1 /*
  2  * Copyright (c) 2018, 2025, 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 
 25 #ifndef SHARE_GC_SHARED_C2_BARRIERSETC2_HPP
 26 #define SHARE_GC_SHARED_C2_BARRIERSETC2_HPP
 27 
 28 #include "memory/allocation.hpp"
 29 #include "oops/accessDecorators.hpp"
 30 #include "opto/loopnode.hpp"
 31 #include "opto/machnode.hpp"
 32 #include "opto/matcher.hpp"
 33 #include "opto/memnode.hpp"
 34 #include "utilities/globalDefinitions.hpp"
 35 
 36 // This means the access is mismatched. This means the value of an access
 37 // is not equivalent to the value pointed to by the address.
 38 const DecoratorSet C2_MISMATCHED             = DECORATOR_LAST << 1;
 39 // The access may not be aligned to its natural size.
 40 const DecoratorSet C2_UNALIGNED              = DECORATOR_LAST << 2;
 41 // The atomic cmpxchg is weak, meaning that spurious false negatives are allowed,
 42 // but never false positives.
 43 const DecoratorSet C2_WEAK_CMPXCHG           = DECORATOR_LAST << 3;
 44 // This denotes that a load has control dependency.
 45 const DecoratorSet C2_CONTROL_DEPENDENT_LOAD = DECORATOR_LAST << 4;
 46 // This denotes that a load that must be pinned, but may float above safepoints.
 47 const DecoratorSet C2_UNKNOWN_CONTROL_LOAD   = DECORATOR_LAST << 5;
 48 // This denotes that the access is produced from the sun.misc.Unsafe intrinsics.
 49 const DecoratorSet C2_UNSAFE_ACCESS          = DECORATOR_LAST << 6;
 50 // This denotes that the access mutates state.
 51 const DecoratorSet C2_WRITE_ACCESS           = DECORATOR_LAST << 7;
 52 // This denotes that the access reads state.
 53 const DecoratorSet C2_READ_ACCESS            = DECORATOR_LAST << 8;
 54 // A nearby allocation?
 55 const DecoratorSet C2_TIGHTLY_COUPLED_ALLOC  = DECORATOR_LAST << 9;
 56 // Loads and stores from an arraycopy being optimized
 57 const DecoratorSet C2_ARRAY_COPY             = DECORATOR_LAST << 10;
 58 // Loads from immutable memory
 59 const DecoratorSet C2_IMMUTABLE_MEMORY       = DECORATOR_LAST << 11;
 60 
 61 class Compile;
 62 class ConnectionGraph;
 63 class GraphKit;
 64 class IdealKit;
 65 class Node;
 66 class PhaseGVN;
 67 class PhaseIdealLoop;
 68 class PhaseMacroExpand;
 69 class Type;
 70 class TypePtr;
 71 class Unique_Node_List;
 72 
 73 // This class wraps a node and a type.
 74 class C2AccessValue: public StackObj {
 75 protected:
 76   Node* _node;
 77   const Type* _type;
 78 
 79 public:
 80   C2AccessValue(Node* node, const Type* type) :
 81     _node(node),
 82     _type(type) {}
 83 
 84   Node* node() const        { return _node; }
 85   const Type* type() const  { return _type; }
 86 
 87   void set_node(Node* node) { _node = node; }
 88 };
 89 
 90 // This class wraps a node and a pointer type.
 91 class C2AccessValuePtr: public C2AccessValue {
 92 
 93 public:
 94   C2AccessValuePtr(Node* node, const TypePtr* type) :
 95     C2AccessValue(node, type) {}
 96 
 97   const TypePtr* type() const { return _type->is_ptr(); }
 98 };
 99 
100 // This class wraps a bunch of context parameters that are passed around in the
101 // BarrierSetC2 backend hierarchy, for loads and stores, to reduce boiler plate.
