1 //===- InlineSizeEstimatorAnalysis.cpp - IR to native size from ML model --===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This implements feature and label extraction for offline supervised learning
11 // of a IR to native size model.
12 //
13 //===----------------------------------------------------------------------===//
14 #include "llvm/Analysis/InlineSizeEstimatorAnalysis.h"
15 
16 #ifdef LLVM_HAVE_TF_API
17 #include "llvm/Analysis/Utils/TFUtils.h"
18 #endif
19 #include "llvm/Analysis/LoopInfo.h"
20 #include "llvm/Analysis/TargetLibraryInfo.h"
21 #include "llvm/Analysis/TargetTransformInfo.h"
22 #include "llvm/IR/BasicBlock.h"
23 #include "llvm/IR/Dominators.h"
24 #include "llvm/IR/Function.h"
25 #include "llvm/IR/Instructions.h"
26 #include "llvm/IR/PassManager.h"
27 #include "llvm/MC/MCAsmLayout.h"
28 #include "llvm/Support/Casting.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/raw_ostream.h"
31 
32 #include <algorithm>
33 #include <deque>
34 
35 using namespace llvm;
36 
37 AnalysisKey InlineSizeEstimatorAnalysis::Key;
38 
39 #define DEBUG_TYPE "inline-size-estimator"
40 
41 #ifdef LLVM_HAVE_TF_API
42 cl::opt<std::string> TFIR2NativeModelPath(
43     "ml-inliner-ir2native-model", cl::Hidden,
44     cl::desc("Path to saved model evaluating native size from IR."));
45 
46 namespace {
47 unsigned getMaxInstructionID() {
48 #define LAST_OTHER_INST(NR) return NR;
49 #include "llvm/IR/Instruction.def"
50 }
51 
52 class IRToNativeSizeLearning {
53 public:
54   enum class NamedFeatureIndex : size_t {
55     InitialSize,
56     Blocks,
57     Calls,
58     IsLocal,
59     IsLinkOnceODR,
60     IsLinkOnce,
61     Loops,
62     MaxLoopDepth,
63     MaxDomTreeLevel,
64 
65     NumNamedFeatures
66   };
67   static const size_t NumNamedFeatures =
68       static_cast<size_t>(NamedFeatureIndex::NumNamedFeatures);
69   struct FunctionFeatures {
70     static const size_t FeatureCount;
71 
72     std::array<int32_t, NumNamedFeatures> NamedFeatures = {0};
73     std::vector<int32_t> InstructionHistogram;
74     std::vector<int32_t> InstructionPairHistogram;
75 
76     void fillTensor(int32_t *Ptr) const;
77     int32_t &operator[](NamedFeatureIndex Pos) {
78       return NamedFeatures[static_cast<size_t>(Pos)];
79     }
80   };
81   IRToNativeSizeLearning() = default;
82 
83   static FunctionFeatures getFunctionFeatures(Function &F,
84                                               FunctionAnalysisManager &FAM);
85 };
86 
87 // This is a point in time - we determined including these pairs of
88 // consecutive instructions (in the IR layout available at inline time) as
89 // features improves the model performance. We want to move away from manual
90 // feature selection.
91 // The array is given in opcode pairs rather than labels because 1) labels
92 // weren't readily available, and 2) the successions were hand - extracted.
93 //
94 // This array must be sorted.
95 static const std::array<std::pair<size_t, size_t>, 137>
96     ImportantInstructionSuccessions{
97         {{1, 1},   {1, 4},   {1, 5},   {1, 7},   {1, 8},   {1, 9},   {1, 11},
98          {1, 12},  {1, 13},  {1, 14},  {1, 18},  {1, 20},  {1, 22},  {1, 24},
99          {1, 25},  {1, 26},  {1, 27},  {1, 28},  {1, 29},  {1, 30},  {1, 31},
100          {1, 32},  {1, 33},  {1, 34},  {1, 39},  {1, 40},  {1, 42},  {1, 45},
101          {2, 1},   {2, 2},   {2, 13},  {2, 28},  {2, 29},  {2, 32},  {2, 33},
102          {2, 34},  {2, 38},  {2, 48},  {2, 49},  {2, 53},  {2, 55},  {2, 56},
103          {13, 2},  {13, 13}, {13, 26}, {13, 33}, {13, 34}, {13, 56}, {15, 27},
104          {28, 2},  {28, 48}, {28, 53}, {29, 2},  {29, 33}, {29, 56}, {31, 31},
105          {31, 33}, {31, 34}, {31, 49}, {32, 1},  {32, 2},  {32, 13}, {32, 15},
106          {32, 28}, {32, 29}, {32, 32}, {32, 33}, {32, 34}, {32, 39}, {32, 40},
107          {32, 48}, {32, 49}, {32, 53}, {32, 56}, {33, 1},  {33, 2},  {33, 32},
108          {33, 33}, {33, 34}, {33, 49}, {33, 53}, {33, 56}, {34, 1},  {34, 2},
109          {34, 32}, {34, 33}, {34, 34}, {34, 49}, {34, 53}, {34, 56}, {38, 34},
110          {39, 57}, {40, 34}, {47, 15}, {47, 49}, {48, 2},  {48, 34}, {48, 56},
111          {49, 1},  {49, 2},  {49, 28}, {49, 32}, {49, 33}, {49, 34}, {49, 39},
112          {49, 49}, {49, 56}, {53, 1},  {53, 2},  {53, 28}, {53, 34}, {53, 53},
113          {53, 57}, {55, 1},  {55, 28}, {55, 34}, {55, 53}, {55, 55}, {55, 56},
114          {56, 1},  {56, 2},  {56, 7},  {56, 13}, {56, 32}, {56, 33}, {56, 34},
115          {56, 49}, {56, 53}, {56, 56}, {56, 64}, {57, 34}, {57, 56}, {57, 57},
116          {64, 1},  {64, 64}, {65, 1},  {65, 65}}};
117 
118 // We have: 9 calculated features (the features here); 1 feature for each
119 // instruction opcode; and 1 feature for each manually-identified sequence.
