb->_dom_depth > (block->_dom_depth + 1))
284 cb = cb->_idom; // Hoist loads as far as we want
285 // The non-null-block should dominate the memory op, too. Live
286 // range spilling will insert a spill in the non-null-block if it is
287 // needs to spill the memory op for an implicit null check.
288 if (cb->_dom_depth == (block->_dom_depth + 1)) {
289 if (cb != not_null_block) continue;
290 cb = cb->_idom;
291 }
292 }
293 if( cb != block ) continue;
295 // Found a memory user; see if it can be hoisted to check-block
296 uint vidx = 0; // Capture index of value into memop
297 uint j;
298 for( j = mach->req()-1; j > 0; j-- ) {
299 if( mach->in(j) == val ) {
300 vidx = j;
301 // Ignore DecodeN val which could be hoisted to where needed.
302 if( is_decoden ) continue;
303 }
304 // Block of memory-op input
305 Block *inb = get_block_for_node(mach->in(j));
306 Block *b = block; // Start from nul check
307 while( b != inb && b->_dom_depth > inb->_dom_depth )
308 b = b->_idom; // search upwards for input
309 // See if input dominates null check
310 if( b != inb )
311 break;
312 }
313 if( j > 0 )
314 continue;
315 Block *mb = get_block_for_node(mach);
316 // Hoisting stores requires more checks for the anti-dependence case.
317 // Give up hoisting if we have to move the store past any load.
318 if( was_store ) {
319 Block *b = mb; // Start searching here for a local load
320 // mach use (faulting) trying to hoist
321 // n might be blocker to hoisting
322 while( b != block ) {
323 uint k;
324 for( k = 1; k < b->number_of_nodes(); k++ ) {
325 Node *n = b->get_node(k);
326 if( n->needs_anti_dependence_check() &&
327 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
328 break; // Found anti-dependent load
329 }
330 if( k < b->number_of_nodes() )
331 break; // Found anti-dependent load
332 // Make sure control does not do a merge (would have to check allpaths)
333 if( b->num_preds() != 2 ) break;
334 b = get_block_for_node(b->pred(1)); // Move up to predecessor block
335 }
336 if( b != block ) continue;
337 }
339 // Make sure this memory op is not already being used for a NullCheck
340 Node *e = mb->end();
341 if( e->is_MachNullCheck() && e->in(1) == mach )
342 continue; // Already being used as a NULL check
344 // Found a candidate! Pick one with least dom depth - the highest
345 // in the dom tree should be closest to the null check.
346 if (best == NULL || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) {
347 best = mach;
348 bidx = vidx;
349 }
350 }
351 // No candidate!
352 if (best == NULL) {
353 return;
354 }
356 // ---- Found an implicit null check
357 extern int implicit_null_checks;
358 implicit_null_checks++;
360 if( is_decoden ) {
361 // Check if we need to hoist decodeHeapOop_not_null first.
362 Block *valb = get_block_for_node(val);
363 if( block != valb && block->_dom_depth < valb->_dom_depth ) {
364 // Hoist it up to the end of the test block.
365 valb->find_remove(val);
366 block->add_inst(val);
367 map_node_to_block(val, block);
368 // DecodeN on x86 may kill flags. Check for flag-killing projections
369 // that also need to be hoisted.
370 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
371 Node* n = val->fast_out(j);
372 if( n->is_MachProj() ) {
373 get_block_for_node(n)->find_remove(n);
374 block->add_inst(n);
375 map_node_to_block(n, block);
376 }
377 }
378 }
379 }
380 // Hoist the memory candidate up to the end of the test block.
381 Block *old_block = get_block_for_node(best);
382 old_block->find_remove(best);
383 block->add_inst(best);
384 map_node_to_block(best, block);
386 // Move the control dependence
387 if (best->in(0) && best->in(0) == old_block->head())
388 best->set_req(0, block->head());
390 // Check for flag-killing projections that also need to be hoisted
391 // Should be DU safe because no edge updates.
392 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
393 Node* n = best->fast_out(j);
394 if( n->is_MachProj() ) {
395 get_block_for_node(n)->find_remove(n);
396 block->add_inst(n);
397 map_node_to_block(n, block);
398 }
399 }
401 // proj==Op_True --> ne test; proj==Op_False --> eq test.
