Mercurial > hg > graal-compiler
comparison src/share/vm/opto/loopTransform.cpp @ 2468:6c97c830fb6f
Merge
author | jrose |
---|---|
date | Sat, 09 Apr 2011 21:16:12 -0700 |
parents | 1d1603768966 3af54845df98 |
children | ae93231c7a1f |
comparison
equal
deleted
inserted
replaced
2439:0930dc920c18 | 2468:6c97c830fb6f |
---|---|
58 // Put loop body on igvn work list | 58 // Put loop body on igvn work list |
59 void IdealLoopTree::record_for_igvn() { | 59 void IdealLoopTree::record_for_igvn() { |
60 for( uint i = 0; i < _body.size(); i++ ) { | 60 for( uint i = 0; i < _body.size(); i++ ) { |
61 Node *n = _body.at(i); | 61 Node *n = _body.at(i); |
62 _phase->_igvn._worklist.push(n); | 62 _phase->_igvn._worklist.push(n); |
63 } | |
64 } | |
65 | |
66 //------------------------------compute_exact_trip_count----------------------- | |
67 // Compute loop exact trip count if possible. Do not recalculate trip count for | |
68 // split loops (pre-main-post) which have their limits and inits behind Opaque node. | |
69 void IdealLoopTree::compute_exact_trip_count( PhaseIdealLoop *phase ) { | |
70 if (!_head->as_Loop()->is_valid_counted_loop()) { | |
71 return; | |
72 } | |
73 CountedLoopNode* cl = _head->as_CountedLoop(); | |
74 // Trip count may become nonexact for iteration split loops since | |
75 // RCE modifies limits. Note, _trip_count value is not reset since | |
76 // it is used to limit unrolling of main loop. | |
77 cl->set_nonexact_trip_count(); | |
78 | |
79 // Loop's test should be part of loop. | |
80 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) | |
81 return; // Infinite loop | |
82 | |
83 #ifdef ASSERT | |
84 BoolTest::mask bt = cl->loopexit()->test_trip(); | |
85 assert(bt == BoolTest::lt || bt == BoolTest::gt || | |
86 bt == BoolTest::ne, "canonical test is expected"); | |
87 #endif | |
88 | |
89 Node* init_n = cl->init_trip(); | |
90 Node* limit_n = cl->limit(); | |
91 if (init_n != NULL && init_n->is_Con() && | |
92 limit_n != NULL && limit_n->is_Con()) { | |
93 // Use longs to avoid integer overflow. | |
94 int stride_con = cl->stride_con(); | |
95 long init_con = cl->init_trip()->get_int(); | |
96 long limit_con = cl->limit()->get_int(); | |
97 int stride_m = stride_con - (stride_con > 0 ? 1 : -1); | |
98 long trip_count = (limit_con - init_con + stride_m)/stride_con; | |
99 if (trip_count > 0 && (julong)trip_count < (julong)max_juint) { | |
100 // Set exact trip count. | |
101 cl->set_exact_trip_count((uint)trip_count); | |
102 } | |
63 } | 103 } |
64 } | 104 } |
65 | 105 |
66 //------------------------------compute_profile_trip_cnt---------------------------- | 106 //------------------------------compute_profile_trip_cnt---------------------------- |
67 // Compute loop trip count from profile data as | 107 // Compute loop trip count from profile data as |
299 // backedges) and then map to the new peeled iteration. This leaves | 339 // backedges) and then map to the new peeled iteration. This leaves |
300 // the pre-loop with only 1 user (the new peeled iteration), but the | 340 // the pre-loop with only 1 user (the new peeled iteration), but the |
301 // peeled-loop backedge has 2 users. | 341 // peeled-loop backedge has 2 users. |
302 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the | 342 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the |
303 // extra backedge user. | 343 // extra backedge user. |
344 // | |
345 // orig | |
346 // | |
347 // stmt1 | |
348 // | | |
349 // v | |
350 // loop predicate | |
351 // | | |
352 // v | |
353 // loop<----+ | |
354 // | | | |
355 // stmt2 | | |
356 // | | | |
357 // v | | |
358 // if ^ | |
359 // / \ | | |
360 // / \ | | |
361 // v v | | |
362 // false true | | |
363 // / \ | | |
364 // / ----+ | |
365 // | | |
366 // v | |
367 // exit | |
368 // | |
369 // | |
370 // after clone loop | |
371 // | |
372 // stmt1 | |
373 // | | |
374 // v | |
375 // loop predicate | |
376 // / \ | |
377 // clone / \ orig | |
378 // / \ | |
379 // / \ | |
380 // v v | |
381 // +---->loop clone loop<----+ | |
382 // | | | | | |
383 // | stmt2 clone stmt2 | | |
384 // | | | | | |
385 // | v v | | |
386 // ^ if clone If ^ | |
387 // | / \ / \ | | |
388 // | / \ / \ | | |
389 // | v v v v | | |
390 // | true false false true | | |
391 // | / \ / \ | | |
392 // +---- \ / ----+ | |
393 // \ / | |
394 // 1v v2 | |
395 // region | |
396 // | | |
397 // v | |
398 // exit | |
399 // | |
400 // | |
401 // after peel and predicate move | |
402 // | |
403 // stmt1 | |
404 // / | |
405 // / | |
406 // clone / orig | |
407 // / | |
408 // / +----------+ | |
409 // / | | | |
410 // / loop predicate | | |
411 // / | | | |
412 // v v | | |
413 // TOP-->loop clone loop<----+ | | |
414 // | | | | | |
415 // stmt2 clone stmt2 | | | |
416 // | | | ^ | |
417 // v v | | | |
418 // if clone If ^ | | |
419 // / \ / \ | | | |
420 // / \ / \ | | | |
421 // v v v v | | | |
422 // true false false true | | | |
423 // | \ / \ | | | |
424 // | \ / ----+ ^ | |
425 // | \ / | | |
426 // | 1v v2 | | |
427 // v region | | |
428 // | | | | |
429 // | v | | |
430 // | exit | | |
431 // | | | |
432 // +--------------->-----------------+ | |
433 // | |
434 // | |
435 // final graph | |
436 // | |
437 // stmt1 | |
438 // | | |
439 // v | |
440 // stmt2 clone | |
441 // | | |
442 // v | |
443 // if clone | |
444 // / | | |
445 // / | | |
446 // v v | |
447 // false true | |
448 // | | | |
449 // | v | |
450 // | loop predicate | |
451 // | | | |
452 // | v | |
453 // | loop<----+ | |
454 // | | | | |
455 // | stmt2 | | |
456 // | | | | |
457 // | v | | |
458 // v if ^ | |
459 // | / \ | | |
460 // | / \ | | |
461 // | v v | | |
462 // | false true | | |
463 // | | \ | | |
464 // v v --+ | |
465 // region | |
466 // | | |
467 // v | |
468 // exit | |
469 // | |
304 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) { | 470 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) { |
305 | 471 |
306 C->set_major_progress(); | 472 C->set_major_progress(); |
307 // Peeling a 'main' loop in a pre/main/post situation obfuscates the | 473 // Peeling a 'main' loop in a pre/main/post situation obfuscates the |
308 // 'pre' loop from the main and the 'pre' can no longer have it's | 474 // 'pre' loop from the main and the 'pre' can no longer have it's |
313 if (TraceLoopOpts) { | 479 if (TraceLoopOpts) { |
314 tty->print("Peel "); | 480 tty->print("Peel "); |
315 loop->dump_head(); | 481 loop->dump_head(); |
316 } | 482 } |
317 #endif | 483 #endif |
318 Node *h = loop->_head; | 484 Node* head = loop->_head; |
319 if (h->is_CountedLoop()) { | 485 bool counted_loop = head->is_CountedLoop(); |
320 CountedLoopNode *cl = h->as_CountedLoop(); | 486 if (counted_loop) { |
487 CountedLoopNode *cl = head->as_CountedLoop(); | |
321 assert(cl->trip_count() > 0, "peeling a fully unrolled loop"); | 488 assert(cl->trip_count() > 0, "peeling a fully unrolled loop"); |
322 cl->set_trip_count(cl->trip_count() - 1); | 489 cl->set_trip_count(cl->trip_count() - 1); |
323 if (cl->is_main_loop()) { | 490 if (cl->is_main_loop()) { |
324 cl->set_normal_loop(); | 491 cl->set_normal_loop(); |
325 #ifndef PRODUCT | 492 #ifndef PRODUCT |
328 loop->dump_head(); | 495 loop->dump_head(); |
329 } | 496 } |
330 #endif | 497 #endif |
331 } | 498 } |
332 } | 499 } |
500 Node* entry = head->in(LoopNode::EntryControl); | |
333 | 501 |
334 // Step 1: Clone the loop body. The clone becomes the peeled iteration. | 502 // Step 1: Clone the loop body. The clone becomes the peeled iteration. |
335 // The pre-loop illegally has 2 control users (old & new loops). | 503 // The pre-loop illegally has 2 control users (old & new loops). |
