1 //===- InlineSizeEstimatorAnalysis.cpp - IR to native size from ML model --===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This implements feature and label extraction for offline supervised learning
10 // of a IR to native size model.
11 //
12 //===----------------------------------------------------------------------===//
13 #include "llvm/Analysis/InlineSizeEstimatorAnalysis.h"
14 
15 #ifdef LLVM_HAVE_TFLITE
16 #include "llvm/Analysis/Utils/TFUtils.h"
17 #endif
18 #include "llvm/IR/Function.h"
19 #include "llvm/IR/PassManager.h"
20 #include "llvm/Support/raw_ostream.h"
21 
22 using namespace llvm;
23 
24 AnalysisKey InlineSizeEstimatorAnalysis::Key;
25 
26 #ifdef LLVM_HAVE_TFLITE
27 #include "llvm/Analysis/LoopInfo.h"
28 #include "llvm/Analysis/TargetLibraryInfo.h"
29 #include "llvm/Analysis/TargetTransformInfo.h"
30 #include "llvm/IR/BasicBlock.h"
31 #include "llvm/IR/Dominators.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/MC/MCAsmLayout.h"
34 #include "llvm/Support/Casting.h"
35 #include "llvm/Support/CommandLine.h"
36 #include <algorithm>
37 #include <deque>
38 #include <optional>
39 
40 cl::opt<std::string> TFIR2NativeModelPath(
41     "ml-inliner-ir2native-model", cl::Hidden,
42     cl::desc("Path to saved model evaluating native size from IR."));
43 
44 #define DEBUG_TYPE "inline-size-estimator"
45 namespace {
46 unsigned getMaxInstructionID() {
47 #define LAST_OTHER_INST(NR) return NR;
48 #include "llvm/IR/Instruction.def"
49 }
50 
51 class IRToNativeSizeLearning {
52 public:
53   enum class NamedFeatureIndex : size_t {
54     InitialSize,
55     Blocks,
56     Calls,
57     IsLocal,
58     IsLinkOnceODR,
59     IsLinkOnce,
60     Loops,
61     MaxLoopDepth,
62     MaxDomTreeLevel,
63 
64     NumNamedFeatures
65   };
66   static const size_t NumNamedFeatures =
67       static_cast<size_t>(NamedFeatureIndex::NumNamedFeatures);
68   struct FunctionFeatures {
69     static const size_t FeatureCount;
70 
71     std::array<int32_t, NumNamedFeatures> NamedFeatures = {0};
72     std::vector<int32_t> InstructionHistogram;
73     std::vector<int32_t> InstructionPairHistogram;
74 
75     void fillTensor(int32_t *Ptr) const;
76     int32_t &operator[](NamedFeatureIndex Pos) {
77       return NamedFeatures[static_cast<size_t>(Pos)];
78     }
79   };
80   IRToNativeSizeLearning() = default;
81 
82   static FunctionFeatures getFunctionFeatures(Function &F,
83                                               FunctionAnalysisManager &FAM);
84 };
85 
86 // This is a point in time - we determined including these pairs of
87 // consecutive instructions (in the IR layout available at inline time) as
88 // features improves the model performance. We want to move away from manual
89 // feature selection.
90 // The array is given in opcode pairs rather than labels because 1) labels
91 // weren't readily available, and 2) the successions were hand - extracted.
92 //
93 // This array must be sorted.
