1 /*
  2  * Copyright (c) 2015, 2021, 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 
 24 #include "precompiled.hpp"
 25 #include "classfile/javaClasses.hpp"
 26 #include "gc/z/c2/zBarrierSetC2.hpp"
 27 #include "gc/z/zBarrierSet.hpp"
 28 #include "gc/z/zBarrierSetAssembler.hpp"
 29 #include "gc/z/zBarrierSetRuntime.hpp"
 30 #include "opto/arraycopynode.hpp"
 31 #include "opto/addnode.hpp"
 32 #include "opto/block.hpp"
 33 #include "opto/compile.hpp"
 34 #include "opto/graphKit.hpp"
 35 #include "opto/machnode.hpp"
 36 #include "opto/macro.hpp"
 37 #include "opto/memnode.hpp"
 38 #include "opto/node.hpp"
 39 #include "opto/output.hpp"
 40 #include "opto/regalloc.hpp"
 41 #include "opto/rootnode.hpp"
 42 #include "opto/runtime.hpp"
 43 #include "opto/type.hpp"
 44 #include "utilities/growableArray.hpp"
 45 #include "utilities/macros.hpp"
 46 
 47 class ZBarrierSetC2State : public ResourceObj {
 48 private:
 49   GrowableArray<ZLoadBarrierStubC2*>* _stubs;
 50   Node_Array                          _live;
 51 
 52 public:
 53   ZBarrierSetC2State(Arena* arena) :
 54     _stubs(new (arena) GrowableArray<ZLoadBarrierStubC2*>(arena, 8,  0, NULL)),
 55     _live(arena) {}
 56 
 57   GrowableArray<ZLoadBarrierStubC2*>* stubs() {
 58     return _stubs;
 59   }
 60 
 61   RegMask* live(const Node* node) {
 62     if (!node->is_Mach()) {
 63       // Don't need liveness for non-MachNodes
 64       return NULL;
 65     }
 66 
 67     const MachNode* const mach = node->as_Mach();
 68     if (mach->barrier_data() == ZLoadBarrierElided) {
 69       // Don't need liveness data for nodes without barriers
 70       return NULL;
 71     }
 72 
 73     RegMask* live = (RegMask*)_live[node->_idx];
 74     if (live == NULL) {
 75       live = new (Compile::current()->comp_arena()->AmallocWords(sizeof(RegMask))) RegMask();
 76       _live.map(node->_idx, (Node*)live);
 77     }
 78 
 79     return live;
 80   }
 81 };
 82 
 83 static ZBarrierSetC2State* barrier_set_state() {
 84   return reinterpret_cast<ZBarrierSetC2State*>(Compile::current()->barrier_set_state());
 85 }
 86 
 87 ZLoadBarrierStubC2* ZLoadBarrierStubC2::create(const MachNode* node, Address ref_addr, Register ref, Register tmp, uint8_t barrier_data) {
 88   ZLoadBarrierStubC2* const stub = new (Compile::current()->comp_arena()) ZLoadBarrierStubC2(node, ref_addr, ref, tmp, barrier_data);
 89   if (!Compile::current()->output()->in_scratch_emit_size()) {
 90     barrier_set_state()->stubs()->append(stub);
 91   }
 92 
 93   return stub;
 94 }
 95 
 96 ZLoadBarrierStubC2::ZLoadBarrierStubC2(const MachNode* node, Address ref_addr, Register ref, Register tmp, uint8_t barrier_data) :
 97     _node(node),
 98     _ref_addr(ref_addr),
 99     _ref(ref),
100     _tmp(tmp),
101     _barrier_data(barrier_data),
102     _entry(),
103     _continuation() {
104   assert_different_registers(ref, ref_addr.base());
105   assert_different_registers(ref, ref_addr.index());
106 }
107 
108 Address ZLoadBarrierStubC2::ref_addr() const {
109   return _ref_addr;
110 }
111 
112 Register ZLoadBarrierStubC2::ref() const {
113   return _ref;
114 }
115 
116 Register ZLoadBarrierStubC2::tmp() const {
117   return _tmp;
118 }
119 
120 address ZLoadBarrierStubC2::slow_path() const {
121   DecoratorSet decorators = DECORATORS_NONE;
122   if (_barrier_data & ZLoadBarrierStrong) {
123     decorators |= ON_STRONG_OOP_REF;
124   }
125   if (_barrier_data & ZLoadBarrierWeak) {
126     decorators |= ON_WEAK_OOP_REF;
127   }
128   if (_barrier_data & ZLoadBarrierPhantom) {
129     decorators |= ON_PHANTOM_OOP_REF;
130   }
131   if (_barrier_data & ZLoadBarrierNoKeepalive) {
132     decorators |= AS_NO_KEEPALIVE;
133   }
134   return ZBarrierSetRuntime::load_barrier_on_oop_field_preloaded_addr(decorators);
135 }
136 
137 RegMask& ZLoadBarrierStubC2::live() const {
138   return *barrier_set_state()->live(_node);
139 }
140 
141 Label* ZLoadBarrierStubC2::entry() {
142   // The _entry will never be bound when in_scratch_emit_size() is true.
