1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
|
// Copyright 2022 the V8 project authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
#include "src/compiler/turboshaft/decompression-optimization.h"
#include "src/base/v8-fallthrough.h"
#include "src/codegen/machine-type.h"
#include "src/compiler/turboshaft/operations.h"
#include "src/compiler/turboshaft/optimization-phase.h"
namespace v8::internal::compiler::turboshaft {
namespace {
// Analyze the uses of values to determine if a compressed value has any uses
// that need it to be decompressed. Since this analysis looks at uses, we
// iterate the graph backwards, updating the analysis state for the inputs of an
// operation. Due to loop phis, we need to compute a fixed-point. Therefore, we
// re-visit the loop if a loop phi backedge changes something. As a performance
// optimization, we keep track of operations (`candidates`) that need to be
// updated potentially, so that we don't have to walk the whole graph again.
struct DecompressionAnalyzer : AnalyzerBase {
using Base = AnalyzerBase;
// We use `uint8_t` instead of `bool` here to avoid the bitvector optimization
// of std::vector.
FixedSidetable<uint8_t> needs_decompression;
ZoneVector<OpIndex> candidates;
DecompressionAnalyzer(const Graph& graph, Zone* phase_zone)
: AnalyzerBase(graph, phase_zone),
needs_decompression(graph.op_id_count(), phase_zone),
candidates(phase_zone) {
candidates.reserve(graph.op_id_count() / 8);
}
void Run() {
for (uint32_t next_block_id = graph.block_count() - 1; next_block_id > 0;) {
BlockIndex block_index = BlockIndex(next_block_id);
--next_block_id;
const Block& block = graph.Get(block_index);
if (block.IsLoop()) {
ProcessBlock<true>(block, &next_block_id);
} else {
ProcessBlock<false>(block, &next_block_id);
}
}
}
bool NeedsDecompression(OpIndex op) { return needs_decompression[op]; }
bool NeedsDecompression(const Operation& op) {
return NeedsDecompression(graph.Index(op));
}
bool MarkAsNeedsDecompression(OpIndex op) {
return (needs_decompression[op] = true);
}
template <bool is_loop>
void ProcessBlock(const Block& block, uint32_t* next_block_id) {
for (const Operation& op : base::Reversed(graph.operations(block))) {
if (is_loop && op.Is<PhiOp>() && NeedsDecompression(op)) {
const PhiOp& phi = op.Cast<PhiOp>();
if (!NeedsDecompression(phi.input(1))) {
Block* backedge = block.LastPredecessor();
*next_block_id =
std::max<uint32_t>(*next_block_id, backedge->index().id());
}
}
ProcessOperation(op);
}
}
void ProcessOperation(const Operation& op);
};
void DecompressionAnalyzer::ProcessOperation(const Operation& op) {
switch (op.opcode) {
case Opcode::kStore: {
auto& store = op.Cast<StoreOp>();
MarkAsNeedsDecompression(store.base());
if (store.index().valid()) MarkAsNeedsDecompression(store.index());
if (!store.stored_rep.IsTagged()) {
MarkAsNeedsDecompression(store.value());
}
break;
}
case Opcode::kFrameState:
// The deopt code knows how to handle compressed inputs.
break;
case Opcode::kPhi: {
// Replicate the phi's state for its inputs.
auto& phi = op.Cast<PhiOp>();
if (NeedsDecompression(op)) {
for (OpIndex input : phi.inputs()) {
MarkAsNeedsDecompression(input);
}
} else {
candidates.push_back(graph.Index(op));
}
break;
}
case Opcode::kEqual: {
auto& equal = op.Cast<EqualOp>();
if (equal.rep == WordRepresentation::Word64()) {
MarkAsNeedsDecompression(equal.left());
MarkAsNeedsDecompression(equal.right());
}
break;
}
case Opcode::kComparison: {
auto& comp = op.Cast<ComparisonOp>();
if (comp.rep == WordRepresentation::Word64()) {
MarkAsNeedsDecompression(comp.left());
MarkAsNeedsDecompression(comp.right());
}
break;
}
case Opcode::kWordBinop: {
auto& binary_op = op.Cast<WordBinopOp>();
if (binary_op.rep == WordRepresentation::Word64()) {
MarkAsNeedsDecompression(binary_op.left());
MarkAsNeedsDecompression(binary_op.right());
}
break;
}
case Opcode::kShift: {
auto& shift_op = op.Cast<ShiftOp>();
if (shift_op.rep == WordRepresentation::Word64()) {
MarkAsNeedsDecompression(shift_op.left());
}
break;
}
case Opcode::kChange: {
auto& change = op.Cast<ChangeOp>();
if (change.to == WordRepresentation::Word64() && NeedsDecompression(op)) {
MarkAsNeedsDecompression(change.input());
}
break;
}
case Opcode::kTaggedBitcast: {
auto& bitcast = op.Cast<TaggedBitcastOp>();
if (NeedsDecompression(op)) {
MarkAsNeedsDecompression(bitcast.input());
} else {
candidates.push_back(graph.Index(op));
}
break;
}
case Opcode::kLoad:
case Opcode::kConstant:
if (!NeedsDecompression(op)) {
candidates.push_back(graph.Index(op));
}
V8_FALLTHROUGH;
default:
for (OpIndex input : op.inputs()) {
MarkAsNeedsDecompression(input);
}
break;
}
}
} // namespace
// Instead of using `OptimizationPhase`, we directly mutate the operations after
// the analysis. Doing it in-place is possible because we only modify operation
// options.
void RunDecompressionOptimization(Graph& graph, Zone* phase_zone) {
DecompressionAnalyzer analyzer(graph, phase_zone);
analyzer.Run();
for (OpIndex op_idx : analyzer.candidates) {
Operation& op = graph.Get(op_idx);
if (analyzer.NeedsDecompression(op)) continue;
switch (op.opcode) {
case Opcode::kConstant: {
auto& constant = op.Cast<ConstantOp>();
if (constant.kind == ConstantOp::Kind::kHeapObject) {
constant.kind = ConstantOp::Kind::kCompressedHeapObject;
}
break;
}
case Opcode::kPhi: {
auto& phi = op.Cast<PhiOp>();
if (phi.rep == RegisterRepresentation::Tagged()) {
phi.rep = RegisterRepresentation::Compressed();
}
break;
}
case Opcode::kLoad: {
auto& load = op.Cast<LoadOp>();
if (load.loaded_rep.IsTagged()) {
DCHECK_EQ(load.result_rep,
any_of(RegisterRepresentation::Tagged(),
RegisterRepresentation::Compressed()));
load.result_rep = RegisterRepresentation::Compressed();
}
break;
}
case Opcode::kTaggedBitcast: {
auto& bitcast = op.Cast<TaggedBitcastOp>();
if (bitcast.from == RegisterRepresentation::Tagged() &&
bitcast.to == RegisterRepresentation::PointerSized()) {
bitcast.from = RegisterRepresentation::Compressed();
bitcast.to = RegisterRepresentation::Word32();
}
break;
}
default:
break;
}
}
}
} // namespace v8::internal::compiler::turboshaft
|