94 static const std::array<std::pair<size_t, size_t>, 137>
95     ImportantInstructionSuccessions{
96         {{1, 1},   {1, 4},   {1, 5},   {1, 7},   {1, 8},   {1, 9},   {1, 11},
97          {1, 12},  {1, 13},  {1, 14},  {1, 18},  {1, 20},  {1, 22},  {1, 24},
98          {1, 25},  {1, 26},  {1, 27},  {1, 28},  {1, 29},  {1, 30},  {1, 31},
99          {1, 32},  {1, 33},  {1, 34},  {1, 39},  {1, 40},  {1, 42},  {1, 45},
100          {2, 1},   {2, 2},   {2, 13},  {2, 28},  {2, 29},  {2, 32},  {2, 33},
101          {2, 34},  {2, 38},  {2, 48},  {2, 49},  {2, 53},  {2, 55},  {2, 56},
102          {13, 2},  {13, 13}, {13, 26}, {13, 33}, {13, 34}, {13, 56}, {15, 27},
103          {28, 2},  {28, 48}, {28, 53}, {29, 2},  {29, 33}, {29, 56}, {31, 31},
104          {31, 33}, {31, 34}, {31, 49}, {32, 1},  {32, 2},  {32, 13}, {32, 15},
105          {32, 28}, {32, 29}, {32, 32}, {32, 33}, {32, 34}, {32, 39}, {32, 40},
106          {32, 48}, {32, 49}, {32, 53}, {32, 56}, {33, 1},  {33, 2},  {33, 32},
107          {33, 33}, {33, 34}, {33, 49}, {33, 53}, {33, 56}, {34, 1},  {34, 2},
108          {34, 32}, {34, 33}, {34, 34}, {34, 49}, {34, 53}, {34, 56}, {38, 34},
109          {39, 57}, {40, 34}, {47, 15}, {47, 49}, {48, 2},  {48, 34}, {48, 56},
110          {49, 1},  {49, 2},  {49, 28}, {49, 32}, {49, 33}, {49, 34}, {49, 39},
111          {49, 49}, {49, 56}, {53, 1},  {53, 2},  {53, 28}, {53, 34}, {53, 53},
112          {53, 57}, {55, 1},  {55, 28}, {55, 34}, {55, 53}, {55, 55}, {55, 56},
113          {56, 1},  {56, 2},  {56, 7},  {56, 13}, {56, 32}, {56, 33}, {56, 34},
114          {56, 49}, {56, 53}, {56, 56}, {56, 64}, {57, 34}, {57, 56}, {57, 57},
115          {64, 1},  {64, 64}, {65, 1},  {65, 65}}};
116 
117 // We have: 9 calculated features (the features here); 1 feature for each
118 // instruction opcode; and 1 feature for each manually-identified sequence.
119 // For the latter 2, we build a histogram: we count the number of
120 // occurrences of each instruction opcode or succession of instructions,
121 // respectively.
122 // Note that instruction opcodes start from 1. For convenience, we also have an
123 // always 0 feature for the '0' opcode, hence the extra 1.
124 const size_t IRToNativeSizeLearning::FunctionFeatures::FeatureCount =
125     ImportantInstructionSuccessions.size() + getMaxInstructionID() + 1 +
126     IRToNativeSizeLearning::NumNamedFeatures;
127 
128 size_t getSize(Function &F, TargetTransformInfo &TTI) {
129   size_t Ret = 0;
130   for (const auto &BB : F)
131     for (const auto &I : BB)
132       Ret += *(TTI.getInstructionCost(
133           &I, TargetTransformInfo::TargetCostKind::TCK_CodeSize).getValue());
134   return Ret;
135 }
136 
137 size_t getSize(Function &F, FunctionAnalysisManager &FAM) {
138   auto &TTI = FAM.getResult<TargetIRAnalysis>(F);
139   return getSize(F, TTI);
140 }
141 
142 unsigned getMaxDominatorTreeDepth(const Function &F,
143                                   const DominatorTree &Tree) {
144   unsigned Ret = 0;
145   for (const auto &BB : F)
146     if (const auto *TN = Tree.getNode(&BB))
147       Ret = std::max(Ret, TN->getLevel());
148   return Ret;
149 }
150 } // namespace
151 
152 IRToNativeSizeLearning::FunctionFeatures
153 IRToNativeSizeLearning::getFunctionFeatures(Function &F,
154                                             FunctionAnalysisManager &FAM) {
155   assert(llvm::is_sorted(ImportantInstructionSuccessions) &&
156          "expected function features are sorted");
157 
158   auto &DomTree = FAM.getResult<DominatorTreeAnalysis>(F);
159   FunctionFeatures FF;
160   size_t InstrCount = getMaxInstructionID() + 1;
161   FF.InstructionHistogram.resize(InstrCount);
162 
163   FF.InstructionPairHistogram.resize(ImportantInstructionSuccessions.size());
164 
165   int StartID = 0;
166   int LastID = StartID;
167   auto getPairIndex = [](size_t a, size_t b) {
168     auto I = llvm::find(ImportantInstructionSuccessions, std::make_pair(a, b));
169     if (I == ImportantInstructionSuccessions.end())
170       return -1;
171     return static_cast<int>(
172         std::distance(ImportantInstructionSuccessions.begin(), I));
173   };
174 
175   // We don't want debug calls, because they'd just add noise.