143   // However, we still need to return a label that is not bound now, but
144   // will eventually be bound. Any lable will do, as it will only act as
145   // a placeholder, so we return the _continuation label.
146   return Compile::current()->output()->in_scratch_emit_size() ? &_continuation : &_entry;
147 }
148 
149 Label* ZLoadBarrierStubC2::continuation() {
150   return &_continuation;
151 }
152 
153 void* ZBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
154   return new (comp_arena) ZBarrierSetC2State(comp_arena);
155 }
156 
157 void ZBarrierSetC2::late_barrier_analysis() const {
158   analyze_dominating_barriers();
159   compute_liveness_at_stubs();
160 }
161 
162 void ZBarrierSetC2::emit_stubs(CodeBuffer& cb) const {
163   MacroAssembler masm(&cb);
164   GrowableArray<ZLoadBarrierStubC2*>* const stubs = barrier_set_state()->stubs();
165 
166   for (int i = 0; i < stubs->length(); i++) {
167     // Make sure there is enough space in the code buffer
168     if (cb.insts()->maybe_expand_to_ensure_remaining(PhaseOutput::MAX_inst_size) && cb.blob() == NULL) {
169       ciEnv::current()->record_failure("CodeCache is full");
170       return;
171     }
172 
173     ZBarrierSet::assembler()->generate_c2_load_barrier_stub(&masm, stubs->at(i));
174   }
175 
176   masm.flush();
177 }
178 
179 int ZBarrierSetC2::estimate_stub_size() const {
180   Compile* const C = Compile::current();
181   BufferBlob* const blob = C->output()->scratch_buffer_blob();
182   GrowableArray<ZLoadBarrierStubC2*>* const stubs = barrier_set_state()->stubs();
183   int size = 0;
184 
185   for (int i = 0; i < stubs->length(); i++) {
186     CodeBuffer cb(blob->content_begin(), (address)C->output()->scratch_locs_memory() - blob->content_begin());
187     MacroAssembler masm(&cb);
188     ZBarrierSet::assembler()->generate_c2_load_barrier_stub(&masm, stubs->at(i));
189     size += cb.insts_size();
190   }
191 
192   return size;
193 }
194 
195 static void set_barrier_data(C2Access& access) {
196   if (ZBarrierSet::barrier_needed(access.decorators(), access.type())) {
197     uint8_t barrier_data = 0;
198 
199     if (access.decorators() & ON_PHANTOM_OOP_REF) {
200       barrier_data |= ZLoadBarrierPhantom;
201     } else if (access.decorators() & ON_WEAK_OOP_REF) {
202       barrier_data |= ZLoadBarrierWeak;
203     } else {
204       barrier_data |= ZLoadBarrierStrong;
205     }
206 
207     if (access.decorators() & AS_NO_KEEPALIVE) {
208       barrier_data |= ZLoadBarrierNoKeepalive;
209     }
210 
211     access.set_barrier_data(barrier_data);
212   }
213 }
214 
215 Node* ZBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
216   set_barrier_data(access);
217   return BarrierSetC2::load_at_resolved(access, val_type);
218 }
219 
220 Node* ZBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
221                                                     Node* new_val, const Type* val_type) const {
222   set_barrier_data(access);
223   return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, val_type);
224 }
225 
226 Node* ZBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
227                                                      Node* new_val, const Type* value_type) const {
228   set_barrier_data(access);
229   return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
230 }
231 
232 Node* ZBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* new_val, const Type* val_type) const {
233   set_barrier_data(access);
234   return BarrierSetC2::atomic_xchg_at_resolved(access, new_val, val_type);
235 }
236 
237 bool ZBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type,
238                                                     bool is_clone, bool is_clone_instance,
239                                                     ArrayCopyPhase phase) const {
240   if (phase == ArrayCopyPhase::Parsing) {
241     return false;
242   }
243   if (phase == ArrayCopyPhase::Optimization) {
244     return is_clone_instance;
245   }
246   // else ArrayCopyPhase::Expansion
247   return type == T_OBJECT || type == T_ARRAY;