336 clone_loop( loop, old_new, dom_depth(loop->_head) ); | 504 clone_loop( loop, old_new, dom_depth(head) ); |
337 | |
338 | 505 |
339 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. | 506 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. |
340 // Do this by making the old-loop fall-in edges act as if they came | 507 // Do this by making the old-loop fall-in edges act as if they came |
341 // around the loopback from the prior iteration (follow the old-loop | 508 // around the loopback from the prior iteration (follow the old-loop |
342 // backedges) and then map to the new peeled iteration. This leaves | 509 // backedges) and then map to the new peeled iteration. This leaves |
343 // the pre-loop with only 1 user (the new peeled iteration), but the | 510 // the pre-loop with only 1 user (the new peeled iteration), but the |
344 // peeled-loop backedge has 2 users. | 511 // peeled-loop backedge has 2 users. |
345 for (DUIterator_Fast jmax, j = loop->_head->fast_outs(jmax); j < jmax; j++) { | 512 Node* new_exit_value = old_new[head->in(LoopNode::LoopBackControl)->_idx]; |
346 Node* old = loop->_head->fast_out(j); | 513 new_exit_value = move_loop_predicates(entry, new_exit_value); |
347 if( old->in(0) == loop->_head && old->req() == 3 && | 514 _igvn.hash_delete(head); |
348 (old->is_Loop() || old->is_Phi()) ) { | 515 head->set_req(LoopNode::EntryControl, new_exit_value); |
349 Node *new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; | 516 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { |
350 if( !new_exit_value ) // Backedge value is ALSO loop invariant? | 517 Node* old = head->fast_out(j); |
518 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) { | |
519 new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; | |
520 if (!new_exit_value ) // Backedge value is ALSO loop invariant? | |
351 // Then loop body backedge value remains the same. | 521 // Then loop body backedge value remains the same. |
352 new_exit_value = old->in(LoopNode::LoopBackControl); | 522 new_exit_value = old->in(LoopNode::LoopBackControl); |
353 _igvn.hash_delete(old); | 523 _igvn.hash_delete(old); |
354 old->set_req(LoopNode::EntryControl, new_exit_value); | 524 old->set_req(LoopNode::EntryControl, new_exit_value); |
355 } | 525 } |
356 } | 526 } |
357 | 527 |
358 | 528 |
359 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the | 529 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the |
360 // extra backedge user. | 530 // extra backedge user. |
361 Node *nnn = old_new[loop->_head->_idx]; | 531 Node* new_head = old_new[head->_idx]; |
362 _igvn.hash_delete(nnn); | 532 _igvn.hash_delete(new_head); |
363 nnn->set_req(LoopNode::LoopBackControl, C->top()); | 533 new_head->set_req(LoopNode::LoopBackControl, C->top()); |
364 for (DUIterator_Fast j2max, j2 = nnn->fast_outs(j2max); j2 < j2max; j2++) { | 534 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) { |
365 Node* use = nnn->fast_out(j2); | 535 Node* use = new_head->fast_out(j2); |
366 if( use->in(0) == nnn && use->req() == 3 && use->is_Phi() ) { | 536 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) { |
367 _igvn.hash_delete(use); | 537 _igvn.hash_delete(use); |
368 use->set_req(LoopNode::LoopBackControl, C->top()); | 538 use->set_req(LoopNode::LoopBackControl, C->top()); |
369 } | 539 } |
370 } | 540 } |
371 | 541 |
372 | 542 |
373 // Step 4: Correct dom-depth info. Set to loop-head depth. | 543 // Step 4: Correct dom-depth info. Set to loop-head depth. |
374 int dd = dom_depth(loop->_head); | 544 int dd = dom_depth(head); |
375 set_idom(loop->_head, loop->_head->in(1), dd); | 545 set_idom(head, head->in(1), dd); |
376 for (uint j3 = 0; j3 < loop->_body.size(); j3++) { | 546 for (uint j3 = 0; j3 < loop->_body.size(); j3++) { |
377 Node *old = loop->_body.at(j3); | 547 Node *old = loop->_body.at(j3); |
378 Node *nnn = old_new[old->_idx]; | 548 Node *nnn = old_new[old->_idx]; |
379 if (!has_ctrl(nnn)) | 549 if (!has_ctrl(nnn)) |
380 set_idom(nnn, idom(nnn), dd-1); | 550 set_idom(nnn, idom(nnn), dd-1); |
381 // While we're at it, remove any SafePoints from the peeled code | 551 // While we're at it, remove any SafePoints from the peeled code |
382 if( old->Opcode() == Op_SafePoint ) { | 552 if (old->Opcode() == Op_SafePoint) { |
383 Node *nnn = old_new[old->_idx]; | 553 Node *nnn = old_new[old->_idx]; |
384 lazy_replace(nnn,nnn->in(TypeFunc::Control)); | 554 lazy_replace(nnn,nnn->in(TypeFunc::Control)); |
385 } | 555 } |
386 } | 556 } |
387 | 557 |
390 peeled_dom_test_elim(loop,old_new); | 560 peeled_dom_test_elim(loop,old_new); |
391 | 561 |
392 loop->record_for_igvn(); | 562 loop->record_for_igvn(); |
393 } | 563 } |
394 | 564 |
565 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop | |
566 | |
395 //------------------------------policy_maximally_unroll------------------------ | 567 //------------------------------policy_maximally_unroll------------------------ |
396 // Return exact loop trip count, or 0 if not maximally unrolling | 568 // Calculate exact loop trip count and return true if loop can be maximally |
569 // unrolled. | |
397 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const { | 570 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const { |
398 CountedLoopNode *cl = _head->as_CountedLoop(); | 571 CountedLoopNode *cl = _head->as_CountedLoop(); |
399 assert(cl->is_normal_loop(), ""); | 572 assert(cl->is_normal_loop(), ""); |
400 | 573 if (!cl->is_valid_counted_loop()) |
401 Node *init_n = cl->init_trip(); | 574 return false; // Malformed counted loop |
402 Node *limit_n = cl->limit(); | 575 |
403 | 576 if (!cl->has_exact_trip_count()) { |
404 // Non-constant bounds | 577 // Trip count is not exact. |
405 if (init_n == NULL || !init_n->is_Con() || | |
406 limit_n == NULL || !limit_n->is_Con() || | |
407 // protect against stride not being a constant | |
408 !cl->stride_is_con()) { | |
409 return false; | 578 return false; |
410 } | 579 } |
411 int init = init_n->get_int(); | 580 |
412 int limit = limit_n->get_int(); | 581 uint trip_count = cl->trip_count(); |
413 int span = limit - init; | 582 // Note, max_juint is used to indicate unknown trip count. |
414 int stride = cl->stride_con(); | 583 assert(trip_count > 1, "one iteration loop should be optimized out already"); |
415 | 584 assert(trip_count < max_juint, "exact trip_count should be less than max_uint."); |
416 if (init >= limit || stride > span) { | |
417 // return a false (no maximally unroll) and the regular unroll/peel | |
418 // route will make a small mess which CCP will fold away. | |
419 return false; | |
420 } | |
421 uint trip_count = span/stride; // trip_count can be greater than 2 Gig. | |
422 assert( (int)trip_count*stride == span, "must divide evenly" ); | |
423 | 585 |
424 // Real policy: if we maximally unroll, does it get too big? | 586 // Real policy: if we maximally unroll, does it get too big? |
425 // Allow the unrolled mess to get larger than standard loop | 587 // Allow the unrolled mess to get larger than standard loop |
426 // size. After all, it will no longer be a loop. | 588 // size. After all, it will no longer be a loop. |
427 uint body_size = _body.size(); | 589 uint body_size = _body.size(); |
428 uint unroll_limit = (uint)LoopUnrollLimit * 4; | 590 uint unroll_limit = (uint)LoopUnrollLimit * 4; |
429 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits"); | 591 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits"); |
430 cl->set_trip_count(trip_count); | |
431 if (trip_count > unroll_limit || body_size > unroll_limit) { | 592 if (trip_count > unroll_limit || body_size > unroll_limit) { |
593 return false; | |
594 } | |
595 | |
596 // Take into account that after unroll conjoined heads and tails will fold, | |
