1 /*
2 * Copyright (c) 1998, 2016, 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 *
249 break; // Failed for some cutout?
250 }
251 x = copy; // Progress, try again
252 }
253
254 // Phis and 2-address instructions cannot change registers so easily - their
255 // outputs must match their input.
256 if( !can_change_regs )
257 return blk_adjust; // Only check stupid copies!
258
259 // Loop backedges won't have a value-mapping yet
260 if( &value == NULL ) return blk_adjust;
261
262 // Skip through all copies to the _value_ being used. Do not change from
263 // int to pointer. This attempts to jump through a chain of copies, where
264 // intermediate copies might be illegal, i.e., value is stored down to stack
265 // then reloaded BUT survives in a register the whole way.
266 Node *val = skip_copies(n->in(k));
267 if (val == x) return blk_adjust; // No progress?
268
269 int n_regs = RegMask::num_registers(val->ideal_reg());
270 uint val_idx = _lrg_map.live_range_id(val);
271 OptoReg::Name val_reg = lrgs(val_idx).reg();
272
273 // See if it happens to already be in the correct register!
274 // (either Phi's direct register, or the common case of the name
275 // never-clobbered original-def register)
276 if (register_contains_value(val, val_reg, n_regs, value)) {
277 blk_adjust += use_prior_register(n,k,regnd[val_reg],current_block,value,regnd);
278 if( n->in(k) == regnd[val_reg] ) // Success! Quit trying
279 return blk_adjust;
280 }
281
282 // See if we can skip the copy by changing registers. Don't change from
283 // using a register to using the stack unless we know we can remove a
284 // copy-load. Otherwise we might end up making a pile of Intel cisc-spill
285 // ops reading from memory instead of just loading once and using the
286 // register.
287
288 // Also handle duplicate copies here.
289 const Type *t = val->is_Con() ? val->bottom_type() : NULL;
290
291 // Scan all registers to see if this value is around already
292 for( uint reg = 0; reg < (uint)_max_reg; reg++ ) {
293 if (reg == (uint)nk_reg) {
294 // Found ourselves so check if there is only one user of this
295 // copy and keep on searching for a better copy if so.
296 bool ignore_self = true;
297 x = n->in(k);
298 DUIterator_Fast imax, i = x->fast_outs(imax);
299 Node* first = x->fast_out(i); i++;
300 while (i < imax && ignore_self) {
301 Node* use = x->fast_out(i); i++;
302 if (use != first) ignore_self = false;
303 }
304 if (ignore_self) continue;
305 }
306
307 Node *vv = value[reg];
308 if (n_regs > 1) { // Doubles and vectors check for aligned-adjacent set
309 uint last = (n_regs-1); // Looking for the last part of a set
310 if ((reg&last) != last) continue; // Wrong part of a set
311 if (!register_contains_value(vv, reg, n_regs, value)) continue; // Different value
312 }
313 if( vv == val || // Got a direct hit?
314 (t && vv && vv->bottom_type() == t && vv->is_Mach() &&
315 vv->as_Mach()->rule() == val->as_Mach()->rule()) ) { // Or same constant?
316 assert( !n->is_Phi(), "cannot change registers at a Phi so easily" );
317 if( OptoReg::is_stack(nk_reg) || // CISC-loading from stack OR
318 OptoReg::is_reg(reg) || // turning into a register use OR
319 regnd[reg]->outcnt()==1 ) { // last use of a spill-load turns into a CISC use
320 blk_adjust += use_prior_register(n,k,regnd[reg],current_block,value,regnd);
321 if( n->in(k) == regnd[reg] ) // Success! Quit trying
322 return blk_adjust;
323 } // End of if not degrading to a stack
324 } // End of if found value in another register
325 } // End of scan all machine registers
326 return blk_adjust;
327 }
328
329
574 for (j = 1; j < block->num_preds(); j++) {
575 Block* pb = _cfg.get_block_for_node(block->pred(j));
576 if (pb == freed) {
577 continue; // Did self already via freelist
578 }
579 Node_List &p_regnd = *blk2regnd[pb->_pre_order];
580 for( uint k = 0; k < (uint)_max_reg; k++ ) {
581 if( regnd[k] != p_regnd[k] ) { // Conflict on reaching defs?