102 class C2Access: public StackObj {
103 protected:
104   DecoratorSet      _decorators;
105   Node*             _base;
106   C2AccessValuePtr& _addr;
107   Node*             _raw_access;
108   BasicType         _type;
109   uint8_t           _barrier_data;
110 
111   void fixup_decorators();
112 
113 public:
114   C2Access(DecoratorSet decorators,
115            BasicType type, Node* base, C2AccessValuePtr& addr) :
116     _decorators(decorators),
117     _base(base),
118     _addr(addr),
119     _raw_access(nullptr),
120     _type(type),
121     _barrier_data(0)
122   {}
123 
124   DecoratorSet decorators() const { return _decorators; }
125   Node* base() const              { return _base; }
126   C2AccessValuePtr& addr() const  { return _addr; }
127   BasicType type() const          { return _type; }
128   bool is_oop() const             { return is_reference_type(_type); }
129   bool is_raw() const             { return (_decorators & AS_RAW) != 0; }
130   Node* raw_access() const        { return _raw_access; }
131 
132   uint8_t barrier_data() const        { return _barrier_data; }
133   void set_barrier_data(uint8_t data) { _barrier_data = data; }
134 
135   void set_raw_access(Node* raw_access) { _raw_access = raw_access; }
136   virtual void set_memory() {} // no-op for normal accesses, but not for atomic accesses.
137 
138   MemNode::MemOrd mem_node_mo() const;
139   bool needs_cpu_membar() const;
140 
141   virtual PhaseGVN& gvn() const = 0;
142   virtual bool is_parse_access() const { return false; }
143   virtual bool is_opt_access() const { return false; }
144 };
145 
146 // C2Access for parse time calls to the BarrierSetC2 backend.
147 class C2ParseAccess: public C2Access {
148 protected:
149   GraphKit*         _kit;
150   Node* _ctl;
151   const InlineTypeNode* _vt; // For flat, atomic accesses that might require GC barriers on oop fields
152 
153   void* barrier_set_state() const;
154 
155 public:
156   C2ParseAccess(GraphKit* kit, DecoratorSet decorators,
157                 BasicType type, Node* base, C2AccessValuePtr& addr,
158                 Node* ctl = nullptr, const InlineTypeNode* vt = nullptr) :
159     C2Access(decorators, type, base, addr),
160     _kit(kit),
161     _ctl(ctl),
162     _vt (vt) {
163     fixup_decorators();
164   }
165 
166   GraphKit* kit() const           { return _kit; }
167   Node* control() const;
168   const InlineTypeNode* vt() const { return _vt; }
169 
170   virtual PhaseGVN& gvn() const;
171   virtual bool is_parse_access() const { return true; }
172 };
173 
174 // This class wraps a bunch of context parameters that are passed around in the
175 // BarrierSetC2 backend hierarchy, for atomic accesses, to reduce boiler plate.
176 class C2AtomicParseAccess: public C2ParseAccess {
177   Node* _memory;
178   uint  _alias_idx;
179 
180 public:
181   C2AtomicParseAccess(GraphKit* kit, DecoratorSet decorators, BasicType type,
182                  Node* base, C2AccessValuePtr& addr, uint alias_idx) :
183     C2ParseAccess(kit, decorators, type, base, addr),
184     _memory(nullptr),
185     _alias_idx(alias_idx) {}
186 
187   // Set the memory node based on the current memory slice.
188   virtual void set_memory();
189 
190   Node* memory() const       { return _memory; }
191   uint alias_idx() const     { return _alias_idx; }
192 };
193 
194 // C2Access for optimization time calls to the BarrierSetC2 backend.
195 class C2OptAccess: public C2Access {
196   PhaseGVN& _gvn;
197   MergeMemNode* _mem;
198   Node* _ctl;
199 
200 public:
201   C2OptAccess(PhaseGVN& gvn, Node* ctl, MergeMemNode* mem, DecoratorSet decorators,
202               BasicType type, Node* base, C2AccessValuePtr& addr) :
203     C2Access(decorators, type, base, addr),
204     _gvn(gvn), _mem(mem), _ctl(ctl) {
205     fixup_decorators();
206   }
207 
208   MergeMemNode* mem() const { return _mem; }
209   Node* ctl() const { return _ctl; }
210 
211   virtual PhaseGVN& gvn() const { return _gvn; }
212   virtual bool is_opt_access() const { return true; }
213 };
214 
215 class BarrierSetC2State : public ArenaObj {
216 protected:
217   Node_Array                      _live;
218 
219 public:
220   BarrierSetC2State(Arena* arena) : _live(arena) {}
221 
222   RegMask* live(const Node* node) {
223     if (!node->is_Mach() || !needs_liveness_data(node->as_Mach())) {
224       // Don't need liveness for non-MachNodes or if the GC doesn't request it
225       return nullptr;
226     }
227     RegMask* live = (RegMask*)_live[node->_idx];
228     if (live == nullptr) {
229       live = new (Compile::current()->comp_arena()->AmallocWords(sizeof(RegMask))) RegMask();
230       _live.map(node->_idx, (Node*)live);
231     }
232 
233     return live;
234   }
235 
236   virtual bool needs_liveness_data(const MachNode* mach) const = 0;
237   virtual bool needs_livein_data() const = 0;
238 };
239 
240 // This class represents the slow path in a C2 barrier. It is defined by a
241 // memory access, an entry point, and a continuation point (typically the end of
242 // the barrier). It provides a set of registers whose value is live across the
243 // barrier, and hence must be preserved across runtime calls from the stub.