597 // otherwise policy_unroll() may allow more unrolling than max unrolling. | |
598 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count; | |
599 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE; | |
600 if (body_size != tst_body_size) // Check for int overflow | |
601 return false; | |
602 if (new_body_size > unroll_limit || | |
603 // Unrolling can result in a large amount of node construction | |
604 new_body_size >= MaxNodeLimit - phase->C->unique()) { | |
432 return false; | 605 return false; |
433 } | 606 } |
434 | 607 |
435 // Currently we don't have policy to optimize one iteration loops. | 608 // Currently we don't have policy to optimize one iteration loops. |
436 // Maximally unrolling transformation is used for that: | 609 // Maximally unrolling transformation is used for that: |
437 // it is peeled and the original loop become non reachable (dead). | 610 // it is peeled and the original loop become non reachable (dead). |
438 if (trip_count == 1) | 611 // Also fully unroll a loop with few iterations regardless next |
612 // conditions since following loop optimizations will split | |
613 // such loop anyway (pre-main-post). | |
614 if (trip_count <= 3) | |
439 return true; | 615 return true; |
440 | 616 |
441 // Do not unroll a loop with String intrinsics code. | 617 // Do not unroll a loop with String intrinsics code. |
442 // String intrinsics are large and have loops. | 618 // String intrinsics are large and have loops. |
443 for (uint k = 0; k < _body.size(); k++) { | 619 for (uint k = 0; k < _body.size(); k++) { |
450 return false; | 626 return false; |
451 } | 627 } |
452 } // switch | 628 } // switch |
453 } | 629 } |
454 | 630 |
455 if (body_size <= unroll_limit) { | 631 return true; // Do maximally unroll |
456 uint new_body_size = body_size * trip_count; | |
457 if (new_body_size <= unroll_limit && | |
458 body_size == new_body_size / trip_count && | |
459 // Unrolling can result in a large amount of node construction | |
460 new_body_size < MaxNodeLimit - phase->C->unique()) { | |
461 return true; // maximally unroll | |
462 } | |
463 } | |
464 | |
465 return false; // Do not maximally unroll | |
466 } | 632 } |
467 | 633 |
468 | 634 |
469 //------------------------------policy_unroll---------------------------------- | 635 //------------------------------policy_unroll---------------------------------- |
470 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if | 636 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if |
472 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const { | 638 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const { |
473 | 639 |
474 CountedLoopNode *cl = _head->as_CountedLoop(); | 640 CountedLoopNode *cl = _head->as_CountedLoop(); |
475 assert(cl->is_normal_loop() || cl->is_main_loop(), ""); | 641 assert(cl->is_normal_loop() || cl->is_main_loop(), ""); |
476 | 642 |
477 // protect against stride not being a constant | 643 if (!cl->is_valid_counted_loop()) |
478 if (!cl->stride_is_con()) return false; | 644 return false; // Malformed counted loop |
479 | 645 |
480 // protect against over-unrolling | 646 // protect against over-unrolling |
481 if (cl->trip_count() <= 1) return false; | 647 if (cl->trip_count() <= 1) return false; |
648 | |
649 // Check for stride being a small enough constant | |
650 if (abs(cl->stride_con()) > (1<<3)) return false; | |
482 | 651 |
483 int future_unroll_ct = cl->unrolled_count() * 2; | 652 int future_unroll_ct = cl->unrolled_count() * 2; |
484 | 653 |
485 // Don't unroll if the next round of unrolling would push us | 654 // Don't unroll if the next round of unrolling would push us |
486 // over the expected trip count of the loop. One is subtracted | 655 // over the expected trip count of the loop. One is subtracted |
558 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true; | 727 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true; |
559 // Normal case: loop too big | 728 // Normal case: loop too big |
560 return false; | 729 return false; |
561 } | 730 } |
562 | 731 |
563 // Check for stride being a small enough constant | |
564 if (abs(cl->stride_con()) > (1<<3)) return false; | |
565 | |
566 // Unroll once! (Each trip will soon do double iterations) | 732 // Unroll once! (Each trip will soon do double iterations) |
567 return true; | 733 return true; |
568 } | 734 } |
569 | 735 |
570 //------------------------------policy_align----------------------------------- | 736 //------------------------------policy_align----------------------------------- |
954 #ifndef PRODUCT | 1120 #ifndef PRODUCT |
955 if (PrintOpto && VerifyLoopOptimizations) { | 1121 if (PrintOpto && VerifyLoopOptimizations) { |
956 tty->print("Unrolling "); | 1122 tty->print("Unrolling "); |
957 loop->dump_head(); | 1123 loop->dump_head(); |
958 } else if (TraceLoopOpts) { | 1124 } else if (TraceLoopOpts) { |
959 tty->print("Unroll %d ", loop_head->unrolled_count()*2); | 1125 if (loop_head->trip_count() < (uint)LoopUnrollLimit) { |
1126 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count()); | |
1127 } else { | |
1128 tty->print("Unroll %d ", loop_head->unrolled_count()*2); | |
1129 } | |
960 loop->dump_head(); | 1130 loop->dump_head(); |
961 } | 1131 } |
962 #endif | 1132 #endif |
963 | 1133 |
964 // Remember loop node count before unrolling to detect | 1134 // Remember loop node count before unrolling to detect |
1629 // Micro-benchmark spamming. Policy is to always remove empty loops. | 1799 // Micro-benchmark spamming. Policy is to always remove empty loops. |
1630 // The 'DO' part is to replace the trip counter with the value it will | 1800 // The 'DO' part is to replace the trip counter with the value it will |
1631 // have on the last iteration. This will break the loop. | 1801 // have on the last iteration. This will break the loop. |
1632 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) { | 1802 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) { |
1633 // Minimum size must be empty loop | 1803 // Minimum size must be empty loop |
1634 if (_body.size() > 7/*number of nodes in an empty loop*/) | 1804 if (_body.size() > EMPTY_LOOP_SIZE) |
1635 return false; | 1805 return false; |
1636 | 1806 |
1637 if (!_head->is_CountedLoop()) | 1807 if (!_head->is_CountedLoop()) |
1638 return false; // Dead loop | 1808 return false; // Dead loop |
1639 CountedLoopNode *cl = _head->as_CountedLoop(); | 1809 CountedLoopNode *cl = _head->as_CountedLoop(); |
1656 #endif | 1826 #endif |
1657 | 1827 |
1658 // main and post loops have explicitly created zero trip guard | 1828 // main and post loops have explicitly created zero trip guard |
1659 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop(); | 1829 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop(); |
1660 if (needs_guard) { | 1830 if (needs_guard) { |
1831 // Skip guard if values not overlap. | |
1832 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int(); | |
1833 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int(); | |
1834 int stride_con = cl->stride_con(); | |
1835 if (stride_con > 0) { | |
1836 needs_guard = (init_t->_hi >= limit_t->_lo); | |
1837 } else { | |
1838 needs_guard = (init_t->_lo <= limit_t->_hi); | |
1839 } | |
1840 } | |
1841 if (needs_guard) { | |
1661 // Check for an obvious zero trip guard. | 1842 // Check for an obvious zero trip guard. |
1662 Node* inctrl = cl->in(LoopNode::EntryControl); | 1843 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl)); |
1663 if (inctrl->Opcode() == Op_IfTrue) { | 1844 if (inctrl->Opcode() == Op_IfTrue) { |