402 // One of two graph shapes got matched:
403 // (IfTrue (If (Bool NE (CmpP ptr NULL))))
404 // (IfFalse (If (Bool EQ (CmpP ptr NULL))))
405 // NULL checks are always branch-if-eq. If we see a IfTrue projection
406 // then we are replacing a 'ne' test with a 'eq' NULL check test.
407 // We need to flip the projections to keep the same semantics.
408 if( proj->Opcode() == Op_IfTrue ) {
409 // Swap order of projections in basic block to swap branch targets
410 Node *tmp1 = block->get_node(block->end_idx()+1);
411 Node *tmp2 = block->get_node(block->end_idx()+2);
412 block->map_node(tmp2, block->end_idx()+1);
413 block->map_node(tmp1, block->end_idx()+2);
414 Node *tmp = new (C) Node(C->top()); // Use not NULL input
415 tmp1->replace_by(tmp);
416 tmp2->replace_by(tmp1);
417 tmp->replace_by(tmp2);
418 tmp->destruct();
419 }
421 // Remove the existing null check; use a new implicit null check instead.
422 // Since schedule-local needs precise def-use info, we need to correct
423 // it as well.
424 Node *old_tst = proj->in(0);
425 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);
426 block->map_node(nul_chk, block->end_idx());
427 map_node_to_block(nul_chk, block);
428 // Redirect users of old_test to nul_chk
429 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
430 old_tst->last_out(i2)->set_req(0, nul_chk);
431 // Clean-up any dead code
432 for (uint i3 = 0; i3 < old_tst->req(); i3++) {
433 Node* in = old_tst->in(i3);
434 old_tst->set_req(i3, NULL);
435 if (in->outcnt() == 0) {
436 // Remove dead input node
437 in->disconnect_inputs(NULL, C);
438 block->find_remove(in);
439 }
440 }
442 latency_from_uses(nul_chk);
443 latency_from_uses(best);
444 }
447 //------------------------------select-----------------------------------------
448 // Select a nice fellow from the worklist to schedule next. If there is only
449 // one choice, then use it. Projections take top priority for correctness
450 // reasons - if I see a projection, then it is next. There are a number of
451 // other special cases, for instructions that consume condition codes, et al.
452 // These are chosen immediately. Some instructions are required to immediately
453 // precede the last instruction in the block, and these are taken last. Of the
454 // remaining cases (most), choose the instruction with the greatest latency
455 // (that is, the most number of pseudo-cycles required to the end of the
456 // routine). If there is a tie, choose the instruction with the most inputs.
457 Node* PhaseCFG::select(Block* block, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) {
459 // If only a single entry on the stack, use it
460 uint cnt = worklist.size();
461 if (cnt == 1) {
462 Node *n = worklist[0];
463 worklist.map(0,worklist.pop());
464 return n;
465 }
467 uint choice = 0; // Bigger is most important
468 uint latency = 0; // Bigger is scheduled first
469 uint score = 0; // Bigger is better
470 int idx = -1; // Index in worklist
471 int cand_cnt = 0; // Candidate count
473 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
474 // Order in worklist is used to break ties.
475 // See caller for how this is used to delay scheduling
476 // of induction variable increments to after the other
477 // uses of the phi are scheduled.
478 Node *n = worklist[i]; // Get Node on worklist
480 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
481 if( n->is_Proj() || // Projections always win
482 n->Opcode()== Op_Con || // So does constant 'Top'
483 iop == Op_CreateEx || // Create-exception must start block
484 iop == Op_CheckCastPP
485 ) {
486 worklist.map(i,worklist.pop());
487 return n;
488 }
490 // Final call in a block must be adjacent to 'catch'
491 Node *e = block->end();
492 if( e->is_Catch() && e->in(0)->in(0) == n )
493 continue;
495 // Memory op for an implicit null check has to be at the end of the block
496 if( e->is_MachNullCheck() && e->in(1) == n )
497 continue;
499 // Schedule IV increment last.