244 class BarrierStubC2 : public ArenaObj {
245 protected:
246   const MachNode* _node;
247   Label           _entry;
248   Label           _continuation;
249   RegMask         _preserve;
250 
251   // Registers that are live-in/live-out of the entire memory access
252   // implementation (possibly including multiple barriers). Whether live-in or
253   // live-out registers are returned depends on
254   // BarrierSetC2State::needs_livein_data().
255   RegMask& live() const;
256 
257 public:
258   BarrierStubC2(const MachNode* node);
259 
260   // Entry point to the stub.
261   Label* entry();
262   // Return point from the stub (typically end of barrier).
263   Label* continuation();
264   // High-level, GC-specific barrier flags.
265   uint8_t barrier_data() const;
266 
267   // Preserve the value in reg across runtime calls in this barrier.
268   void preserve(Register reg);
269   // Do not preserve the value in reg across runtime calls in this barrier.
270   void dont_preserve(Register reg);
271   // Set of registers whose value needs to be preserved across runtime calls in this barrier.
272   const RegMask& preserve_set() const;
273 };
274 
275 // This is the top-level class for the backend of the Access API in C2.
276 // The top-level class is responsible for performing raw accesses. The
277 // various GC barrier sets inherit from the BarrierSetC2 class to sprinkle
278 // barriers into the accesses.
279 class BarrierSetC2: public CHeapObj<mtGC> {
280 protected:
281   virtual void resolve_address(C2Access& access) const;
282   virtual Node* store_at_resolved(C2Access& access, C2AccessValue& val) const;
283   virtual Node* load_at_resolved(C2Access& access, const Type* val_type) const;
284 
285   virtual Node* atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
286                                                Node* new_val, const Type* val_type) const;
287   virtual Node* atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
288                                                 Node* new_val, const Type* value_type) const;
289   virtual Node* atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* new_val, const Type* val_type) const;
290   virtual Node* atomic_add_at_resolved(C2AtomicParseAccess& access, Node* new_val, const Type* val_type) const;
291   void pin_atomic_op(C2AtomicParseAccess& access) const;
292   void clone_in_runtime(PhaseMacroExpand* phase, ArrayCopyNode* ac,
293                         address call_addr, const char* call_name) const;
294 
295 public:
296   // This is the entry-point for the backend to perform accesses through the Access API.