1664 // The test should look like just the backedge of a CountedLoop | 1845 // The test should look like just the backedge of a CountedLoop |
1665 Node* iff = inctrl->in(0); | 1846 Node* iff = inctrl->in(0); |
1666 if (iff->is_If()) { | 1847 if (iff->is_If()) { |
1667 Node* bol = iff->in(1); | 1848 Node* bol = iff->in(1); |
1700 phase->_igvn.replace_node(phi,final); | 1881 phase->_igvn.replace_node(phi,final); |
1701 phase->C->set_major_progress(); | 1882 phase->C->set_major_progress(); |
1702 return true; | 1883 return true; |
1703 } | 1884 } |
1704 | 1885 |
1886 //------------------------------policy_do_one_iteration_loop------------------- | |
1887 // Convert one iteration loop into normal code. | |
1888 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) { | |
1889 if (!_head->as_Loop()->is_valid_counted_loop()) | |
1890 return false; // Only for counted loop | |
1891 | |
1892 CountedLoopNode *cl = _head->as_CountedLoop(); | |
1893 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) { | |
1894 return false; | |
1895 } | |
1896 | |
1897 #ifndef PRODUCT | |
1898 if(TraceLoopOpts) { | |
1899 tty->print("OneIteration "); | |
1900 this->dump_head(); | |
1901 } | |
1902 #endif | |
1903 | |
1904 Node *init_n = cl->init_trip(); | |
1905 #ifdef ASSERT | |
1906 // Loop boundaries should be constant since trip count is exact. | |
1907 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration"); | |
1908 #endif | |
1909 // Replace the phi at loop head with the value of the init_trip. | |
1910 // Then the CountedLoopEnd will collapse (backedge will not be taken) | |
1911 // and all loop-invariant uses of the exit values will be correct. | |
1912 phase->_igvn.replace_node(cl->phi(), cl->init_trip()); | |
1913 phase->C->set_major_progress(); | |
1914 return true; | |
1915 } | |
1705 | 1916 |
1706 //============================================================================= | 1917 //============================================================================= |
1707 //------------------------------iteration_split_impl--------------------------- | 1918 //------------------------------iteration_split_impl--------------------------- |
1708 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) { | 1919 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) { |
1920 // Compute exact loop trip count if possible. | |
1921 compute_exact_trip_count(phase); | |
1922 | |
1923 // Convert one iteration loop into normal code. | |
1924 if (policy_do_one_iteration_loop(phase)) | |
1925 return true; | |
1926 | |
1709 // Check and remove empty loops (spam micro-benchmarks) | 1927 // Check and remove empty loops (spam micro-benchmarks) |
1710 if( policy_do_remove_empty_loop(phase) ) | 1928 if (policy_do_remove_empty_loop(phase)) |
1711 return true; // Here we removed an empty loop | 1929 return true; // Here we removed an empty loop |
1712 | 1930 |
1713 bool should_peel = policy_peeling(phase); // Should we peel? | 1931 bool should_peel = policy_peeling(phase); // Should we peel? |
1714 | 1932 |
1715 bool should_unswitch = policy_unswitching(phase); | 1933 bool should_unswitch = policy_unswitching(phase); |
1716 | 1934 |
1717 // Non-counted loops may be peeled; exactly 1 iteration is peeled. | 1935 // Non-counted loops may be peeled; exactly 1 iteration is peeled. |
1718 // This removes loop-invariant tests (usually null checks). | 1936 // This removes loop-invariant tests (usually null checks). |
1719 if( !_head->is_CountedLoop() ) { // Non-counted loop | 1937 if (!_head->is_CountedLoop()) { // Non-counted loop |
1720 if (PartialPeelLoop && phase->partial_peel(this, old_new)) { | 1938 if (PartialPeelLoop && phase->partial_peel(this, old_new)) { |
1721 // Partial peel succeeded so terminate this round of loop opts | 1939 // Partial peel succeeded so terminate this round of loop opts |
1722 return false; | 1940 return false; |
1723 } | 1941 } |
1724 if( should_peel ) { // Should we peel? | 1942 if (should_peel) { // Should we peel? |
1725 #ifndef PRODUCT | 1943 #ifndef PRODUCT |
1726 if (PrintOpto) tty->print_cr("should_peel"); | 1944 if (PrintOpto) tty->print_cr("should_peel"); |
1727 #endif | 1945 #endif |
1728 phase->do_peeling(this,old_new); | 1946 phase->do_peeling(this,old_new); |
1729 } else if( should_unswitch ) { | 1947 } else if (should_unswitch) { |
1730 phase->do_unswitching(this, old_new); | 1948 phase->do_unswitching(this, old_new); |
1731 } | 1949 } |
1732 return true; | 1950 return true; |
1733 } | 1951 } |
1734 CountedLoopNode *cl = _head->as_CountedLoop(); | 1952 CountedLoopNode *cl = _head->as_CountedLoop(); |
1735 | 1953 |
1736 if( !cl->loopexit() ) return true; // Ignore various kinds of broken loops | 1954 if (!cl->loopexit()) return true; // Ignore various kinds of broken loops |
1737 | 1955 |
1738 // Do nothing special to pre- and post- loops | 1956 // Do nothing special to pre- and post- loops |
1739 if( cl->is_pre_loop() || cl->is_post_loop() ) return true; | 1957 if (cl->is_pre_loop() || cl->is_post_loop()) return true; |
1740 | 1958 |
1741 // Compute loop trip count from profile data | 1959 // Compute loop trip count from profile data |
1742 compute_profile_trip_cnt(phase); | 1960 compute_profile_trip_cnt(phase); |
1743 | 1961 |
1744 // Before attempting fancy unrolling, RCE or alignment, see if we want | 1962 // Before attempting fancy unrolling, RCE or alignment, see if we want |
1745 // to completely unroll this loop or do loop unswitching. | 1963 // to completely unroll this loop or do loop unswitching. |
1746 if( cl->is_normal_loop() ) { | 1964 if (cl->is_normal_loop()) { |
1747 if (should_unswitch) { | 1965 if (should_unswitch) { |
1748 phase->do_unswitching(this, old_new); | 1966 phase->do_unswitching(this, old_new); |
1749 return true; | 1967 return true; |
1750 } | 1968 } |
1751 bool should_maximally_unroll = policy_maximally_unroll(phase); | 1969 bool should_maximally_unroll = policy_maximally_unroll(phase); |
1752 if( should_maximally_unroll ) { | 1970 if (should_maximally_unroll) { |
1753 // Here we did some unrolling and peeling. Eventually we will | 1971 // Here we did some unrolling and peeling. Eventually we will |
1754 // completely unroll this loop and it will no longer be a loop. | 1972 // completely unroll this loop and it will no longer be a loop. |
1755 phase->do_maximally_unroll(this,old_new); | 1973 phase->do_maximally_unroll(this,old_new); |
1756 return true; | 1974 return true; |
1757 } | 1975 } |
1758 } | 1976 } |
1759 | 1977 |
1978 // Skip next optimizations if running low on nodes. Note that | |
1979 // policy_unswitching and policy_maximally_unroll have this check. | |
1980 uint nodes_left = MaxNodeLimit - phase->C->unique(); | |
1981 if ((2 * _body.size()) > nodes_left) { | |
1982 return true; | |
1983 } | |
1760 | 1984 |
1761 // Counted loops may be peeled, may need some iterations run up | 1985 // Counted loops may be peeled, may need some iterations run up |
1762 // front for RCE, and may want to align loop refs to a cache | 1986 // front for RCE, and may want to align loop refs to a cache |
1763 // line. Thus we clone a full loop up front whose trip count is | 1987 // line. Thus we clone a full loop up front whose trip count is |
1764 // at least 1 (if peeling), but may be several more. | 1988 // at least 1 (if peeling), but may be several more. |
1785 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align; | 2009 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align; |
1786 | 2010 |
1787 // If we have any of these conditions (RCE, alignment, unrolling) met, then | 2011 // If we have any of these conditions (RCE, alignment, unrolling) met, then |
1788 // we switch to the pre-/main-/post-loop model. This model also covers | 2012 // we switch to the pre-/main-/post-loop model. This model also covers |