500 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd &&
501 e->in(1)->in(1) == n && n->is_iteratively_computed())
502 continue;
504 uint n_choice = 2;
506 // See if this instruction is consumed by a branch. If so, then (as the
507 // branch is the last instruction in the basic block) force it to the
508 // end of the basic block
509 if ( must_clone[iop] ) {
510 // See if any use is a branch
511 bool found_machif = false;
513 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
514 Node* use = n->fast_out(j);
516 // The use is a conditional branch, make them adjacent
517 if (use->is_MachIf() && get_block_for_node(use) == block) {
518 found_machif = true;
519 break;
520 }
522 // More than this instruction pending for successor to be ready,
523 // don't choose this if other opportunities are ready
524 if (ready_cnt.at(use->_idx) > 1)
525 n_choice = 1;
526 }
528 // loop terminated, prefer not to use this instruction
529 if (found_machif)
530 continue;
531 }
533 // See if this has a predecessor that is "must_clone", i.e. sets the
534 // condition code. If so, choose this first
535 for (uint j = 0; j < n->req() ; j++) {
536 Node *inn = n->in(j);
537 if (inn) {
538 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
539 n_choice = 3;
540 break;
541 }
542 }
543 }
545 // MachTemps should be scheduled last so they are near their uses
546 if (n->is_MachTemp()) {
547 n_choice = 1;
548 }
550 uint n_latency = get_latency_for_node(n);
551 uint n_score = n->req(); // Many inputs get high score to break ties
553 // Keep best latency found
554 cand_cnt++;
555 if (choice < n_choice ||
556 (choice == n_choice &&
557 ((StressLCM && Compile::randomized_select(cand_cnt)) ||
558 (!StressLCM &&
559 (latency < n_latency ||
560 (latency == n_latency &&
561 (score < n_score))))))) {
562 choice = n_choice;
563 latency = n_latency;
564 score = n_score;
565 idx = i; // Also keep index in worklist
566 }
567 } // End of for all ready nodes in worklist
569 assert(idx >= 0, "index should be set");
570 Node *n = worklist[(uint)idx]; // Get the winner
572 worklist.map((uint)idx, worklist.pop()); // Compress worklist
573 return n;
574 }
577 //------------------------------set_next_call----------------------------------
578 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) {
579 if( next_call.test_set(n->_idx) ) return;
580 for( uint i=0; i<n->len(); i++ ) {
581 Node *m = n->in(i);
582 if( !m ) continue; // must see all nodes in block that precede call
583 if (get_block_for_node(m) == block) {
584 set_next_call(block, m, next_call);
585 }
586 }
587 }
589 //------------------------------needed_for_next_call---------------------------
590 // Set the flag 'next_call' for each Node that is needed for the next call to
591 // be scheduled. This flag lets me bias scheduling so Nodes needed for the
592 // next subroutine call get priority - basically it moves things NOT needed
593 // for the next call till after the call. This prevents me from trying to
594 // carry lots of stuff live across a call.
595 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) {
596 // Find the next control-defining Node in this block
597 Node* call = NULL;
598 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
599 Node* m = this_call->fast_out(i);
600 if (get_block_for_node(m) == block && // Local-block user
601 m != this_call && // Not self-start node
602 m->is_MachCall()) {
603 call = m;
604 break;
605 }
606 }
607 if (call == NULL) return; // No next call (e.g., block end is near)
608 // Set next-call for all inputs to this call
609 set_next_call(block, call, next_call);
610 }
612 //------------------------------add_call_kills-------------------------------------
613 // helper function that adds caller save registers to MachProjNode
614 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) {
615 // Fill in the kill mask for the call
616 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
617 if( !regs.Member(r) ) { // Not already defined by the call
618 // Save-on-call register?
619 if ((save_policy[r] == 'C') ||
620 (save_policy[r] == 'A') ||
621 ((save_policy[r] == 'E') && exclude_soe)) {
622 proj->_rout.Insert(r);
623 }
624 }
625 }
626 }
629 //------------------------------sched_call-------------------------------------
630 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) {
631 RegMask regs;
633 // Schedule all the users of the call right now. All the users are
634 // projection Nodes, so they must be scheduled next to the call.
635 // Collect all the defined registers.
636 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
637 Node* n = mcall->fast_out(i);
638 assert( n->is_MachProj(), "" );
639 int n_cnt = ready_cnt.at(n->_idx)-1;
640 ready_cnt.at_put(n->_idx, n_cnt);
641 assert( n_cnt == 0, "" );
642 // Schedule next to call
643 block->map_node(n, node_cnt++);
644 // Collect defined registers
645 regs.OR(n->out_RegMask());
646 // Check for scheduling the next control-definer
647 if( n->bottom_type() == Type::CONTROL )
648 // Warm up next pile of heuristic bits
649 needed_for_next_call(block, n, next_call);
651 // Children of projections are now all ready
652 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
653 Node* m = n->fast_out(j); // Get user
654 if(get_block_for_node(m) != block) {
655 continue;
656 }
657 if( m->is_Phi() ) continue;
658 int m_cnt = ready_cnt.at(m->_idx)-1;
659 ready_cnt.at_put(m->_idx, m_cnt);
660 if( m_cnt == 0 )
661 worklist.push(m);
662 }
664 }
666 // Act as if the call defines the Frame Pointer.