120 // For the latter 2, we build a histogram: we count the number of
121 // occurrences of each instruction opcode or succession of instructions,
122 // respectively.
123 // Note that instruction opcodes start from 1. For convenience, we also have an
124 // always 0 feature for the '0' opcode, hence the extra 1.
125 const size_t IRToNativeSizeLearning::FunctionFeatures::FeatureCount =
126     ImportantInstructionSuccessions.size() + getMaxInstructionID() + 1 +
127     IRToNativeSizeLearning::NumNamedFeatures;
128 
129 size_t getSize(Function &F, TargetTransformInfo &TTI) {
130   size_t Ret = 0;
131   for (const auto &BB : F)
132     for (const auto &I : BB)
133       Ret += *(TTI.getInstructionCost(
134           &I, TargetTransformInfo::TargetCostKind::TCK_CodeSize).getValue());
135   return Ret;
136 }
137 
138 size_t getSize(Function &F, FunctionAnalysisManager &FAM) {
139   auto &TTI = FAM.getResult<TargetIRAnalysis>(F);
140   return getSize(F, TTI);
141 }
142 
143 unsigned getMaxDominatorTreeDepth(const Function &F,
144                                   const DominatorTree &Tree) {
145   unsigned Ret = 0;
146   for (const auto &BB : F)
147     if (const auto *TN = Tree.getNode(&BB))
148       Ret = std::max(Ret, TN->getLevel());
149   return Ret;
150 }
151 } // namespace
152 
153 IRToNativeSizeLearning::FunctionFeatures
154 IRToNativeSizeLearning::getFunctionFeatures(Function &F,
155                                             FunctionAnalysisManager &FAM) {
156   assert(llvm::is_sorted(ImportantInstructionSuccessions) &&
157          "expected function features are sorted");
158 
159   auto &DomTree = FAM.getResult<DominatorTreeAnalysis>(F);
160   FunctionFeatures FF;
161   size_t InstrCount = getMaxInstructionID() + 1;
162   FF.InstructionHistogram.resize(InstrCount);
163 
164   FF.InstructionPairHistogram.resize(ImportantInstructionSuccessions.size());
165 
166   int StartID = 0;
167   int LastID = StartID;
168   auto getPairIndex = [](size_t a, size_t b) {
169     auto I = llvm::find(ImportantInstructionSuccessions, std::make_pair(a, b));
170     if (I == ImportantInstructionSuccessions.end())
171       return -1;
172     return static_cast<int>(
173         std::distance(ImportantInstructionSuccessions.begin(), I));
174   };
175 
176   // We don't want debug calls, because they'd just add noise.
