1 | /* |
2 | * Copyright (C) 2015-2017 Apple Inc. All rights reserved. |
3 | * |
4 | * Redistribution and use in source and binary forms, with or without |
5 | * modification, are permitted provided that the following conditions |
6 | * are met: |
7 | * 1. Redistributions of source code must retain the above copyright |
8 | * notice, this list of conditions and the following disclaimer. |
9 | * 2. Redistributions in binary form must reproduce the above copyright |
10 | * notice, this list of conditions and the following disclaimer in the |
11 | * documentation and/or other materials provided with the distribution. |
12 | * |
13 | * THIS SOFTWARE IS PROVIDED BY APPLE INC. ``AS IS'' AND ANY |
14 | * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
15 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR |
16 | * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL APPLE INC. OR |
17 | * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, |
18 | * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, |
19 | * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR |
20 | * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY |
21 | * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT |
22 | * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE |
23 | * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
24 | */ |
25 | |
26 | #include "config.h" |
27 | #include "AirOptimizeBlockOrder.h" |
28 | |
29 | #if ENABLE(B3_JIT) |
30 | |
31 | #include "AirBlockWorklist.h" |
32 | #include "AirCode.h" |
33 | #include "AirInstInlines.h" |
34 | #include "AirPhaseScope.h" |
35 | #include <wtf/BubbleSort.h> |
36 | |
37 | namespace JSC { namespace B3 { namespace Air { |
38 | |
39 | namespace { |
40 | |
41 | class SortedSuccessors { |
42 | public: |
43 | SortedSuccessors() |
44 | { |
45 | } |
46 | |
47 | void append(BasicBlock* block) |
48 | { |
49 | m_successors.append(block); |
50 | } |
51 | |
52 | void process(BlockWorklist& worklist) |
53 | { |
54 | // We prefer a stable sort, and we don't want it to go off the rails if we see NaN. Also, the number |
55 | // of successors is bounded. In fact, it currently cannot be more than 2. :-) |
56 | bubbleSort( |
57 | m_successors.begin(), m_successors.end(), |
58 | [] (BasicBlock* left, BasicBlock* right) { |
59 | return left->frequency() < right->frequency(); |
60 | }); |
61 | |
62 | // Pushing the successors in ascending order of frequency ensures that the very next block we visit |
63 | // is our highest-frequency successor (unless that successor has already been visited). |
64 | for (unsigned i = 0; i < m_successors.size(); ++i) |
65 | worklist.push(m_successors[i]); |
66 | |
67 | m_successors.shrink(0); |
68 | } |
69 | |
70 | private: |
71 | Vector<BasicBlock*, 2> m_successors; |
72 | }; |
73 | |
74 | } // anonymous namespace |
75 | |
76 | Vector<BasicBlock*> blocksInOptimizedOrder(Code& code) |
77 | { |
78 | Vector<BasicBlock*> blocksInOrder; |
79 | |
80 | BlockWorklist fastWorklist; |
81 | SortedSuccessors sortedSuccessors; |
82 | SortedSuccessors sortedSlowSuccessors; |
83 | |
84 | // We expect entrypoint lowering to have already happened. |
85 | RELEASE_ASSERT(code.numEntrypoints()); |
86 | |
87 | auto appendSuccessor = [&] (const FrequentedBlock& block) { |
88 | if (block.isRare()) |
89 | sortedSlowSuccessors.append(block.block()); |
90 | else |
91 | sortedSuccessors.append(block.block()); |
92 | }; |
93 | |
94 | // For everything but the first entrypoint, we push them in order of frequency and frequency |
95 | // class. |
96 | for (unsigned i = 1; i < code.numEntrypoints(); ++i) |
97 | appendSuccessor(code.entrypoint(i)); |
98 | |
99 | // Always push the primary successor last so that it gets highest priority. |
100 | fastWorklist.push(code.entrypoint(0).block()); |
