1 /*
2  * Copyright (C) 2016 The Android Open Source Project
3  *
4  * Licensed under the Apache License, Version 2.0 (the "License");
5  * you may not use this file except in compliance with the License.
6  * You may obtain a copy of the License at
7  *
8  *      http://www.apache.org/licenses/LICENSE-2.0
9  *
10  * Unless required by applicable law or agreed to in writing, software
11  * distributed under the License is distributed on an "AS IS" BASIS,
12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13  * See the License for the specific language governing permissions and
14  * limitations under the License.
15  */
16 
17 #ifndef ART_COMPILER_OPTIMIZING_REGISTER_ALLOCATOR_GRAPH_COLOR_H_
18 #define ART_COMPILER_OPTIMIZING_REGISTER_ALLOCATOR_GRAPH_COLOR_H_
19 
20 #include "arch/instruction_set.h"
21 #include "base/arena_object.h"
22 #include "base/array_ref.h"
23 #include "base/macros.h"
24 #include "base/scoped_arena_containers.h"
25 #include "register_allocator.h"
26 
27 namespace art {
28 
29 class CodeGenerator;
30 class HBasicBlock;
31 class HGraph;
32 class HInstruction;
33 class HParallelMove;
34 class Location;
35 class SsaLivenessAnalysis;
36 class InterferenceNode;
37 struct CoalesceOpportunity;
38 enum class CoalesceKind;
39 
40 /**
41  * A graph coloring register allocator.
42  *
43  * The algorithm proceeds as follows:
44  * (1) Build an interference graph, where nodes represent live intervals, and edges represent
45  *     interferences between two intervals. Coloring this graph with k colors is isomorphic to
46  *     finding a valid register assignment with k registers.
47  * (2) To color the graph, first prune all nodes with degree less than k, since these nodes are
48  *     guaranteed a color. (No matter how we color their adjacent nodes, we can give them a
49  *     different color.) As we prune nodes from the graph, more nodes may drop below degree k,
50  *     enabling further pruning. The key is to maintain the pruning order in a stack, so that we
51  *     can color the nodes in the reverse order.
52  *     When there are no more nodes with degree less than k, we start pruning alternate nodes based
53  *     on heuristics. Since these nodes are not guaranteed a color, we are careful to
54  *     prioritize nodes that require a register. We also prioritize short intervals, because
55  *     short intervals cannot be split very much if coloring fails (see below). "Prioritizing"
56  *     a node amounts to pruning it later, since it will have fewer interferences if we prune other
57  *     nodes first.
58  * (3) We color nodes in the reverse order in which we pruned them. If we cannot assign
59  *     a node a color, we do one of two things:
60  *     - If the node requires a register, we consider the current coloring attempt a failure.
61  *       However, we split the node's live interval in order to make the interference graph
62  *       sparser, so that future coloring attempts may succeed.
63  *     - If the node does not require a register, we simply assign it a location on the stack.
64  *
65  * If iterative move coalescing is enabled, the algorithm also attempts to conservatively
66  * combine nodes in the graph that would prefer to have the same color. (For example, the output
67  * of a phi instruction would prefer to have the same register as at least one of its inputs.)
68  * There are several additional steps involved with this:
69  * - We look for coalesce opportunities by examining each live interval, a step similar to that
70  *   used by linear scan when looking for register hints.
71  * - When pruning the graph, we maintain a worklist of coalesce opportunities, as well as a worklist
72  *   of low degree nodes that have associated coalesce opportunities. Only when we run out of
73  *   coalesce opportunities do we start pruning coalesce-associated nodes.
74  * - When pruning a node, if any nodes transition from high degree to low degree, we add
75  *   associated coalesce opportunities to the worklist, since these opportunities may now succeed.
76  * - Whether two nodes can be combined is decided by two different heuristics--one used when
77  *   coalescing uncolored nodes, and one used for coalescing an uncolored node with a colored node.
78  *   It is vital that we only combine two nodes if the node that remains is guaranteed to receive
79  *   a color. This is because additionally spilling is more costly than failing to coalesce.
80  * - Even if nodes are not coalesced while pruning, we keep the coalesce opportunities around
81  *   to be used as last-chance register hints when coloring. If nothing else, we try to use
82  *   caller-save registers before callee-save registers.
83  *
84  * A good reference for graph coloring register allocation is
85  * "Modern Compiler Implementation in Java" (Andrew W. Appel, 2nd Edition).