177   for (const auto &BB : F) {
178     for (const auto &I : BB.instructionsWithoutDebug()) {
179       auto ID = I.getOpcode();
180 
181       ++FF.InstructionHistogram[ID];
182       int PairIndex = getPairIndex(LastID, ID);
183       if (PairIndex >= 0)
184         ++FF.InstructionPairHistogram[PairIndex];
185       LastID = ID;
186       if (isa<CallBase>(I))
187         ++FF[NamedFeatureIndex::Calls];
188     }
189   }
190 
191   FF[NamedFeatureIndex::InitialSize] = getSize(F, FAM);
192   FF[NamedFeatureIndex::IsLocal] = F.hasLocalLinkage();
193   FF[NamedFeatureIndex::IsLinkOnceODR] = F.hasLinkOnceODRLinkage();
194   FF[NamedFeatureIndex::IsLinkOnce] = F.hasLinkOnceLinkage();
195   FF[NamedFeatureIndex::Blocks] =
196       std::distance(F.getBasicBlockList().begin(), F.getBasicBlockList().end());
197   auto &LI = FAM.getResult<LoopAnalysis>(F);
198   FF[NamedFeatureIndex::Loops] = std::distance(LI.begin(), LI.end());
199   for (auto &L : LI)
200     FF[NamedFeatureIndex::MaxLoopDepth] =
201         std::max(FF[NamedFeatureIndex::MaxLoopDepth],
202                  static_cast<int32_t>(L->getLoopDepth()));
203   FF[NamedFeatureIndex::MaxDomTreeLevel] = getMaxDominatorTreeDepth(F, DomTree);
204   return FF;
205 }
206 
207 void IRToNativeSizeLearning::FunctionFeatures::fillTensor(int32_t *Ptr) const {
208   std::copy(NamedFeatures.begin(), NamedFeatures.end(), Ptr);
209   Ptr += NamedFeatures.size();
210   std::copy(InstructionHistogram.begin(), InstructionHistogram.end(), Ptr);
211   Ptr += InstructionHistogram.size();
212   std::copy(InstructionPairHistogram.begin(), InstructionPairHistogram.end(),
213             Ptr);
214 }
215 
216 bool InlineSizeEstimatorAnalysis::isEvaluatorRequested() {
217   return !TFIR2NativeModelPath.empty();
218 }
219 
220 InlineSizeEstimatorAnalysis::InlineSizeEstimatorAnalysis() {
221   if (!isEvaluatorRequested()) {
222     return;
223   }
224   std::vector<TensorSpec> InputSpecs{TensorSpec::createSpec<int32_t>(
225       "serving_default_input_1",
226       {1, static_cast<int64_t>(
227               IRToNativeSizeLearning::FunctionFeatures::FeatureCount)})};
228   std::vector<TensorSpec> OutputSpecs{
229       TensorSpec::createSpec<float>("StatefulPartitionedCall", {1})};
230   Evaluator = std::make_unique<TFModelEvaluator>(
231       TFIR2NativeModelPath.getValue().c_str(), InputSpecs, OutputSpecs);
232   if (!Evaluator || !Evaluator->isValid()) {
233     Evaluator.reset();
234     return;
235   }
236 }
237 
238 InlineSizeEstimatorAnalysis::Result
239 InlineSizeEstimatorAnalysis::run(const Function &F,
240                                  FunctionAnalysisManager &FAM) {
241   if (!Evaluator)
242     return None;
243   auto Features = IRToNativeSizeLearning::getFunctionFeatures(
244       const_cast<Function &>(F), FAM);
245   int32_t *V = Evaluator->getInput<int32_t>(0);
246   Features.fillTensor(V);
247   auto ER = Evaluator->evaluate();
248   if (!ER)
249     return None;
250   float Ret = *ER->getTensorValue<float>(0);
251   if (Ret < 0.0)
252     Ret = 0.0;
253   return static_cast<size_t>(Ret);
254 }
255 
256 InlineSizeEstimatorAnalysis::~InlineSizeEstimatorAnalysis() {}
257 InlineSizeEstimatorAnalysis::InlineSizeEstimatorAnalysis(
258     InlineSizeEstimatorAnalysis &&Other)
259     : Evaluator(std::move(Other.Evaluator)) {}
260 
261 #else
262 namespace llvm {
263 class TFModelEvaluator {};
264 } // namespace llvm
265 InlineSizeEstimatorAnalysis::InlineSizeEstimatorAnalysis() {}
266 InlineSizeEstimatorAnalysis ::InlineSizeEstimatorAnalysis(
267     InlineSizeEstimatorAnalysis &&) {}
268 InlineSizeEstimatorAnalysis::~InlineSizeEstimatorAnalysis() {}
269 InlineSizeEstimatorAnalysis::Result
270 InlineSizeEstimatorAnalysis::run(const Function &F,
271                                  FunctionAnalysisManager &FAM) {
272   return None;
273 }
274 bool InlineSizeEstimatorAnalysis::isEvaluatorRequested() { return false; }
275 #endif
276 
277 PreservedAnalyses
278 InlineSizeEstimatorAnalysisPrinterPass::run(Function &F,
279                                             FunctionAnalysisManager &AM) {
280   OS << "[InlineSizeEstimatorAnalysis] size estimate for " << F.getName()
281      << ": " << AM.getResult<InlineSizeEstimatorAnalysis>(F) << "\n";
282   return PreservedAnalyses::all();
283 }
284