248 }
249 
250 // This TypeFunc assumes a 64bit system
251 static const TypeFunc* clone_type() {
252   // Create input type (domain)
253   const Type** domain_fields = TypeTuple::fields(4);
254   domain_fields[TypeFunc::Parms + 0] = TypeInstPtr::NOTNULL;  // src
255   domain_fields[TypeFunc::Parms + 1] = TypeInstPtr::NOTNULL;  // dst
256   domain_fields[TypeFunc::Parms + 2] = TypeLong::LONG;        // size lower
257   domain_fields[TypeFunc::Parms + 3] = Type::HALF;            // size upper
258   const TypeTuple* domain = TypeTuple::make(TypeFunc::Parms + 4, domain_fields);
259 
260   // Create result type (range)
261   const Type** range_fields = TypeTuple::fields(0);
262   const TypeTuple* range = TypeTuple::make(TypeFunc::Parms + 0, range_fields);
263 
264   return TypeFunc::make(domain, range);
265 }
266 
267 #define XTOP LP64_ONLY(COMMA phase->top())
268 
269 void ZBarrierSetC2::clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const {
270   Node* const src = ac->in(ArrayCopyNode::Src);
271   const TypeAryPtr* ary_ptr = src->get_ptr_type()->isa_aryptr();
272 
273   if (ac->is_clone_array() && ary_ptr != NULL) {
274     BasicType bt = ary_ptr->elem()->array_element_basic_type();
275     if (is_reference_type(bt)) {
276       // Clone object array
277       bt = T_OBJECT;
278     } else {
279       // Clone primitive array
280       bt = T_LONG;
281     }
282 
283     Node* ctrl = ac->in(TypeFunc::Control);
284     Node* mem = ac->in(TypeFunc::Memory);
285     Node* src = ac->in(ArrayCopyNode::Src);
286     Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
287     Node* dest = ac->in(ArrayCopyNode::Dest);
288     Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
289     Node* length = ac->in(ArrayCopyNode::Length);
290 
291     if (bt == T_OBJECT) {
292       // BarrierSetC2::clone sets the offsets via BarrierSetC2::arraycopy_payload_base_offset
293       // which 8-byte aligns them to allow for word size copies. Make sure the offsets point
294       // to the first element in the array when cloning object arrays. Otherwise, load
295       // barriers are applied to parts of the header. Also adjust the length accordingly.
296       assert(src_offset == dest_offset, "should be equal");
297       jlong offset = src_offset->get_long();
298       if (offset != arrayOopDesc::base_offset_in_bytes(T_OBJECT)) {
299         assert(!UseCompressedClassPointers, "should only happen without compressed class pointers");
300         assert((arrayOopDesc::base_offset_in_bytes(T_OBJECT) - offset) == BytesPerLong, "unexpected offset");
301         length = phase->transform_later(new SubLNode(length, phase->longcon(1))); // Size is in longs
302         src_offset = phase->longcon(arrayOopDesc::base_offset_in_bytes(T_OBJECT));
303         dest_offset = src_offset;
304       }
305     }
306     Node* payload_src = phase->basic_plus_adr(src, src_offset);
307     Node* payload_dst = phase->basic_plus_adr(dest, dest_offset);
308 
309     const char* copyfunc_name = "arraycopy";
310     address     copyfunc_addr = phase->basictype2arraycopy(bt, NULL, NULL, true, copyfunc_name, true);
311 
312     const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
313     const TypeFunc* call_type = OptoRuntime::fast_arraycopy_Type();
314 
315     Node* call = phase->make_leaf_call(ctrl, mem, call_type, copyfunc_addr, copyfunc_name, raw_adr_type, payload_src, payload_dst, length XTOP);
316     phase->transform_later(call);
317 
318     phase->igvn().replace_node(ac, call);
319     return;
320   }
321 
322   // Clone instance
323   Node* const ctrl       = ac->in(TypeFunc::Control);
324   Node* const mem        = ac->in(TypeFunc::Memory);
325   Node* const dst        = ac->in(ArrayCopyNode::Dest);
326   Node* const size       = ac->in(ArrayCopyNode::Length);
327 
328   assert(size->bottom_type()->is_long(), "Should be long");
329 
330   // The native clone we are calling here expects the instance size in words
331   // Add header/offset size to payload size to get instance size.