86  */
87 class RegisterAllocatorGraphColor : public RegisterAllocator {
88  public:
89   RegisterAllocatorGraphColor(ScopedArenaAllocator* allocator,
90                               CodeGenerator* codegen,
91                               const SsaLivenessAnalysis& analysis,
92                               bool iterative_move_coalescing = true);
93   ~RegisterAllocatorGraphColor() override;
94 
95   void AllocateRegisters() override;
96 
97   bool Validate(bool log_fatal_on_failure) override;
98 
99  private:
100   // Collect all intervals and prepare for register allocation.
101   void ProcessInstructions();
102   void ProcessInstruction(HInstruction* instruction);
103 
104   // If any inputs require specific registers, block those registers
105   // at the position of this instruction.
106   void CheckForFixedInputs(HInstruction* instruction);
107 
108   // If the output of an instruction requires a specific register, split
109   // the interval and assign the register to the first part.
110   void CheckForFixedOutput(HInstruction* instruction);
111 
112   // Add all applicable safepoints to a live interval.
113   // Currently depends on instruction processing order.
114   void AddSafepointsFor(HInstruction* instruction);
115 
116   // Collect all live intervals associated with the temporary locations
117   // needed by an instruction.
118   void CheckForTempLiveIntervals(HInstruction* instruction);
119 
120   // If a safe point is needed, add a synthesized interval to later record
121   // the number of live registers at this point.
122   void CheckForSafepoint(HInstruction* instruction);
123 
124   // Split an interval, but only if `position` is inside of `interval`.
125   // Return either the new interval, or the original interval if not split.
126   static LiveInterval* TrySplit(LiveInterval* interval, size_t position);
127 
128   // To ensure every graph can be colored, split live intervals
129   // at their register defs and uses. This creates short intervals with low
130   // degree in the interference graph, which are prioritized during graph
131   // coloring.
132   void SplitAtRegisterUses(LiveInterval* interval);
133 
134   // If the given instruction is a catch phi, give it a spill slot.
135   void AllocateSpillSlotForCatchPhi(HInstruction* instruction);
136 
137   // Ensure that the given register cannot be allocated for a given range.
138   void BlockRegister(Location location, size_t start, size_t end);
139   void BlockRegisters(size_t start, size_t end, bool caller_save_only = false);
140 
141   bool IsCallerSave(size_t reg, bool processing_core_regs);
142 
143   // Assigns stack slots to a list of intervals, ensuring that interfering intervals are not
144   // assigned the same stack slot.
145   void ColorSpillSlots(ArrayRef<LiveInterval* const> nodes, /* out */ size_t* num_stack_slots_used);
146 
147   // Provide stack slots to nodes that need them.
148   void AllocateSpillSlots(ArrayRef<InterferenceNode* const> nodes);
149 
150   // Whether iterative move coalescing should be performed. Iterative move coalescing
151   // improves code quality, but increases compile time.
152   const bool iterative_move_coalescing_;
153 
154   // Live intervals, split by kind (core and floating point).
155   // These should not contain high intervals, as those are represented by
156   // the corresponding low interval throughout register allocation.
157   ScopedArenaVector<LiveInterval*> core_intervals_;
158   ScopedArenaVector<LiveInterval*> fp_intervals_;
159 
160   // Intervals for temporaries, saved for special handling in the resolution phase.
161   ScopedArenaVector<LiveInterval*> temp_intervals_;
162 
163   // Safepoints, saved for special handling while processing instructions.
164   ScopedArenaVector<HInstruction*> safepoints_;
165 
166   // Interference nodes representing specific registers. These are "pre-colored" nodes
167   // in the interference graph.
168   ScopedArenaVector<InterferenceNode*> physical_core_nodes_;
169   ScopedArenaVector<InterferenceNode*> physical_fp_nodes_;
170 
171   // Allocated stack slot counters.
172   size_t num_int_spill_slots_;
173   size_t num_double_spill_slots_;
174   size_t num_float_spill_slots_;
175   size_t num_long_spill_slots_;
176   size_t catch_phi_spill_slot_counter_;
177 
178   // Number of stack slots needed for the pointer to the current method.
179   // This is 1 for 32-bit architectures, and 2 for 64-bit architectures.
180   const size_t reserved_art_method_slots_;
181 
182   // Number of stack slots needed for outgoing arguments.
183   const size_t reserved_out_slots_;
184 
185   friend class ColoringIteration;
186 
187   DISALLOW_COPY_AND_ASSIGN(RegisterAllocatorGraphColor);
188 };
189 
190 }  // namespace art
191 
192 #endif  // ART_COMPILER_OPTIMIZING_REGISTER_ALLOCATOR_GRAPH_COLOR_H_
193