667 // Certainly the FP is alive and well after the call.
668 regs.Insert(_matcher.c_frame_pointer());
670 // Set all registers killed and not already defined by the call.
671 uint r_cnt = mcall->tf()->range()->cnt();
672 int op = mcall->ideal_Opcode();
673 MachProjNode *proj = new (C) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
674 map_node_to_block(proj, block);
675 block->insert_node(proj, node_cnt++);
677 // Select the right register save policy.
678 const char *save_policy = NULL;
679 switch (op) {
680 case Op_CallRuntime:
681 case Op_CallLeaf:
682 case Op_CallLeafNoFP:
683 // Calling C code so use C calling convention
684 save_policy = _matcher._c_reg_save_policy;
685 break;
687 case Op_CallStaticJava:
688 case Op_CallDynamicJava:
689 // Calling Java code so use Java calling convention
690 save_policy = _matcher._register_save_policy;
691 break;
693 default:
694 ShouldNotReachHere();
695 }
697 // When using CallRuntime mark SOE registers as killed by the call
698 // so values that could show up in the RegisterMap aren't live in a
699 // callee saved register since the register wouldn't know where to
700 // find them. CallLeaf and CallLeafNoFP are ok because they can't
701 // have debug info on them. Strictly speaking this only needs to be
702 // done for oops since idealreg2debugmask takes care of debug info
703 // references but there no way to handle oops differently than other
704 // pointers as far as the kill mask goes.
705 bool exclude_soe = op == Op_CallRuntime;
707 // If the call is a MethodHandle invoke, we need to exclude the
708 // register which is used to save the SP value over MH invokes from
709 // the mask. Otherwise this register could be used for
710 // deoptimization information.
711 if (op == Op_CallStaticJava) {
712 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall;
713 if (mcallstaticjava->_method_handle_invoke)
714 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
715 }
717 add_call_kills(proj, regs, save_policy, exclude_soe);
719 return node_cnt;
720 }
723 //------------------------------schedule_local---------------------------------
724 // Topological sort within a block. Someday become a real scheduler.
725 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call) {
726 // Already "sorted" are the block start Node (as the first entry), and
727 // the block-ending Node and any trailing control projections. We leave
728 // these alone. PhiNodes and ParmNodes are made to follow the block start
729 // Node. Everything else gets topo-sorted.
731 #ifndef PRODUCT
732 if (trace_opto_pipelining()) {
733 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order);
734 for (uint i = 0;i < block->number_of_nodes(); i++) {
735 tty->print("# ");
736 block->get_node(i)->fast_dump();
737 }
738 tty->print_cr("#");
739 }
740 #endif
742 // RootNode is already sorted
743 if (block->number_of_nodes() == 1) {
744 return true;
745 }
747 // Move PhiNodes and ParmNodes from 1 to cnt up to the start
748 uint node_cnt = block->end_idx();
749 uint phi_cnt = 1;
750 uint i;
751 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
752 Node *n = block->get_node(i);
753 if( n->is_Phi() || // Found a PhiNode or ParmNode
754 (n->is_Proj() && n->in(0) == block->head()) ) {
755 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
756 block->map_node(block->get_node(phi_cnt), i);
757 block->map_node(n, phi_cnt++); // swap Phi/Parm up front
758 } else { // All others
759 // Count block-local inputs to 'n'
760 uint cnt = n->len(); // Input count
761 uint local = 0;
762 for( uint j=0; j<cnt; j++ ) {
763 Node *m = n->in(j);
764 if( m && get_block_for_node(m) == block && !m->is_top() )
765 local++; // One more block-local input
766 }
767 ready_cnt.at_put(n->_idx, local); // Count em up
769 #ifdef ASSERT
770 if( UseConcMarkSweepGC || UseG1GC ) {
771 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
772 // Check the precedence edges
773 for (uint prec = n->req(); prec < n->len(); prec++) {
774 Node* oop_store = n->in(prec);
775 if (oop_store != NULL) {
776 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark");
777 }
778 }
779 }
780 }
781 #endif
783 // A few node types require changing a required edge to a precedence edge
784 // before allocation.