1789 // peeling. | 2013 // peeling. |
1790 if( should_rce || should_align || should_unroll ) { | 2014 if (should_rce || should_align || should_unroll) { |
1791 if( cl->is_normal_loop() ) // Convert to 'pre/main/post' loops | 2015 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops |
1792 phase->insert_pre_post_loops(this,old_new, !may_rce_align); | 2016 phase->insert_pre_post_loops(this,old_new, !may_rce_align); |
1793 | 2017 |
1794 // Adjust the pre- and main-loop limits to let the pre and post loops run | 2018 // Adjust the pre- and main-loop limits to let the pre and post loops run |
1795 // with full checks, but the main-loop with no checks. Remove said | 2019 // with full checks, but the main-loop with no checks. Remove said |
1796 // checks from the main body. | 2020 // checks from the main body. |
1797 if( should_rce ) | 2021 if (should_rce) |
1798 phase->do_range_check(this,old_new); | 2022 phase->do_range_check(this,old_new); |
1799 | 2023 |
1800 // Double loop body for unrolling. Adjust the minimum-trip test (will do | 2024 // Double loop body for unrolling. Adjust the minimum-trip test (will do |
1801 // twice as many iterations as before) and the main body limit (only do | 2025 // twice as many iterations as before) and the main body limit (only do |
1802 // an even number of trips). If we are peeling, we might enable some RCE | 2026 // an even number of trips). If we are peeling, we might enable some RCE |
1803 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if | 2027 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if |
1804 // peeling. | 2028 // peeling. |
1805 if( should_unroll && !should_peel ) | 2029 if (should_unroll && !should_peel) |
1806 phase->do_unroll(this,old_new, true); | 2030 phase->do_unroll(this,old_new, true); |
1807 | 2031 |
1808 // Adjust the pre-loop limits to align the main body | 2032 // Adjust the pre-loop limits to align the main body |
1809 // iterations. | 2033 // iterations. |
1810 if( should_align ) | 2034 if (should_align) |
1811 Unimplemented(); | 2035 Unimplemented(); |
1812 | 2036 |
1813 } else { // Else we have an unchanged counted loop | 2037 } else { // Else we have an unchanged counted loop |
1814 if( should_peel ) // Might want to peel but do nothing else | 2038 if (should_peel) // Might want to peel but do nothing else |
1815 phase->do_peeling(this,old_new); | 2039 phase->do_peeling(this,old_new); |
1816 } | 2040 } |
1817 return true; | 2041 return true; |
1818 } | 2042 } |
1819 | 2043 |
1859 if (_next && !_next->iteration_split(phase, old_new)) | 2083 if (_next && !_next->iteration_split(phase, old_new)) |
1860 return false; | 2084 return false; |
1861 return true; | 2085 return true; |
1862 } | 2086 } |
1863 | 2087 |
1864 //-------------------------------is_uncommon_trap_proj---------------------------- | 2088 |
1865 // Return true if proj is the form of "proj->[region->..]call_uct" | 2089 //============================================================================= |
1866 bool PhaseIdealLoop::is_uncommon_trap_proj(ProjNode* proj, Deoptimization::DeoptReason reason) { | |
1867 int path_limit = 10; | |
1868 assert(proj, "invalid argument"); | |
1869 Node* out = proj; | |
1870 for (int ct = 0; ct < path_limit; ct++) { | |
1871 out = out->unique_ctrl_out(); | |
1872 if (out == NULL || out->is_Root() || out->is_Start()) | |
1873 return false; | |
1874 if (out->is_CallStaticJava()) { | |
1875 int req = out->as_CallStaticJava()->uncommon_trap_request(); | |
1876 if (req != 0) { | |
1877 Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req); | |
1878 if (trap_reason == reason || reason == Deoptimization::Reason_none) { | |
1879 return true; | |
1880 } | |
1881 } | |
1882 return false; // don't do further after call | |
1883 } | |
1884 } | |
1885 return false; | |
1886 } | |
1887 | |
1888 //-------------------------------is_uncommon_trap_if_pattern------------------------- | |
1889 // Return true for "if(test)-> proj -> ... | |
1890 // | | |
1891 // V | |
1892 // other_proj->[region->..]call_uct" | |
1893 // | |
1894 // "must_reason_predicate" means the uct reason must be Reason_predicate | |
1895 bool PhaseIdealLoop::is_uncommon_trap_if_pattern(ProjNode *proj, Deoptimization::DeoptReason reason) { | |
1896 Node *in0 = proj->in(0); | |
1897 if (!in0->is_If()) return false; | |
1898 // Variation of a dead If node. | |
1899 if (in0->outcnt() < 2) return false; | |
1900 IfNode* iff = in0->as_If(); | |
1901 | |
1902 // we need "If(Conv2B(Opaque1(...)))" pattern for reason_predicate | |
1903 if (reason != Deoptimization::Reason_none) { | |
1904 if (iff->in(1)->Opcode() != Op_Conv2B || | |
1905 iff->in(1)->in(1)->Opcode() != Op_Opaque1) { | |
1906 return false; | |
1907 } | |
1908 } | |
1909 | |
1910 ProjNode* other_proj = iff->proj_out(1-proj->_con)->as_Proj(); | |
1911 return is_uncommon_trap_proj(other_proj, reason); | |
1912 } | |
1913 | |
1914 //-------------------------------register_control------------------------- | |
1915 void PhaseIdealLoop::register_control(Node* n, IdealLoopTree *loop, Node* pred) { | |
1916 assert(n->is_CFG(), "must be control node"); | |
1917 _igvn.register_new_node_with_optimizer(n); | |
1918 loop->_body.push(n); | |
1919 set_loop(n, loop); | |
1920 // When called from beautify_loops() idom is not constructed yet. | |
1921 if (_idom != NULL) { | |
1922 set_idom(n, pred, dom_depth(pred)); | |
1923 } | |
1924 } | |
1925 | |
1926 //------------------------------create_new_if_for_predicate------------------------ | |
1927 // create a new if above the uct_if_pattern for the predicate to be promoted. | |
1928 // | |
1929 // before after | |
1930 // ---------- ---------- | |
1931 // ctrl ctrl | |
1932 // | | | |
1933 // | | | |
1934 // v v | |
1935 // iff new_iff | |
1936 // / \ / \ | |
1937 // / \ / \ | |
1938 // v v v v | |
1939 // uncommon_proj cont_proj if_uct if_cont | |
1940 // \ | | | | | |
1941 // \ | | | | | |
1942 // v v v | v | |
1943 // rgn loop | iff | |
1944 // | | / \ | |
1945 // | | / \ | |
1946 // v | v v | |
1947 // uncommon_trap | uncommon_proj cont_proj | |
1948 // \ \ | | | |
1949 // \ \ | | | |
1950 // v v v v | |
1951 // rgn loop | |
1952 // | | |
1953 // | | |
1954 // v | |
1955 // uncommon_trap | |
1956 // | |
1957 // | |
1958 // We will create a region to guard the uct call if there is no one there. | |
1959 // The true projecttion (if_cont) of the new_iff is returned. | |
1960 // This code is also used to clone predicates to clonned loops. | |
1961 ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, | |
1962 Deoptimization::DeoptReason reason) { | |
1963 assert(is_uncommon_trap_if_pattern(cont_proj, reason), "must be a uct if pattern!"); | |
1964 IfNode* iff = cont_proj->in(0)->as_If(); | |
1965 | |
1966 ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con); | |
1967 Node *rgn = uncommon_proj->unique_ctrl_out(); | |
1968 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct"); | |
1969 | |
1970 if (!rgn->is_Region()) { // create a region to guard the call | |
1971 assert(rgn->is_Call(), "must be call uct"); | |
1972 CallNode* call = rgn->as_Call(); | |
1973 IdealLoopTree* loop = get_loop(call); | |
1974 rgn = new (C, 1) RegionNode(1); | |
1975 rgn->add_req(uncommon_proj); | |
1976 register_control(rgn, loop, uncommon_proj); | |
1977 _igvn.hash_delete(call); | |
1978 call->set_req(0, rgn); | |
1979 // When called from beautify_loops() idom is not constructed yet. | |
1980 if (_idom != NULL) { | |
1981 set_idom(call, rgn, dom_depth(rgn)); | |
1982 } | |
1983 } | |
1984 | |
1985 Node* entry = iff->in(0); | |
1986 if (new_entry != NULL) { | |