332   Node* const base_offset = phase->longcon(arraycopy_payload_base_offset(ac->is_clone_array()) >> LogBytesPerLong);
333   Node* const full_size = phase->transform_later(new AddLNode(size, base_offset));
334 
335   Node* const call = phase->make_leaf_call(ctrl,
336                                            mem,
337                                            clone_type(),
338                                            ZBarrierSetRuntime::clone_addr(),
339                                            "ZBarrierSetRuntime::clone",
340                                            TypeRawPtr::BOTTOM,
341                                            src,
342                                            dst,
343                                            full_size,
344                                            phase->top());
345   phase->transform_later(call);
346   phase->igvn().replace_node(ac, call);
347 }
348 
349 #undef XTOP
350 
351 // == Dominating barrier elision ==
352 
353 static bool block_has_safepoint(const Block* block, uint from, uint to) {
354   for (uint i = from; i < to; i++) {
355     if (block->get_node(i)->is_MachSafePoint()) {
356       // Safepoint found
357       return true;
358     }
359   }
360 
361   // Safepoint not found
362   return false;
363 }
364 
365 static bool block_has_safepoint(const Block* block) {
366   return block_has_safepoint(block, 0, block->number_of_nodes());
367 }
368 
369 static uint block_index(const Block* block, const Node* node) {
370   for (uint j = 0; j < block->number_of_nodes(); ++j) {
371     if (block->get_node(j) == node) {
372       return j;
373     }
374   }
375   ShouldNotReachHere();
376   return 0;
377 }
378 
379 void ZBarrierSetC2::analyze_dominating_barriers() const {
380   ResourceMark rm;
381   Compile* const C = Compile::current();
382   PhaseCFG* const cfg = C->cfg();
383   Block_List worklist;
384   Node_List mem_ops;
385   Node_List barrier_loads;
386 
387   // Step 1 - Find accesses, and track them in lists
388   for (uint i = 0; i < cfg->number_of_blocks(); ++i) {
389     const Block* const block = cfg->get_block(i);
390     for (uint j = 0; j < block->number_of_nodes(); ++j) {
391       const Node* const node = block->get_node(j);
392       if (!node->is_Mach()) {
393         continue;
394       }
395 
396       MachNode* const mach = node->as_Mach();
397       switch (mach->ideal_Opcode()) {
398       case Op_LoadP:
399         if ((mach->barrier_data() & ZLoadBarrierStrong) != 0) {
400           barrier_loads.push(mach);
401         }
402         if ((mach->barrier_data() & (ZLoadBarrierStrong | ZLoadBarrierNoKeepalive)) ==
403             ZLoadBarrierStrong) {
404           mem_ops.push(mach);
405         }
406         break;
407       case Op_CompareAndExchangeP:
408       case Op_CompareAndSwapP:
409       case Op_GetAndSetP:
410         if ((mach->barrier_data() & ZLoadBarrierStrong) != 0) {
411           barrier_loads.push(mach);
412         }
413       case Op_StoreP:
414         mem_ops.push(mach);
415         break;
416 
417       default:
418         break;
419       }
420     }
421   }
422 
423   // Step 2 - Find dominating accesses for each load
424   for (uint i = 0; i < barrier_loads.size(); i++) {
425     MachNode* const load = barrier_loads.at(i)->as_Mach();
426     const TypePtr* load_adr_type = NULL;
427     intptr_t load_offset = 0;
428     const Node* const load_obj = load->get_base_and_disp(load_offset, load_adr_type);
429     Block* const load_block = cfg->get_block_for_node(load);
430     const uint load_index = block_index(load_block, load);
431 
432     for (uint j = 0; j < mem_ops.size(); j++) {
433       MachNode* mem = mem_ops.at(j)->as_Mach();
434       const TypePtr* mem_adr_type = NULL;
435       intptr_t mem_offset = 0;
436       const Node* mem_obj = mem->get_base_and_disp(mem_offset, mem_adr_type);
437       Block* mem_block = cfg->get_block_for_node(mem);
438       uint mem_index = block_index(mem_block, mem);
439 
440       if (load_obj == NodeSentinel || mem_obj == NodeSentinel ||
441           load_obj == NULL || mem_obj == NULL ||
442           load_offset < 0 || mem_offset < 0) {
443         continue;
444       }
445 
446       if (mem_obj != load_obj || mem_offset != load_offset) {