582 value.map(k,NULL); // Then no value handy
583 regnd.map(k,NULL);
584 }
585 }
586 }
587 }
588
589 // For all Phi's
590 for (j = 1; j < phi_dex; j++) {
591 uint k;
592 Node *phi = block->get_node(j);
593 uint pidx = _lrg_map.live_range_id(phi);
594 OptoReg::Name preg = lrgs(_lrg_map.live_range_id(phi)).reg();
595
596 // Remove copies remaining on edges. Check for junk phi.
597 Node *u = NULL;
598 for (k = 1; k < phi->req(); k++) {
599 Node *x = phi->in(k);
600 if( phi != x && u != x ) // Found a different input
601 u = u ? NodeSentinel : x; // Capture unique input, or NodeSentinel for 2nd input
602 }
603 if (u != NodeSentinel) { // Junk Phi. Remove
604 phi->replace_by(u);
605 j -= yank_if_dead(phi, block, &value, ®nd);
606 phi_dex--;
607 continue;
608 }
609 // Note that if value[pidx] exists, then we merged no new values here
610 // and the phi is useless. This can happen even with the above phi
611 // removal for complex flows. I cannot keep the better known value here
612 // because locally the phi appears to define a new merged value. If I
613 // keep the better value then a copy of the phi, being unable to use the
614 // global flow analysis, can't "peek through" the phi to the original
615 // reaching value and so will act like it's defining a new value. This
616 // can lead to situations where some uses are from the old and some from
617 // the new values. Not illegal by itself but throws the over-strong
618 // assert in scheduling.
619 if( pidx ) {
620 value.map(preg,phi);
621 regnd.map(preg,phi);
622 int n_regs = RegMask::num_registers(phi->ideal_reg());
623 for (int l = 1; l < n_regs; l++) {
624 OptoReg::Name preg_lo = OptoReg::add(preg,-l);
625 value.map(preg_lo,phi);
626 regnd.map(preg_lo,phi);
627 }
628 }
629 }
630
631 // For all remaining instructions
632 for (j = phi_dex; j < block->number_of_nodes(); j++) {
633 Node* n = block->get_node(j);
634
635 if(n->outcnt() == 0 && // Dead?
636 n != C->top() && // (ignore TOP, it has no du info)
637 !n->is_Proj() ) { // fat-proj kills
638 j -= yank_if_dead(n, block, &value, ®nd);
639 continue;
640 }
641
642 // Improve reaching-def info. Occasionally post-alloc's liveness gives
646 // advantage of this info to set a reaching def for the use-reg.
647 uint k;
648 for (k = 1; k < n->req(); k++) {
649 Node *def = n->in(k); // n->in(k) is a USE; def is the DEF for this USE
650 guarantee(def != NULL, "no disconnected nodes at this point");
651 uint useidx = _lrg_map.live_range_id(def); // useidx is the live range index for this USE
652
653 if( useidx ) {
654 OptoReg::Name ureg = lrgs(useidx).reg();
655 if( !value[ureg] ) {
656 int idx; // Skip occasional useless copy
657 while( (idx=def->is_Copy()) != 0 &&
658 def->in(idx) != NULL && // NULL should not happen
659 ureg == lrgs(_lrg_map.live_range_id(def->in(idx))).reg())
660 def = def->in(idx);
661 Node *valdef = skip_copies(def); // tighten up val through non-useless copies
662 value.map(ureg,valdef); // record improved reaching-def info
663 regnd.map(ureg, def);
664 // Record other half of doubles
665 uint def_ideal_reg = def->ideal_reg();
666 int n_regs = RegMask::num_registers(def_ideal_reg);
667 for (int l = 1; l < n_regs; l++) {
668 OptoReg::Name ureg_lo = OptoReg::add(ureg,-l);
669 if (!value[ureg_lo] &&
670 (!RegMask::can_represent(ureg_lo) ||
671 lrgs(useidx).mask().Member(ureg_lo))) { // Nearly always adjacent
672 value.map(ureg_lo,valdef); // record improved reaching-def info
673 regnd.map(ureg_lo, def);
674 }
675 }
676 }
677 }
678 }
679
680 const uint two_adr = n->is_Mach() ? n->as_Mach()->two_adr() : 0;
681
682 // Remove copies along input edges
683 for (k = 1; k < n->req(); k++) {
684 j -= elide_copy(n, k, block, value, regnd, two_adr != k);
685 }
686
690 continue;
691 }
692
693 // Update the register defined by this instruction
694 OptoReg::Name nreg = lrgs(lidx).reg();
695 // Skip through all copies to the _value_ being defined.