1987 // Clonning the predicate to new location. | |
1988 entry = new_entry; | |
1989 } | |
1990 // Create new_iff | |
1991 IdealLoopTree* lp = get_loop(entry); | |
1992 IfNode *new_iff = new (C, 2) IfNode(entry, NULL, iff->_prob, iff->_fcnt); | |
1993 register_control(new_iff, lp, entry); | |
1994 Node *if_cont = new (C, 1) IfTrueNode(new_iff); | |
1995 Node *if_uct = new (C, 1) IfFalseNode(new_iff); | |
1996 if (cont_proj->is_IfFalse()) { | |
1997 // Swap | |
1998 Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp; | |
1999 } | |
2000 register_control(if_cont, lp, new_iff); | |
2001 register_control(if_uct, get_loop(rgn), new_iff); | |
2002 | |
2003 // if_uct to rgn | |
2004 _igvn.hash_delete(rgn); | |
2005 rgn->add_req(if_uct); | |
2006 // When called from beautify_loops() idom is not constructed yet. | |
2007 if (_idom != NULL) { | |
2008 Node* ridom = idom(rgn); | |
2009 Node* nrdom = dom_lca(ridom, new_iff); | |
2010 set_idom(rgn, nrdom, dom_depth(rgn)); | |
2011 } | |
2012 // rgn must have no phis | |
2013 assert(!rgn->as_Region()->has_phi(), "region must have no phis"); | |
2014 | |
2015 if (new_entry == NULL) { | |
2016 // Attach if_cont to iff | |
2017 _igvn.hash_delete(iff); | |
2018 iff->set_req(0, if_cont); | |
2019 if (_idom != NULL) { | |
2020 set_idom(iff, if_cont, dom_depth(iff)); | |
2021 } | |
2022 } | |
2023 return if_cont->as_Proj(); | |
2024 } | |
2025 | |
2026 //--------------------------find_predicate_insertion_point------------------- | |
2027 // Find a good location to insert a predicate | |
2028 ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason) { | |
2029 if (start_c == NULL || !start_c->is_Proj()) | |
2030 return NULL; | |
2031 if (is_uncommon_trap_if_pattern(start_c->as_Proj(), reason)) { | |
2032 return start_c->as_Proj(); | |
2033 } | |
2034 return NULL; | |
2035 } | |
2036 | |
2037 //--------------------------find_predicate------------------------------------ | |
2038 // Find a predicate | |
2039 Node* PhaseIdealLoop::find_predicate(Node* entry) { | |
2040 Node* predicate = NULL; | |
2041 if (UseLoopPredicate) { | |
2042 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); | |
2043 if (predicate != NULL) { // right pattern that can be used by loop predication | |
2044 assert(entry->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be"); | |
2045 return entry; | |
2046 } | |
2047 } | |
2048 return NULL; | |
2049 } | |
2050 | |
2051 //------------------------------Invariance----------------------------------- | |
2052 // Helper class for loop_predication_impl to compute invariance on the fly and | |
2053 // clone invariants. | |
2054 class Invariance : public StackObj { | |
2055 VectorSet _visited, _invariant; | |
2056 Node_Stack _stack; | |
2057 VectorSet _clone_visited; | |
2058 Node_List _old_new; // map of old to new (clone) | |
2059 IdealLoopTree* _lpt; | |
2060 PhaseIdealLoop* _phase; | |
2061 | |
2062 // Helper function to set up the invariance for invariance computation | |
2063 // If n is a known invariant, set up directly. Otherwise, look up the | |
2064 // the possibility to push n onto the stack for further processing. | |
2065 void visit(Node* use, Node* n) { | |
2066 if (_lpt->is_invariant(n)) { // known invariant | |
2067 _invariant.set(n->_idx); | |
2068 } else if (!n->is_CFG()) { | |
2069 Node *n_ctrl = _phase->ctrl_or_self(n); | |
2070 Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG | |
2071 if (_phase->is_dominator(n_ctrl, u_ctrl)) { | |
2072 _stack.push(n, n->in(0) == NULL ? 1 : 0); | |
2073 } | |
2074 } | |
2075 } | |
2076 | |
2077 // Compute invariance for "the_node" and (possibly) all its inputs recursively | |
2078 // on the fly | |
2079 void compute_invariance(Node* n) { | |
2080 assert(_visited.test(n->_idx), "must be"); | |
2081 visit(n, n); | |
2082 while (_stack.is_nonempty()) { | |
2083 Node* n = _stack.node(); | |
2084 uint idx = _stack.index(); | |
2085 if (idx == n->req()) { // all inputs are processed | |
2086 _stack.pop(); | |
2087 // n is invariant if it's inputs are all invariant | |
2088 bool all_inputs_invariant = true; | |
2089 for (uint i = 0; i < n->req(); i++) { | |
2090 Node* in = n->in(i); | |
2091 if (in == NULL) continue; | |
2092 assert(_visited.test(in->_idx), "must have visited input"); | |
2093 if (!_invariant.test(in->_idx)) { // bad guy | |
2094 all_inputs_invariant = false; | |
2095 break; | |
2096 } | |
2097 } | |
2098 if (all_inputs_invariant) { | |
2099 _invariant.set(n->_idx); // I am a invariant too | |
2100 } | |
2101 } else { // process next input | |
2102 _stack.set_index(idx + 1); | |
2103 Node* m = n->in(idx); | |
2104 if (m != NULL && !_visited.test_set(m->_idx)) { | |
2105 visit(n, m); | |
2106 } | |
2107 } | |
2108 } | |
2109 } | |
2110 | |
2111 // Helper function to set up _old_new map for clone_nodes. | |
2112 // If n is a known invariant, set up directly ("clone" of n == n). | |
2113 // Otherwise, push n onto the stack for real cloning. | |
2114 void clone_visit(Node* n) { | |
2115 assert(_invariant.test(n->_idx), "must be invariant"); | |
2116 if (_lpt->is_invariant(n)) { // known invariant | |
2117 _old_new.map(n->_idx, n); | |
2118 } else{ // to be cloned | |
2119 assert (!n->is_CFG(), "should not see CFG here"); | |
2120 _stack.push(n, n->in(0) == NULL ? 1 : 0); | |
2121 } | |
2122 } | |
2123 | |
2124 // Clone "n" and (possibly) all its inputs recursively | |
2125 void clone_nodes(Node* n, Node* ctrl) { | |
2126 clone_visit(n); | |
2127 while (_stack.is_nonempty()) { | |
2128 Node* n = _stack.node(); | |
2129 uint idx = _stack.index(); | |
2130 if (idx == n->req()) { // all inputs processed, clone n! | |
2131 _stack.pop(); | |
2132 // clone invariant node | |
2133 Node* n_cl = n->clone(); | |
2134 _old_new.map(n->_idx, n_cl); | |
2135 _phase->register_new_node(n_cl, ctrl); | |
2136 for (uint i = 0; i < n->req(); i++) { | |
2137 Node* in = n_cl->in(i); | |
2138 if (in == NULL) continue; | |
2139 n_cl->set_req(i, _old_new[in->_idx]); | |
2140 } | |
2141 } else { // process next input | |
2142 _stack.set_index(idx + 1); | |
2143 Node* m = n->in(idx); | |
2144 if (m != NULL && !_clone_visited.test_set(m->_idx)) { | |
2145 clone_visit(m); // visit the input | |
2146 } | |
2147 } | |
2148 } | |
2149 } | |
2150 | |
2151 public: | |
2152 Invariance(Arena* area, IdealLoopTree* lpt) : | |
2153 _lpt(lpt), _phase(lpt->_phase), | |
2154 _visited(area), _invariant(area), _stack(area, 10 /* guess */), | |
2155 _clone_visited(area), _old_new(area) | |
2156 {} | |
2157 | |
2158 // Map old to n for invariance computation and clone | |
2159 void map_ctrl(Node* old, Node* n) { | |
2160 assert(old->is_CFG() && n->is_CFG(), "must be"); | |
2161 _old_new.map(old->_idx, n); // "clone" of old is n | |
2162 _invariant.set(old->_idx); // old is invariant | |
2163 _clone_visited.set(old->_idx); | |
2164 } | |
2165 | |
2166 // Driver function to compute invariance | |
2167 bool is_invariant(Node* n) { | |
2168 if (!_visited.test_set(n->_idx)) | |
2169 compute_invariance(n); | |
2170 return (_invariant.test(n->_idx) != 0); | |
2171 } | |
2172 | |
2173 // Driver function to clone invariant | |
2174 Node* clone(Node* n, Node* ctrl) { | |
2175 assert(ctrl->is_CFG(), "must be"); | |
2176 assert(_invariant.test(n->_idx), "must be an invariant"); | |
2177 if (!_clone_visited.test(n->_idx)) | |
2178 clone_nodes(n, ctrl); | |
2179 return _old_new[n->_idx]; | |
2180 } | |
2181 }; | |
2182 | |
2183 //------------------------------is_range_check_if ----------------------------------- | |
2184 // Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format | |
2185 // Note: this function is particularly designed for loop predication. We require load_range | |