297   virtual Node* store_at(C2Access& access, C2AccessValue& val) const;
298   virtual Node* load_at(C2Access& access, const Type* val_type) const;
299 
300   virtual Node* atomic_cmpxchg_val_at(C2AtomicParseAccess& access, Node* expected_val,
301                                       Node* new_val, const Type* val_type) const;
302   virtual Node* atomic_cmpxchg_bool_at(C2AtomicParseAccess& access, Node* expected_val,
303                                        Node* new_val, const Type* val_type) const;
304   virtual Node* atomic_xchg_at(C2AtomicParseAccess& access, Node* new_val, const Type* value_type) const;
305   virtual Node* atomic_add_at(C2AtomicParseAccess& access, Node* new_val, const Type* value_type) const;
306 
307   virtual void clone(GraphKit* kit, Node* src_base, Node* dst_base, Node* size, bool is_array) const;
308 
309   virtual Node* obj_allocate(PhaseMacroExpand* macro, Node* mem, Node* toobig_false, Node* size_in_bytes,
310                              Node*& i_o, Node*& needgc_ctrl,
311                              Node*& fast_oop_ctrl, Node*& fast_oop_rawmem,
312                              intx prefetch_lines) const;
313 
314   virtual Node* ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const { return nullptr; }
315 
316   // These are general helper methods used by C2
317   enum ArrayCopyPhase {
318     Parsing,
319     Optimization,
320     Expansion
321   };
322 
323   virtual bool array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, bool is_clone_instance, ArrayCopyPhase phase) const { return false; }
324   virtual void clone_at_expansion(PhaseMacroExpand* phase, ArrayCopyNode* ac) const;
325 
326   // Support for GC barriers emitted during parsing
327   virtual bool has_load_barrier_nodes() const { return false; }
328   virtual bool is_gc_pre_barrier_node(Node* node) const { return false; }
329   virtual bool is_gc_barrier_node(Node* node) const { return false; }
330   virtual Node* step_over_gc_barrier(Node* c) const { return c; }
331 
332   // Support for macro expanded GC barriers
333   virtual void register_potential_barrier_node(Node* node) const { }
334   virtual void unregister_potential_barrier_node(Node* node) const { }
335   virtual void eliminate_gc_barrier(PhaseIterGVN* igvn, Node* node) const { }
336   virtual void eliminate_gc_barrier_data(Node* node) const { }
337   virtual void enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const {}
338   virtual void eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const {}
339 
340   // Allow barrier sets to have shared state that is preserved across a compilation unit.
341   // This could for example comprise macro nodes to be expanded during macro expansion.
342   virtual void* create_barrier_state(Arena* comp_arena) const { return nullptr; }
343   // If the BarrierSetC2 state has barrier nodes in its compilation
344   // unit state to be expanded later, then now is the time to do so.
345   virtual bool expand_barriers(Compile* C, PhaseIterGVN& igvn) const { return false; }
346   virtual bool optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const { return false; }
347   virtual bool strip_mined_loops_expanded(LoopOptsMode mode) const { return false; }
348   virtual bool is_gc_specific_loop_opts_pass(LoopOptsMode mode) const { return false; }
349   // Estimated size of the node barrier in number of C2 Ideal nodes.
350   // This is used to guide heuristics in C2, e.g. whether to unroll a loop.
351   virtual uint estimated_barrier_size(const Node* node) const { return 0; }
352   // Whether the given store can be used to initialize a newly allocated object.
353   virtual bool can_initialize_object(const StoreNode* store) const { return true; }
354 
355   enum CompilePhase {
356     BeforeOptimize,
357     BeforeMacroExpand,
358     BeforeCodeGen
359   };
360 
361 #ifdef ASSERT
362   virtual void verify_gc_barriers(Compile* compile, CompilePhase phase) const {}
363 #endif
364 
365   virtual bool final_graph_reshaping(Compile* compile, Node* n, uint opcode, Unique_Node_List& dead_nodes) const { return false; }
366 
367   virtual bool escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const { return false; }
368   virtual bool escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const { return false; }
369   virtual bool escape_has_out_with_unsafe_object(Node* n) const { return false; }
370 
371   virtual bool matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const { return false; };
372   virtual bool matcher_is_store_load_barrier(Node* x, uint xop) const { return false; }
373 
374   // Whether the given phi node joins OOPs from fast and slow allocation paths.
375   static bool is_allocation(const Node* node);
376   // Elide GC barriers from a Mach node according to elide_dominated_barriers().
377   virtual void elide_dominated_barrier(MachNode* mach) const { }
378   // Elide GC barriers from instructions in 'accesses' if they are dominated by
379   // instructions in 'access_dominators' (according to elide_mach_barrier()) and
380   // there is no safepoint poll in between.
381   void elide_dominated_barriers(Node_List& accesses, Node_List& access_dominators) const;
382   virtual void late_barrier_analysis() const { }
383   virtual void compute_liveness_at_stubs() const;
384   virtual int estimate_stub_size() const { return 0; }
385   virtual void emit_stubs(CodeBuffer& cb) const { }
386 
387   static int arraycopy_payload_base_offset(bool is_array);
388 
389 #ifndef PRODUCT
390   virtual void dump_barrier_data(const MachNode* mach, outputStream* st) const {
391     st->print("%x", mach->barrier_data());
392   };
393 #endif
394 };
395 
396 #endif // SHARE_GC_SHARED_C2_BARRIERSETC2_HPP