176   for (const auto &BB : F) {
177     for (const auto &I : BB.instructionsWithoutDebug()) {
178       auto ID = I.getOpcode();
179 
180       ++FF.InstructionHistogram[ID];
181       int PairIndex = getPairIndex(LastID, ID);
182       if (PairIndex >= 0)
183         ++FF.InstructionPairHistogram[PairIndex];
184       LastID = ID;
185       if (isa<CallBase>(I))
186         ++FF[NamedFeatureIndex::Calls];
187     }
188   }
189 
190   FF[NamedFeatureIndex::InitialSize] = getSize(F, FAM);
191   FF[NamedFeatureIndex::IsLocal] = F.hasLocalLinkage();
192   FF[NamedFeatureIndex::IsLinkOnceODR] = F.hasLinkOnceODRLinkage();
193   FF[NamedFeatureIndex::IsLinkOnce] = F.hasLinkOnceLinkage();
194   FF[NamedFeatureIndex::Blocks] = F.size();
195   auto &LI = FAM.getResult<LoopAnalysis>(F);
196   FF[NamedFeatureIndex::Loops] = std::distance(LI.begin(), LI.end());
197   for (auto &L : LI)
198     FF[NamedFeatureIndex::MaxLoopDepth] =
199         std::max(FF[NamedFeatureIndex::MaxLoopDepth],
200                  static_cast<int32_t>(L->getLoopDepth()));
201   FF[NamedFeatureIndex::MaxDomTreeLevel] = getMaxDominatorTreeDepth(F, DomTree);
202   return FF;
203 }
204 
205 void IRToNativeSizeLearning::FunctionFeatures::fillTensor(int32_t *Ptr) const {
206   std::copy(NamedFeatures.begin(), NamedFeatures.end(), Ptr);
207   Ptr += NamedFeatures.size();
208   std::copy(InstructionHistogram.begin(), InstructionHistogram.end(), Ptr);
209   Ptr += InstructionHistogram.size();
210   std::copy(InstructionPairHistogram.begin(), InstructionPairHistogram.end(),
211             Ptr);
212 }
213 
214 bool InlineSizeEstimatorAnalysis::isEvaluatorRequested() {
215   return !TFIR2NativeModelPath.empty();
216 }
217 
218 InlineSizeEstimatorAnalysis::InlineSizeEstimatorAnalysis() {
219   if (!isEvaluatorRequested()) {
220     return;
221   }
222   std::vector<TensorSpec> InputSpecs{TensorSpec::createSpec<int32_t>(
223       "serving_default_input_1",
224       {1, static_cast<int64_t>(
225               IRToNativeSizeLearning::FunctionFeatures::FeatureCount)})};
226   std::vector<TensorSpec> OutputSpecs{
227       TensorSpec::createSpec<float>("StatefulPartitionedCall", {1})};
228   Evaluator = std::make_unique<TFModelEvaluator>(
229       TFIR2NativeModelPath.getValue().c_str(), InputSpecs, OutputSpecs);
230   if (!Evaluator || !Evaluator->isValid()) {
231     Evaluator.reset();
232     return;
233   }
234 }
235 
236 InlineSizeEstimatorAnalysis::Result
237 InlineSizeEstimatorAnalysis::run(const Function &F,
238                                  FunctionAnalysisManager &FAM) {
239   if (!Evaluator)
240     return std::nullopt;
241   auto Features = IRToNativeSizeLearning::getFunctionFeatures(
242       const_cast<Function &>(F), FAM);
243   int32_t *V = Evaluator->getInput<int32_t>(0);
244   Features.fillTensor(V);
245   auto ER = Evaluator->evaluate();
246   if (!ER)
247     return std::nullopt;
248   float Ret = *ER->getTensorValue<float>(0);
249   if (Ret < 0.0)
250     Ret = 0.0;
251   return static_cast<size_t>(Ret);
252 }
253 
254 InlineSizeEstimatorAnalysis::~InlineSizeEstimatorAnalysis() {}
255 InlineSizeEstimatorAnalysis::InlineSizeEstimatorAnalysis(
256     InlineSizeEstimatorAnalysis &&Other)
257     : Evaluator(std::move(Other.Evaluator)) {}
258 
259 #else
260 namespace llvm {
261 class TFModelEvaluator {};
262 } // namespace llvm
263 InlineSizeEstimatorAnalysis::InlineSizeEstimatorAnalysis() = default;
264 InlineSizeEstimatorAnalysis ::InlineSizeEstimatorAnalysis(
265     InlineSizeEstimatorAnalysis &&) {}
266 InlineSizeEstimatorAnalysis::~InlineSizeEstimatorAnalysis() = default;
267 InlineSizeEstimatorAnalysis::Result
268 InlineSizeEstimatorAnalysis::run(const Function &F,
269                                  FunctionAnalysisManager &FAM) {
270   return std::nullopt;
271 }
272 bool InlineSizeEstimatorAnalysis::isEvaluatorRequested() { return false; }
273 #endif
274 
275 PreservedAnalyses
276 InlineSizeEstimatorAnalysisPrinterPass::run(Function &F,
277                                             FunctionAnalysisManager &AM) {
278   OS << "[InlineSizeEstimatorAnalysis] size estimate for " << F.getName()
279      << ": " << AM.getResult<InlineSizeEstimatorAnalysis>(F) << "\n";
280   return PreservedAnalyses::all();
281 }
282