2186 // and offset to be loop invariant computed on the fly by "invar" | |
2187 bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const { | |
2188 if (!is_loop_exit(iff)) { | |
2189 return false; | |
2190 } | |
2191 if (!iff->in(1)->is_Bool()) { | |
2192 return false; | |
2193 } | |
2194 const BoolNode *bol = iff->in(1)->as_Bool(); | |
2195 if (bol->_test._test != BoolTest::lt) { | |
2196 return false; | |
2197 } | |
2198 if (!bol->in(1)->is_Cmp()) { | |
2199 return false; | |
2200 } | |
2201 const CmpNode *cmp = bol->in(1)->as_Cmp(); | |
2202 if (cmp->Opcode() != Op_CmpU ) { | |
2203 return false; | |
2204 } | |
2205 Node* range = cmp->in(2); | |
2206 if (range->Opcode() != Op_LoadRange) { | |
2207 const TypeInt* tint = phase->_igvn.type(range)->isa_int(); | |
2208 if (!OptimizeFill || tint == NULL || tint->empty() || tint->_lo < 0) { | |
2209 // Allow predication on positive values that aren't LoadRanges. | |
2210 // This allows optimization of loops where the length of the | |
2211 // array is a known value and doesn't need to be loaded back | |
2212 // from the array. | |
2213 return false; | |
2214 } | |
2215 } | |
2216 if (!invar.is_invariant(range)) { | |
2217 return false; | |
2218 } | |
2219 Node *iv = _head->as_CountedLoop()->phi(); | |
2220 int scale = 0; | |
2221 Node *offset = NULL; | |
2222 if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) { | |
2223 return false; | |
2224 } | |
2225 if(offset && !invar.is_invariant(offset)) { // offset must be invariant | |
2226 return false; | |
2227 } | |
2228 return true; | |
2229 } | |
2230 | |
2231 //------------------------------rc_predicate----------------------------------- | |
2232 // Create a range check predicate | |
2233 // | |
2234 // for (i = init; i < limit; i += stride) { | |
2235 // a[scale*i+offset] | |
2236 // } | |
2237 // | |
2238 // Compute max(scale*i + offset) for init <= i < limit and build the predicate | |
2239 // as "max(scale*i + offset) u< a.length". | |
2240 // | |
2241 // There are two cases for max(scale*i + offset): | |
2242 // (1) stride*scale > 0 | |
2243 // max(scale*i + offset) = scale*(limit-stride) + offset | |
2244 // (2) stride*scale < 0 | |
2245 // max(scale*i + offset) = scale*init + offset | |
2246 BoolNode* PhaseIdealLoop::rc_predicate(Node* ctrl, | |
2247 int scale, Node* offset, | |
2248 Node* init, Node* limit, Node* stride, | |
2249 Node* range, bool upper) { | |
2250 DEBUG_ONLY(ttyLocker ttyl); | |
2251 if (TraceLoopPredicate) tty->print("rc_predicate "); | |
2252 | |
2253 Node* max_idx_expr = init; | |
2254 int stride_con = stride->get_int(); | |
2255 if ((stride_con > 0) == (scale > 0) == upper) { | |
2256 max_idx_expr = new (C, 3) SubINode(limit, stride); | |
2257 register_new_node(max_idx_expr, ctrl); | |
2258 if (TraceLoopPredicate) tty->print("(limit - stride) "); | |
2259 } else { | |
2260 if (TraceLoopPredicate) tty->print("init "); | |
2261 } | |
2262 | |
2263 if (scale != 1) { | |
2264 ConNode* con_scale = _igvn.intcon(scale); | |
2265 max_idx_expr = new (C, 3) MulINode(max_idx_expr, con_scale); | |
2266 register_new_node(max_idx_expr, ctrl); | |
2267 if (TraceLoopPredicate) tty->print("* %d ", scale); | |
2268 } | |
2269 | |
2270 if (offset && (!offset->is_Con() || offset->get_int() != 0)){ | |
2271 max_idx_expr = new (C, 3) AddINode(max_idx_expr, offset); | |
2272 register_new_node(max_idx_expr, ctrl); | |
2273 if (TraceLoopPredicate) | |
2274 if (offset->is_Con()) tty->print("+ %d ", offset->get_int()); | |
2275 else tty->print("+ offset "); | |
2276 } | |
2277 | |
2278 CmpUNode* cmp = new (C, 3) CmpUNode(max_idx_expr, range); | |
2279 register_new_node(cmp, ctrl); | |
2280 BoolNode* bol = new (C, 2) BoolNode(cmp, BoolTest::lt); | |
2281 register_new_node(bol, ctrl); | |
2282 | |
2283 if (TraceLoopPredicate) tty->print_cr("<u range"); | |
2284 return bol; | |
2285 } | |
2286 | |
2287 //------------------------------ loop_predication_impl-------------------------- | |
2288 // Insert loop predicates for null checks and range checks | |
2289 bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) { | |
2290 if (!UseLoopPredicate) return false; | |
2291 | |
2292 if (!loop->_head->is_Loop()) { | |
2293 // Could be a simple region when irreducible loops are present. | |
2294 return false; | |
2295 } | |
2296 | |
2297 if (loop->_head->unique_ctrl_out()->Opcode() == Op_NeverBranch) { | |
2298 // do nothing for infinite loops | |
2299 return false; | |
2300 } | |
2301 | |
2302 CountedLoopNode *cl = NULL; | |
2303 if (loop->_head->is_CountedLoop()) { | |
2304 cl = loop->_head->as_CountedLoop(); | |
2305 // do nothing for iteration-splitted loops | |
2306 if (!cl->is_normal_loop()) return false; | |
2307 } | |
2308 | |
2309 LoopNode *lpn = loop->_head->as_Loop(); | |
2310 Node* entry = lpn->in(LoopNode::EntryControl); | |
2311 | |
2312 ProjNode *predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); | |
2313 if (!predicate_proj) { | |
2314 #ifndef PRODUCT | |
2315 if (TraceLoopPredicate) { | |
2316 tty->print("missing predicate:"); | |
2317 loop->dump_head(); | |
2318 lpn->dump(1); | |
2319 } | |
2320 #endif | |
2321 return false; | |
2322 } | |
2323 ConNode* zero = _igvn.intcon(0); | |
2324 set_ctrl(zero, C->root()); | |
2325 | |
2326 ResourceArea *area = Thread::current()->resource_area(); | |
2327 Invariance invar(area, loop); | |
2328 | |
2329 // Create list of if-projs such that a newer proj dominates all older | |
2330 // projs in the list, and they all dominate loop->tail() | |
2331 Node_List if_proj_list(area); | |
2332 LoopNode *head = loop->_head->as_Loop(); | |
2333 Node *current_proj = loop->tail(); //start from tail | |
2334 while ( current_proj != head ) { | |
2335 if (loop == get_loop(current_proj) && // still in the loop ? | |
2336 current_proj->is_Proj() && // is a projection ? | |
2337 current_proj->in(0)->Opcode() == Op_If) { // is a if projection ? | |
2338 if_proj_list.push(current_proj); | |
2339 } | |
2340 current_proj = idom(current_proj); | |
2341 } | |
2342 | |
2343 bool hoisted = false; // true if at least one proj is promoted | |
2344 while (if_proj_list.size() > 0) { | |
2345 // Following are changed to nonnull when a predicate can be hoisted | |
2346 ProjNode* new_predicate_proj = NULL; | |
2347 | |
2348 ProjNode* proj = if_proj_list.pop()->as_Proj(); | |
2349 IfNode* iff = proj->in(0)->as_If(); | |
2350 | |
2351 if (!is_uncommon_trap_if_pattern(proj, Deoptimization::Reason_none)) { | |
2352 if (loop->is_loop_exit(iff)) { | |
2353 // stop processing the remaining projs in the list because the execution of them | |
2354 // depends on the condition of "iff" (iff->in(1)). | |
2355 break; | |
2356 } else { | |
2357 // Both arms are inside the loop. There are two cases: | |
2358 // (1) there is one backward branch. In this case, any remaining proj | |
2359 // in the if_proj list post-dominates "iff". So, the condition of "iff" | |
2360 // does not determine the execution the remining projs directly, and we | |
2361 // can safely continue. | |
2362 // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj" | |
2363 // does not dominate loop->tail(), so it can not be in the if_proj list. | |
2364 continue; | |
2365 } | |
2366 } | |
2367 | |
2368 Node* test = iff->in(1); | |
2369 if (!test->is_Bool()){ //Conv2B, ... | |
2370 continue; | |
2371 } | |
2372 BoolNode* bol = test->as_Bool(); | |
2373 if (invar.is_invariant(bol)) { | |
2374 // Invariant test | |
2375 new_predicate_proj = create_new_if_for_predicate(predicate_proj, NULL, | |
2376 Deoptimization::Reason_predicate); | |
2377 Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0); | |
2378 BoolNode* new_predicate_bol = invar.clone(bol, ctrl)->as_Bool(); | |