696 // Do not change from int to pointer
697 Node *val = skip_copies(n);
698
699 // Clear out a dead definition before starting so that the
700 // elimination code doesn't have to guard against it. The
701 // definition could in fact be a kill projection with a count of
702 // 0 which is safe but since those are uninteresting for copy
703 // elimination just delete them as well.
704 if (regnd[nreg] != NULL && regnd[nreg]->outcnt() == 0) {
705 regnd.map(nreg, NULL);
706 value.map(nreg, NULL);
707 }
708
709 uint n_ideal_reg = n->ideal_reg();
710 int n_regs = RegMask::num_registers(n_ideal_reg);
711 if (n_regs == 1) {
712 // If Node 'n' does not change the value mapped by the register,
713 // then 'n' is a useless copy. Do not update the register->node
714 // mapping so 'n' will go dead.
715 if( value[nreg] != val ) {
716 if (eliminate_copy_of_constant(val, n, block, value, regnd, nreg, OptoReg::Bad)) {
717 j -= replace_and_yank_if_dead(n, nreg, block, value, regnd);
718 } else {
719 // Update the mapping: record new Node defined by the register
720 regnd.map(nreg,n);
721 // Update mapping for defined *value*, which is the defined
722 // Node after skipping all copies.
723 value.map(nreg,val);
724 }
725 } else if( !may_be_copy_of_callee(n) ) {
726 assert(n->is_Copy(), "");
727 j -= replace_and_yank_if_dead(n, nreg, block, value, regnd);
728 }
729 } else if (RegMask::is_vector(n_ideal_reg)) {
730 // If Node 'n' does not change the value mapped by the register,
|
1 /*
2 * Copyright (c) 1998, 2020, 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 *
249 break; // Failed for some cutout?
250 }
251 x = copy; // Progress, try again
252 }
253
254 // Phis and 2-address instructions cannot change registers so easily - their
255 // outputs must match their input.
256 if( !can_change_regs )
257 return blk_adjust; // Only check stupid copies!
258
259 // Loop backedges won't have a value-mapping yet
260 if( &value == NULL ) return blk_adjust;
261
262 // Skip through all copies to the _value_ being used. Do not change from
263 // int to pointer. This attempts to jump through a chain of copies, where
264 // intermediate copies might be illegal, i.e., value is stored down to stack
265 // then reloaded BUT survives in a register the whole way.
266 Node *val = skip_copies(n->in(k));
267 if (val == x) return blk_adjust; // No progress?
268
269 uint val_idx = _lrg_map.live_range_id(val);
270 OptoReg::Name val_reg = lrgs(val_idx).reg();
271 int n_regs = RegMask::num_registers(val->ideal_reg(), lrgs(val_idx));
272
273 // See if it happens to already be in the correct register!
274 // (either Phi's direct register, or the common case of the name
275 // never-clobbered original-def register)
276 if (register_contains_value(val, val_reg, n_regs, value)) {
277 blk_adjust += use_prior_register(n,k,regnd[val_reg],current_block,value,regnd);
278 if( n->in(k) == regnd[val_reg] ) // Success! Quit trying
279 return blk_adjust;
280 }
281
282 // See if we can skip the copy by changing registers. Don't change from
283 // using a register to using the stack unless we know we can remove a
284 // copy-load. Otherwise we might end up making a pile of Intel cisc-spill
285 // ops reading from memory instead of just loading once and using the
286 // register.
287
288 // Also handle duplicate copies here.
289 const Type *t = val->is_Con() ? val->bottom_type() : NULL;
290
291 // Scan all registers to see if this value is around already
292 for( uint reg = 0; reg < (uint)_max_reg; reg++ ) {
293 if (reg == (uint)nk_reg) {
294 // Found ourselves so check if there is only one user of this
295 // copy and keep on searching for a better copy if so.
296 bool ignore_self = true;
297 x = n->in(k);
298 DUIterator_Fast imax, i = x->fast_outs(imax);
299 Node* first = x->fast_out(i); i++;
300 while (i < imax && ignore_self) {
301 Node* use = x->fast_out(i); i++;
302 if (use != first) ignore_self = false;
303 }
304 if (ignore_self) continue;
305 }
306
307 Node *vv = value[reg];
308 // For scalable register, number of registers may be inconsistent between
309 // "val_reg" and "reg". For example, when "val" resides in register
310 // but "reg" is located in stack.