785 if( n->is_Mach() && n->req() > TypeFunc::Parms &&
786 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
787 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
788 // MemBarAcquire could be created without Precedent edge.
789 // del_req() replaces the specified edge with the last input edge
790 // and then removes the last edge. If the specified edge > number of
791 // edges the last edge will be moved outside of the input edges array
792 // and the edge will be lost. This is why this code should be
793 // executed only when Precedent (== TypeFunc::Parms) edge is present.
794 Node *x = n->in(TypeFunc::Parms);
795 n->del_req(TypeFunc::Parms);
796 n->add_prec(x);
797 }
798 }
799 }
800 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count
801 ready_cnt.at_put(block->get_node(i2)->_idx, 0);
803 // All the prescheduled guys do not hold back internal nodes
804 uint i3;
805 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled
806 Node *n = block->get_node(i3); // Get pre-scheduled
807 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
808 Node* m = n->fast_out(j);
809 if (get_block_for_node(m) == block) { // Local-block user
810 int m_cnt = ready_cnt.at(m->_idx)-1;
811 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count
812 }
813 }
814 }
816 Node_List delay;
817 // Make a worklist
818 Node_List worklist;
819 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
820 Node *m = block->get_node(i4);
821 if( !ready_cnt.at(m->_idx) ) { // Zero ready count?
822 if (m->is_iteratively_computed()) {
823 // Push induction variable increments last to allow other uses
824 // of the phi to be scheduled first. The select() method breaks
825 // ties in scheduling by worklist order.
826 delay.push(m);
827 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
828 // Force the CreateEx to the top of the list so it's processed
829 // first and ends up at the start of the block.
830 worklist.insert(0, m);
831 } else {
832 worklist.push(m); // Then on to worklist!
833 }
834 }
835 }
836 while (delay.size()) {
837 Node* d = delay.pop();
838 worklist.push(d);
839 }
841 // Warm up the 'next_call' heuristic bits
842 needed_for_next_call(block, block->head(), next_call);
844 #ifndef PRODUCT
845 if (trace_opto_pipelining()) {
846 for (uint j=0; j< block->number_of_nodes(); j++) {
847 Node *n = block->get_node(j);
848 int idx = n->_idx;
849 tty->print("# ready cnt:%3d ", ready_cnt.at(idx));
850 tty->print("latency:%3d ", get_latency_for_node(n));
851 tty->print("%4d: %s\n", idx, n->Name());
852 }
853 }
854 #endif
856 uint max_idx = (uint)ready_cnt.length();
857 // Pull from worklist and schedule
858 while( worklist.size() ) { // Worklist is not ready
860 #ifndef PRODUCT
861 if (trace_opto_pipelining()) {
862 tty->print("# ready list:");
863 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
864 Node *n = worklist[i]; // Get Node on worklist
865 tty->print(" %d", n->_idx);
866 }
867 tty->cr();
868 }
869 #endif
871 // Select and pop a ready guy from worklist
872 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt);
873 block->map_node(n, phi_cnt++); // Schedule him next
875 #ifndef PRODUCT
876 if (trace_opto_pipelining()) {
877 tty->print("# select %d: %s", n->_idx, n->Name());
878 tty->print(", latency:%d", get_latency_for_node(n));
879 n->dump();
880 if (Verbose) {
881 tty->print("# ready list:");
882 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
883 Node *n = worklist[i]; // Get Node on worklist
884 tty->print(" %d", n->_idx);
885 }
886 tty->cr();
887 }
888 }
890 #endif
891 if( n->is_MachCall() ) {
892 MachCallNode *mcall = n->as_MachCall();
893 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call);
894 continue;
895 }
897 if (n->is_Mach() && n->as_Mach()->has_call()) {
898 RegMask regs;
899 regs.Insert(_matcher.c_frame_pointer());
900 regs.OR(n->out_RegMask());
902 MachProjNode *proj = new (C) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj );
903 map_node_to_block(proj, block);
904 block->insert_node(proj, phi_cnt++);
906 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false);
907 }
909 // Children are now all ready
910 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
911 Node* m = n->fast_out(i5); // Get user
912 if (get_block_for_node(m) != block) {
913 continue;
914 }
915 if( m->is_Phi() ) continue;
916 if (m->_idx >= max_idx) { // new node, skip it
917 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types");
918 continue;
919 }
920 int m_cnt = ready_cnt.at(m->_idx)-1;
921 ready_cnt.at_put(m->_idx, m_cnt);
922 if( m_cnt == 0 )
923 worklist.push(m);
924 }
925 }
927 if( phi_cnt != block->end_idx() ) {
928 // did not schedule all. Retry, Bailout, or Die
929 if (C->subsume_loads() == true && !C->failing()) {
930 // Retry with subsume_loads == false
931 // If this is the first failure, the sentinel string will "stick"
932 // to the Compile object, and the C2Compiler will see it and retry.