2379 | |
2380 // Negate test if necessary | |
2381 bool negated = false; | |
2382 if (proj->_con != predicate_proj->_con) { | |
2383 new_predicate_bol = new (C, 2) BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate()); | |
2384 register_new_node(new_predicate_bol, ctrl); | |
2385 negated = true; | |
2386 } | |
2387 IfNode* new_predicate_iff = new_predicate_proj->in(0)->as_If(); | |
2388 _igvn.hash_delete(new_predicate_iff); | |
2389 new_predicate_iff->set_req(1, new_predicate_bol); | |
2390 #ifndef PRODUCT | |
2391 if (TraceLoopPredicate) { | |
2392 tty->print("Predicate invariant if%s: %d ", negated ? " negated" : "", new_predicate_iff->_idx); | |
2393 loop->dump_head(); | |
2394 } else if (TraceLoopOpts) { | |
2395 tty->print("Predicate IC "); | |
2396 loop->dump_head(); | |
2397 } | |
2398 #endif | |
2399 } else if (cl != NULL && loop->is_range_check_if(iff, this, invar)) { | |
2400 assert(proj->_con == predicate_proj->_con, "must match"); | |
2401 | |
2402 // Range check for counted loops | |
2403 const Node* cmp = bol->in(1)->as_Cmp(); | |
2404 Node* idx = cmp->in(1); | |
2405 assert(!invar.is_invariant(idx), "index is variant"); | |
2406 assert(cmp->in(2)->Opcode() == Op_LoadRange || OptimizeFill, "must be"); | |
2407 Node* rng = cmp->in(2); | |
2408 assert(invar.is_invariant(rng), "range must be invariant"); | |
2409 int scale = 1; | |
2410 Node* offset = zero; | |
2411 bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset); | |
2412 assert(ok, "must be index expression"); | |
2413 | |
2414 Node* init = cl->init_trip(); | |
2415 Node* limit = cl->limit(); | |
2416 Node* stride = cl->stride(); | |
2417 | |
2418 // Build if's for the upper and lower bound tests. The | |
2419 // lower_bound test will dominate the upper bound test and all | |
2420 // cloned or created nodes will use the lower bound test as | |
2421 // their declared control. | |
2422 ProjNode* lower_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate); | |
2423 ProjNode* upper_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate); | |
2424 assert(upper_bound_proj->in(0)->as_If()->in(0) == lower_bound_proj, "should dominate"); | |
2425 Node *ctrl = lower_bound_proj->in(0)->as_If()->in(0); | |
2426 | |
2427 // Perform cloning to keep Invariance state correct since the | |
2428 // late schedule will place invariant things in the loop. | |
2429 rng = invar.clone(rng, ctrl); | |
2430 if (offset && offset != zero) { | |
2431 assert(invar.is_invariant(offset), "offset must be loop invariant"); | |
2432 offset = invar.clone(offset, ctrl); | |
2433 } | |
2434 | |
2435 // Test the lower bound | |
2436 Node* lower_bound_bol = rc_predicate(ctrl, scale, offset, init, limit, stride, rng, false); | |
2437 IfNode* lower_bound_iff = lower_bound_proj->in(0)->as_If(); | |
2438 _igvn.hash_delete(lower_bound_iff); | |
2439 lower_bound_iff->set_req(1, lower_bound_bol); | |
2440 if (TraceLoopPredicate) tty->print_cr("lower bound check if: %d", lower_bound_iff->_idx); | |
2441 | |
2442 // Test the upper bound | |
2443 Node* upper_bound_bol = rc_predicate(ctrl, scale, offset, init, limit, stride, rng, true); | |
2444 IfNode* upper_bound_iff = upper_bound_proj->in(0)->as_If(); | |
2445 _igvn.hash_delete(upper_bound_iff); | |
2446 upper_bound_iff->set_req(1, upper_bound_bol); | |
2447 if (TraceLoopPredicate) tty->print_cr("upper bound check if: %d", lower_bound_iff->_idx); | |
2448 | |
2449 // Fall through into rest of the clean up code which will move | |
2450 // any dependent nodes onto the upper bound test. | |
2451 new_predicate_proj = upper_bound_proj; | |
2452 | |
2453 #ifndef PRODUCT | |
2454 if (TraceLoopOpts && !TraceLoopPredicate) { | |
2455 tty->print("Predicate RC "); | |
2456 loop->dump_head(); | |
2457 } | |
2458 #endif | |
2459 } else { | |
2460 // Loop variant check (for example, range check in non-counted loop) | |
2461 // with uncommon trap. | |
2462 continue; | |
2463 } | |
2464 assert(new_predicate_proj != NULL, "sanity"); | |
2465 // Success - attach condition (new_predicate_bol) to predicate if | |
2466 invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate | |
2467 | |
2468 // Eliminate the old If in the loop body | |
2469 dominated_by( new_predicate_proj, iff, proj->_con != new_predicate_proj->_con ); | |
2470 | |
2471 hoisted = true; | |
2472 C->set_major_progress(); | |
2473 } // end while | |
2474 | |
2475 #ifndef PRODUCT | |
2476 // report that the loop predication has been actually performed | |
2477 // for this loop | |
2478 if (TraceLoopPredicate && hoisted) { | |
2479 tty->print("Loop Predication Performed:"); | |
2480 loop->dump_head(); | |
2481 } | |
2482 #endif | |
2483 | |
2484 return hoisted; | |
2485 } | |
2486 | |
2487 //------------------------------loop_predication-------------------------------- | |
2488 // driver routine for loop predication optimization | |
2489 bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) { | |
2490 bool hoisted = false; | |
2491 // Recursively promote predicates | |
2492 if ( _child ) { | |
2493 hoisted = _child->loop_predication( phase); | |
2494 } | |
2495 | |
2496 // self | |
2497 if (!_irreducible && !tail()->is_top()) { | |
2498 hoisted |= phase->loop_predication_impl(this); | |
2499 } | |
2500 | |
2501 if ( _next ) { //sibling | |
2502 hoisted |= _next->loop_predication( phase); | |
2503 } | |
2504 | |
2505 return hoisted; | |
2506 } | |
2507 | |
2508 | |
2509 // Process all the loops in the loop tree and replace any fill | 2090 // Process all the loops in the loop tree and replace any fill |
2510 // patterns with an intrisc version. | 2091 // patterns with an intrisc version. |
2511 bool PhaseIdealLoop::do_intrinsify_fill() { | 2092 bool PhaseIdealLoop::do_intrinsify_fill() { |
2512 bool changed = false; | 2093 bool changed = false; |
2513 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { | 2094 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { |
2623 } | 2204 } |
2624 #endif | 2205 #endif |
2625 if (value != head->phi()) { | 2206 if (value != head->phi()) { |
2626 msg = "unhandled shift in address"; | 2207 msg = "unhandled shift in address"; |
2627 } else { | 2208 } else { |
2628 found_index = true; | 2209 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) { |
2629 shift = n; | 2210 msg = "scale doesn't match"; |
2630 assert(type2aelembytes(store->as_Mem()->memory_type(), true) == 1 << shift->in(2)->get_int(), "scale should match"); | 2211 } else { |
2212 found_index = true; | |
2213 shift = n; | |
2214 } | |
2631 } | 2215 } |
2632 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) { | 2216 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) { |
2633 if (n->in(1) == head->phi()) { | 2217 if (n->in(1) == head->phi()) { |
2634 found_index = true; | 2218 found_index = true; |
2635 conv = n; | 2219 conv = n; |
2759 Node* shift = NULL; | 2343 Node* shift = NULL; |
2760 Node* offset = NULL; | 2344 Node* offset = NULL; |
2761 if (!match_fill_loop(lpt, store, store_value, shift, offset)) { | 2345 if (!match_fill_loop(lpt, store, store_value, shift, offset)) { |
2762 return false; | 2346 return false; |
2763 } | 2347 } |
2348 | |
2349 #ifndef PRODUCT | |
2350 if (TraceLoopOpts) { | |
2351 tty->print("ArrayFill "); | |
2352 lpt->dump_head(); | |
2353 } | |
2354 #endif | |
2764 | 2355 |
2765 // Now replace the whole loop body by a call to a fill routine that | 2356 // Now replace the whole loop body by a call to a fill routine that |
2766 // covers the same region as the loop. | 2357 // covers the same region as the loop. |
2767 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base); | 2358 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base); |
2768 | 2359 |