311 if (lrgs(val_idx).is_scalable()) {
312 assert(val->ideal_reg() == Op_VecA, "scalable vector register");
313 if (OptoReg::is_stack(reg)) {
314 n_regs = lrgs(val_idx).scalable_reg_slots();
315 } else {
316 n_regs = RegMask::SlotsPerVecA;
317 }
318 }
319 if (n_regs > 1) { // Doubles and vectors check for aligned-adjacent set
320 uint last;
321 if (lrgs(val_idx).is_scalable()) {
322 assert(val->ideal_reg() == Op_VecA, "scalable vector register");
323 // For scalable vector register, regmask is always SlotsPerVecA bits aligned
324 last = RegMask::SlotsPerVecA - 1;
325 } else {
326 last = (n_regs-1); // Looking for the last part of a set
327 }
328 if ((reg&last) != last) continue; // Wrong part of a set
329 if (!register_contains_value(vv, reg, n_regs, value)) continue; // Different value
330 }
331 if( vv == val || // Got a direct hit?
332 (t && vv && vv->bottom_type() == t && vv->is_Mach() &&
333 vv->as_Mach()->rule() == val->as_Mach()->rule()) ) { // Or same constant?
334 assert( !n->is_Phi(), "cannot change registers at a Phi so easily" );
335 if( OptoReg::is_stack(nk_reg) || // CISC-loading from stack OR
336 OptoReg::is_reg(reg) || // turning into a register use OR
337 regnd[reg]->outcnt()==1 ) { // last use of a spill-load turns into a CISC use
338 blk_adjust += use_prior_register(n,k,regnd[reg],current_block,value,regnd);
339 if( n->in(k) == regnd[reg] ) // Success! Quit trying
340 return blk_adjust;
341 } // End of if not degrading to a stack
342 } // End of if found value in another register
343 } // End of scan all machine registers
344 return blk_adjust;
345 }
346
347
592 for (j = 1; j < block->num_preds(); j++) {
593 Block* pb = _cfg.get_block_for_node(block->pred(j));
594 if (pb == freed) {
595 continue; // Did self already via freelist
596 }
597 Node_List &p_regnd = *blk2regnd[pb->_pre_order];
598 for( uint k = 0; k < (uint)_max_reg; k++ ) {
599 if( regnd[k] != p_regnd[k] ) { // Conflict on reaching defs?
600 value.map(k,NULL); // Then no value handy
601 regnd.map(k,NULL);
602 }
603 }
604 }
605 }
606
607 // For all Phi's
608 for (j = 1; j < phi_dex; j++) {
609 uint k;
610 Node *phi = block->get_node(j);
611 uint pidx = _lrg_map.live_range_id(phi);
612 OptoReg::Name preg = lrgs(pidx).reg();
613
614 // Remove copies remaining on edges. Check for junk phi.
615 Node *u = NULL;
616 for (k = 1; k < phi->req(); k++) {
617 Node *x = phi->in(k);
618 if( phi != x && u != x ) // Found a different input
619 u = u ? NodeSentinel : x; // Capture unique input, or NodeSentinel for 2nd input
620 }
621 if (u != NodeSentinel) { // Junk Phi. Remove
622 phi->replace_by(u);
623 j -= yank_if_dead(phi, block, &value, ®nd);
624 phi_dex--;
625 continue;
626 }
627 // Note that if value[pidx] exists, then we merged no new values here
628 // and the phi is useless. This can happen even with the above phi
629 // removal for complex flows. I cannot keep the better known value here
630 // because locally the phi appears to define a new merged value. If I
631 // keep the better value then a copy of the phi, being unable to use the
632 // global flow analysis, can't "peek through" the phi to the original
633 // reaching value and so will act like it's defining a new value. This
634 // can lead to situations where some uses are from the old and some from
635 // the new values. Not illegal by itself but throws the over-strong
636 // assert in scheduling.
637 if( pidx ) {
638 value.map(preg,phi);
639 regnd.map(preg,phi);
640 int n_regs = RegMask::num_registers(phi->ideal_reg(), lrgs(pidx));
641 for (int l = 1; l < n_regs; l++) {
642 OptoReg::Name preg_lo = OptoReg::add(preg,-l);
643 value.map(preg_lo,phi);
644 regnd.map(preg_lo,phi);
645 }
646 }
647 }
648
649 // For all remaining instructions
650 for (j = phi_dex; j < block->number_of_nodes(); j++) {
651 Node* n = block->get_node(j);
652
653 if(n->outcnt() == 0 && // Dead?