933 C->record_failure(C2Compiler::retry_no_subsuming_loads());
934 }
935 // assert( phi_cnt == end_idx(), "did not schedule all" );
936 return false;
937 }
939 #ifndef PRODUCT
940 if (trace_opto_pipelining()) {
941 tty->print_cr("#");
942 tty->print_cr("# after schedule_local");
943 for (uint i = 0;i < block->number_of_nodes();i++) {
944 tty->print("# ");
945 block->get_node(i)->fast_dump();
946 }
947 tty->cr();
948 }
949 #endif
952 return true;
953 }
955 //--------------------------catch_cleanup_fix_all_inputs-----------------------
956 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
957 for (uint l = 0; l < use->len(); l++) {
958 if (use->in(l) == old_def) {
959 if (l < use->req()) {
960 use->set_req(l, new_def);
961 } else {
962 use->rm_prec(l);
963 use->add_prec(new_def);
964 l--;
965 }
966 }
967 }
968 }
970 //------------------------------catch_cleanup_find_cloned_def------------------
971 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
972 assert( use_blk != def_blk, "Inter-block cleanup only");
974 // The use is some block below the Catch. Find and return the clone of the def
975 // that dominates the use. If there is no clone in a dominating block, then
976 // create a phi for the def in a dominating block.
978 // Find which successor block dominates this use. The successor
979 // blocks must all be single-entry (from the Catch only; I will have
980 // split blocks to make this so), hence they all dominate.
981 while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
982 use_blk = use_blk->_idom;
984 // Find the successor
985 Node *fixup = NULL;
987 uint j;
988 for( j = 0; j < def_blk->_num_succs; j++ )
989 if( use_blk == def_blk->_succs[j] )
990 break;
992 if( j == def_blk->_num_succs ) {
993 // Block at same level in dom-tree is not a successor. It needs a
994 // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
995 Node_Array inputs = new Node_List(Thread::current()->resource_area());
996 for(uint k = 1; k < use_blk->num_preds(); k++) {
997 Block* block = get_block_for_node(use_blk->pred(k));
998 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx));
999 }
1001 // Check to see if the use_blk already has an identical phi inserted.
1002 // If it exists, it will be at the first position since all uses of a
1003 // def are processed together.
1004 Node *phi = use_blk->get_node(1);
1005 if( phi->is_Phi() ) {
1006 fixup = phi;
1007 for (uint k = 1; k < use_blk->num_preds(); k++) {
1008 if (phi->in(k) != inputs[k]) {
1009 // Not a match
1010 fixup = NULL;
1011 break;
1012 }
1013 }
1014 }
1016 // If an existing PhiNode was not found, make a new one.
1017 if (fixup == NULL) {
1018 Node *new_phi = PhiNode::make(use_blk->head(), def);
1019 use_blk->insert_node(new_phi, 1);
1020 map_node_to_block(new_phi, use_blk);
1021 for (uint k = 1; k < use_blk->num_preds(); k++) {
1022 new_phi->set_req(k, inputs[k]);
1023 }
1024 fixup = new_phi;
1025 }
1027 } else {
1028 // Found the use just below the Catch. Make it use the clone.
1029 fixup = use_blk->get_node(n_clone_idx);
1030 }
1032 return fixup;
1033 }
1035 //--------------------------catch_cleanup_intra_block--------------------------
1036 // Fix all input edges in use that reference "def". The use is in the same
1037 // block as the def and both have been cloned in each successor block.
1038 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
1040 // Both the use and def have been cloned. For each successor block,
1041 // get the clone of the use, and make its input the clone of the def
1042 // found in that block.