101 | |
102 | while (BasicBlock* block = fastWorklist.pop()) { |
103 | blocksInOrder.append(block); |
104 | for (FrequentedBlock& successor : block->successors()) |
105 | appendSuccessor(successor); |
106 | sortedSuccessors.process(fastWorklist); |
107 | } |
108 | |
109 | BlockWorklist slowWorklist; |
110 | sortedSlowSuccessors.process(slowWorklist); |
111 | |
112 | while (BasicBlock* block = slowWorklist.pop()) { |
113 | // We might have already processed this block. |
114 | if (fastWorklist.saw(block)) |
115 | continue; |
116 | |
117 | blocksInOrder.append(block); |
118 | for (BasicBlock* successor : block->successorBlocks()) |
119 | sortedSuccessors.append(successor); |
120 | sortedSuccessors.process(slowWorklist); |
121 | } |
122 | |
123 | ASSERT(fastWorklist.isEmpty()); |
124 | ASSERT(slowWorklist.isEmpty()); |
125 | |
126 | return blocksInOrder; |
127 | } |
128 | |
129 | void optimizeBlockOrder(Code& code) |
130 | { |
131 | PhaseScope phaseScope(code, "optimizeBlockOrder" ); |
132 | |
133 | Vector<BasicBlock*> blocksInOrder = blocksInOptimizedOrder(code); |
134 | |
135 | // Place blocks into Code's block list according to the ordering in blocksInOrder. We do this by leaking |
136 | // all of the blocks and then readopting them. |
137 | for (auto& entry : code.blockList()) |
138 | entry.release(); |
139 | |
140 | code.blockList().shrink(0); |
141 | |
142 | for (unsigned i = 0; i < blocksInOrder.size(); ++i) { |
143 | BasicBlock* block = blocksInOrder[i]; |
144 | block->setIndex(i); |
145 | code.blockList().append(std::unique_ptr<BasicBlock>(block)); |
146 | } |
147 | |
148 | // Finally, flip any branches that we recognize. It's most optimal if the taken successor does not point |
149 | // at the next block. |
150 | for (BasicBlock* block : code) { |
151 | Inst& branch = block->last(); |
152 | |
153 | // It's somewhat tempting to just say that if the block has two successors and the first arg is |
154 | // invertible, then we can do the optimization. But that's wagging the dog. The fact that an |
155 | // instruction happens to have an argument that is invertible doesn't mean it's a branch, even though |
156 | // it is true that currently only branches have invertible arguments. It's also tempting to say that |
157 | // the /branch flag in AirOpcode.opcodes tells us that something is a branch - except that there, |
158 | // /branch also means Jump. The approach taken here means that if you add new branch instructions and |
159 | // forget about this phase, then at worst your new instructions won't opt into the inversion |
160 | // optimization. You'll probably realize that as soon as you look at the disassembly, and it |
161 | // certainly won't cause any correctness issues. |
162 | |
163 | switch (branch.kind.opcode) { |
164 | case Branch8: |
165 | case Branch32: |
166 | case Branch64: |
167 | case BranchTest8: |
168 | case BranchTest32: |
169 | case BranchTest64: |
170 | case BranchFloat: |
171 | case BranchDouble: |
172 | case BranchAdd32: |
173 | case BranchAdd64: |
174 | case BranchMul32: |
175 | case BranchMul64: |
176 | case BranchSub32: |
177 | case BranchSub64: |
178 | case BranchNeg32: |
179 | case BranchNeg64: |
180 | case BranchAtomicStrongCAS8: |
181 | case BranchAtomicStrongCAS16: |
182 | case BranchAtomicStrongCAS32: |
183 | case BranchAtomicStrongCAS64: |
184 | if (code.findNextBlock(block) == block->successorBlock(0) && branch.args[0].isInvertible()) { |
185 | std::swap(block->successor(0), block->successor(1)); |
186 | branch.args[0] = branch.args[0].inverted(); |
187 | } |
188 | break; |
189 | |
190 | default: |
191 | break; |
192 | } |
193 | } |
194 | } |
195 | |
196 | } } } // namespace JSC::B3::Air |
197 | |
198 | #endif // ENABLE(B3_JIT) |
199 | |