447         // Not the same addresses, not a candidate
448         continue;
449       }
450 
451       if (load_block == mem_block) {
452         // Earlier accesses in the same block
453         if (mem_index < load_index && !block_has_safepoint(mem_block, mem_index + 1, load_index)) {
454           load->set_barrier_data(ZLoadBarrierElided);
455         }
456       } else if (mem_block->dominates(load_block)) {
457         // Dominating block? Look around for safepoints
458         ResourceMark rm;
459         Block_List stack;
460         VectorSet visited;
461         stack.push(load_block);
462         bool safepoint_found = block_has_safepoint(load_block);
463         while (!safepoint_found && stack.size() > 0) {
464           Block* block = stack.pop();
465           if (visited.test_set(block->_pre_order)) {
466             continue;
467           }
468           if (block_has_safepoint(block)) {
469             safepoint_found = true;
470             break;
471           }
472           if (block == mem_block) {
473             continue;
474           }
475 
476           // Push predecessor blocks
477           for (uint p = 1; p < block->num_preds(); ++p) {
478             Block* pred = cfg->get_block_for_node(block->pred(p));
479             stack.push(pred);
480           }
481         }
482 
483         if (!safepoint_found) {
484           load->set_barrier_data(ZLoadBarrierElided);
485         }
486       }
487     }
488   }
489 }
490 
491 // == Reduced spilling optimization ==
492 
493 void ZBarrierSetC2::compute_liveness_at_stubs() const {
494   ResourceMark rm;
495   Compile* const C = Compile::current();
496   Arena* const A = Thread::current()->resource_area();
497   PhaseCFG* const cfg = C->cfg();
498   PhaseRegAlloc* const regalloc = C->regalloc();
499   RegMask* const live = NEW_ARENA_ARRAY(A, RegMask, cfg->number_of_blocks() * sizeof(RegMask));
500   ZBarrierSetAssembler* const bs = ZBarrierSet::assembler();
501   Block_List worklist;
502 
503   for (uint i = 0; i < cfg->number_of_blocks(); ++i) {
504     new ((void*)(live + i)) RegMask();
505     worklist.push(cfg->get_block(i));
506   }
507 
508   while (worklist.size() > 0) {
509     const Block* const block = worklist.pop();
510     RegMask& old_live = live[block->_pre_order];
511     RegMask new_live;
512 
513     // Initialize to union of successors
514     for (uint i = 0; i < block->_num_succs; i++) {
515       const uint succ_id = block->_succs[i]->_pre_order;
516       new_live.OR(live[succ_id]);
517     }
518 
519     // Walk block backwards, computing liveness
520     for (int i = block->number_of_nodes() - 1; i >= 0; --i) {
521       const Node* const node = block->get_node(i);
522 
523       // Remove def bits
524       const OptoReg::Name first = bs->refine_register(node, regalloc->get_reg_first(node));
525       const OptoReg::Name second = bs->refine_register(node, regalloc->get_reg_second(node));
526       if (first != OptoReg::Bad) {
527         new_live.Remove(first);
528       }
529       if (second != OptoReg::Bad) {
530         new_live.Remove(second);
531       }
532 
533       // Add use bits
534       for (uint j = 1; j < node->req(); ++j) {
535         const Node* const use = node->in(j);
536         const OptoReg::Name first = bs->refine_register(use, regalloc->get_reg_first(use));
537         const OptoReg::Name second = bs->refine_register(use, regalloc->get_reg_second(use));
538         if (first != OptoReg::Bad) {
539           new_live.Insert(first);
540         }
541         if (second != OptoReg::Bad) {
542           new_live.Insert(second);
543         }
544       }
545 
546       // If this node tracks liveness, update it
547       RegMask* const regs = barrier_set_state()->live(node);
548       if (regs != NULL) {
549         regs->OR(new_live);
550       }
551     }
552 
553     // Now at block top, see if we have any changes
554     new_live.SUBTRACT(old_live);
555     if (new_live.is_NotEmpty()) {
556       // Liveness has refined, update and propagate to prior blocks
557       old_live.OR(new_live);
558       for (uint i = 1; i < block->num_preds(); ++i) {
559         Block* const pred = cfg->get_block_for_node(block->pred(i));
560         worklist.push(pred);
561       }
562     }
563   }
564 }