654 n != C->top() && // (ignore TOP, it has no du info)
655 !n->is_Proj() ) { // fat-proj kills
656 j -= yank_if_dead(n, block, &value, ®nd);
657 continue;
658 }
659
660 // Improve reaching-def info. Occasionally post-alloc's liveness gives
664 // advantage of this info to set a reaching def for the use-reg.
665 uint k;
666 for (k = 1; k < n->req(); k++) {
667 Node *def = n->in(k); // n->in(k) is a USE; def is the DEF for this USE
668 guarantee(def != NULL, "no disconnected nodes at this point");
669 uint useidx = _lrg_map.live_range_id(def); // useidx is the live range index for this USE
670
671 if( useidx ) {
672 OptoReg::Name ureg = lrgs(useidx).reg();
673 if( !value[ureg] ) {
674 int idx; // Skip occasional useless copy
675 while( (idx=def->is_Copy()) != 0 &&
676 def->in(idx) != NULL && // NULL should not happen
677 ureg == lrgs(_lrg_map.live_range_id(def->in(idx))).reg())
678 def = def->in(idx);
679 Node *valdef = skip_copies(def); // tighten up val through non-useless copies
680 value.map(ureg,valdef); // record improved reaching-def info
681 regnd.map(ureg, def);
682 // Record other half of doubles
683 uint def_ideal_reg = def->ideal_reg();
684 int n_regs = RegMask::num_registers(def_ideal_reg, lrgs(_lrg_map.live_range_id(def)));
685 for (int l = 1; l < n_regs; l++) {
686 OptoReg::Name ureg_lo = OptoReg::add(ureg,-l);
687 if (!value[ureg_lo] &&
688 (!RegMask::can_represent(ureg_lo) ||
689 lrgs(useidx).mask().Member(ureg_lo))) { // Nearly always adjacent
690 value.map(ureg_lo,valdef); // record improved reaching-def info
691 regnd.map(ureg_lo, def);
692 }
693 }
694 }
695 }
696 }
697
698 const uint two_adr = n->is_Mach() ? n->as_Mach()->two_adr() : 0;
699
700 // Remove copies along input edges
701 for (k = 1; k < n->req(); k++) {
702 j -= elide_copy(n, k, block, value, regnd, two_adr != k);
703 }
704
708 continue;
709 }
710
711 // Update the register defined by this instruction
712 OptoReg::Name nreg = lrgs(lidx).reg();
713 // Skip through all copies to the _value_ being defined.
714 // Do not change from int to pointer
715 Node *val = skip_copies(n);
716
717 // Clear out a dead definition before starting so that the
718 // elimination code doesn't have to guard against it. The
719 // definition could in fact be a kill projection with a count of
720 // 0 which is safe but since those are uninteresting for copy
721 // elimination just delete them as well.
722 if (regnd[nreg] != NULL && regnd[nreg]->outcnt() == 0) {
723 regnd.map(nreg, NULL);
724 value.map(nreg, NULL);
725 }
726
727 uint n_ideal_reg = n->ideal_reg();
728 int n_regs = RegMask::num_registers(n_ideal_reg, lrgs(lidx));
729 if (n_regs == 1) {
730 // If Node 'n' does not change the value mapped by the register,
731 // then 'n' is a useless copy. Do not update the register->node
732 // mapping so 'n' will go dead.
733 if( value[nreg] != val ) {
734 if (eliminate_copy_of_constant(val, n, block, value, regnd, nreg, OptoReg::Bad)) {
735 j -= replace_and_yank_if_dead(n, nreg, block, value, regnd);
736 } else {
737 // Update the mapping: record new Node defined by the register
738 regnd.map(nreg,n);
739 // Update mapping for defined *value*, which is the defined
740 // Node after skipping all copies.
741 value.map(nreg,val);
742 }
743 } else if( !may_be_copy_of_callee(n) ) {
744 assert(n->is_Copy(), "");
745 j -= replace_and_yank_if_dead(n, nreg, block, value, regnd);
746 }
747 } else if (RegMask::is_vector(n_ideal_reg)) {
748 // If Node 'n' does not change the value mapped by the register,
|