1044 uint use_idx = blk->find_node(use);
1045 uint offset_idx = use_idx - beg;
1046 for( uint k = 0; k < blk->_num_succs; k++ ) {
1047 // Get clone in each successor block
1048 Block *sb = blk->_succs[k];
1049 Node *clone = sb->get_node(offset_idx+1);
1050 assert( clone->Opcode() == use->Opcode(), "" );
1052 // Make use-clone reference the def-clone
1053 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx));
1054 }
1055 }
1057 //------------------------------catch_cleanup_inter_block---------------------
1058 // Fix all input edges in use that reference "def". The use is in a different
1059 // block than the def.
1060 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
1061 if( !use_blk ) return; // Can happen if the use is a precedence edge
1063 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx);
1064 catch_cleanup_fix_all_inputs(use, def, new_def);
1065 }
1067 //------------------------------call_catch_cleanup-----------------------------
1068 // If we inserted any instructions between a Call and his CatchNode,
1069 // clone the instructions on all paths below the Catch.
1070 void PhaseCFG::call_catch_cleanup(Block* block) {
1072 // End of region to clone
1073 uint end = block->end_idx();
1074 if( !block->get_node(end)->is_Catch() ) return;
1075 // Start of region to clone
1076 uint beg = end;
1077 while(!block->get_node(beg-1)->is_MachProj() ||
1078 !block->get_node(beg-1)->in(0)->is_MachCall() ) {
1079 beg--;
1080 assert(beg > 0,"Catch cleanup walking beyond block boundary");
1081 }
1082 // Range of inserted instructions is [beg, end)
1083 if( beg == end ) return;
1085 // Clone along all Catch output paths. Clone area between the 'beg' and
1086 // 'end' indices.
1087 for( uint i = 0; i < block->_num_succs; i++ ) {
1088 Block *sb = block->_succs[i];
1089 // Clone the entire area; ignoring the edge fixup for now.
1090 for( uint j = end; j > beg; j-- ) {
1091 // It is safe here to clone a node with anti_dependence
1092 // since clones dominate on each path.
1093 Node *clone = block->get_node(j-1)->clone();
1094 sb->insert_node(clone, 1);
1095 map_node_to_block(clone, sb);
1096 }
1097 }
1100 // Fixup edges. Check the def-use info per cloned Node
1101 for(uint i2 = beg; i2 < end; i2++ ) {
1102 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
1103 Node *n = block->get_node(i2); // Node that got cloned
1104 // Need DU safe iterator because of edge manipulation in calls.
1105 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
1106 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
1107 out->push(n->fast_out(j1));
1108 }
1109 uint max = out->size();
1110 for (uint j = 0; j < max; j++) {// For all users
1111 Node *use = out->pop();
1112 Block *buse = get_block_for_node(use);
1113 if( use->is_Phi() ) {
1114 for( uint k = 1; k < use->req(); k++ )
1115 if( use->in(k) == n ) {
1116 Block* b = get_block_for_node(buse->pred(k));
1117 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx);
1118 use->set_req(k, fixup);
1119 }
1120 } else {
1121 if (block == buse) {
1122 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx);
1123 } else {
1124 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx);
1125 }
1126 }
1127 } // End for all users
1129 } // End of for all Nodes in cloned area
1131 // Remove the now-dead cloned ops
1132 for(uint i3 = beg; i3 < end; i3++ ) {
1133 block->get_node(beg)->disconnect_inputs(NULL, C);
1134 block->remove_node(beg);
1135 }
1137 // If the successor blocks have a CreateEx node, move it back to the top
1138 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) {
1139 Block *sb = block->_succs[i4];
1140 uint new_cnt = end - beg;
1141 // Remove any newly created, but dead, nodes.
1142 for( uint j = new_cnt; j > 0; j-- ) {
1143 Node *n = sb->get_node(j);
1144 if (n->outcnt() == 0 &&
1145 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
1146 n->disconnect_inputs(NULL, C);
1147 sb->remove_node(j);
1148 new_cnt--;
1149 }
1150 }
1151 // If any newly created nodes remain, move the CreateEx node to the top
1152 if (new_cnt > 0) {
1153 Node *cex = sb->get_node(1+new_cnt);
1154 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
1155 sb->remove_node(1+new_cnt);
1156 sb->insert_node(cex, 1);
1157 }
1158 }
1159 }
1160 }