| File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp |
| Warning: | line 883, column 48 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===- PGOInstrumentation.cpp - MST-based PGO Instrumentation -------------===// | |||
| 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 file implements PGO instrumentation using a minimum spanning tree based | |||
| 10 | // on the following paper: | |||
| 11 | // [1] Donald E. Knuth, Francis R. Stevenson. Optimal measurement of points | |||
| 12 | // for program frequency counts. BIT Numerical Mathematics 1973, Volume 13, | |||
| 13 | // Issue 3, pp 313-322 | |||
| 14 | // The idea of the algorithm based on the fact that for each node (except for | |||
| 15 | // the entry and exit), the sum of incoming edge counts equals the sum of | |||
| 16 | // outgoing edge counts. The count of edge on spanning tree can be derived from | |||
| 17 | // those edges not on the spanning tree. Knuth proves this method instruments | |||
| 18 | // the minimum number of edges. | |||
| 19 | // | |||
| 20 | // The minimal spanning tree here is actually a maximum weight tree -- on-tree | |||
| 21 | // edges have higher frequencies (more likely to execute). The idea is to | |||
| 22 | // instrument those less frequently executed edges to reduce the runtime | |||
| 23 | // overhead of instrumented binaries. | |||
| 24 | // | |||
| 25 | // This file contains two passes: | |||
| 26 | // (1) Pass PGOInstrumentationGen which instruments the IR to generate edge | |||
| 27 | // count profile, and generates the instrumentation for indirect call | |||
| 28 | // profiling. | |||
| 29 | // (2) Pass PGOInstrumentationUse which reads the edge count profile and | |||
| 30 | // annotates the branch weights. It also reads the indirect call value | |||
| 31 | // profiling records and annotate the indirect call instructions. | |||
| 32 | // | |||
| 33 | // To get the precise counter information, These two passes need to invoke at | |||
| 34 | // the same compilation point (so they see the same IR). For pass | |||
| 35 | // PGOInstrumentationGen, the real work is done in instrumentOneFunc(). For | |||
| 36 | // pass PGOInstrumentationUse, the real work in done in class PGOUseFunc and | |||
| 37 | // the profile is opened in module level and passed to each PGOUseFunc instance. | |||
| 38 | // The shared code for PGOInstrumentationGen and PGOInstrumentationUse is put | |||
| 39 | // in class FuncPGOInstrumentation. | |||
| 40 | // | |||
| 41 | // Class PGOEdge represents a CFG edge and some auxiliary information. Class | |||
| 42 | // BBInfo contains auxiliary information for each BB. These two classes are used | |||
| 43 | // in pass PGOInstrumentationGen. Class PGOUseEdge and UseBBInfo are the derived | |||
| 44 | // class of PGOEdge and BBInfo, respectively. They contains extra data structure | |||
| 45 | // used in populating profile counters. | |||
| 46 | // The MST implementation is in Class CFGMST (CFGMST.h). | |||
| 47 | // | |||
| 48 | //===----------------------------------------------------------------------===// | |||
| 49 | ||||
| 50 | #include "llvm/Transforms/Instrumentation/PGOInstrumentation.h" | |||
| 51 | #include "CFGMST.h" | |||
| 52 | #include "ValueProfileCollector.h" | |||
| 53 | #include "llvm/ADT/APInt.h" | |||
| 54 | #include "llvm/ADT/ArrayRef.h" | |||
| 55 | #include "llvm/ADT/MapVector.h" | |||
| 56 | #include "llvm/ADT/STLExtras.h" | |||
| 57 | #include "llvm/ADT/SmallVector.h" | |||
| 58 | #include "llvm/ADT/Statistic.h" | |||
| 59 | #include "llvm/ADT/StringRef.h" | |||
| 60 | #include "llvm/ADT/Triple.h" | |||
| 61 | #include "llvm/ADT/Twine.h" | |||
| 62 | #include "llvm/ADT/iterator.h" | |||
| 63 | #include "llvm/ADT/iterator_range.h" | |||
| 64 | #include "llvm/Analysis/BlockFrequencyInfo.h" | |||
| 65 | #include "llvm/Analysis/BranchProbabilityInfo.h" | |||
| 66 | #include "llvm/Analysis/CFG.h" | |||
| 67 | #include "llvm/Analysis/EHPersonalities.h" | |||
| 68 | #include "llvm/Analysis/LoopInfo.h" | |||
| 69 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | |||
| 70 | #include "llvm/Analysis/ProfileSummaryInfo.h" | |||
| 71 | #include "llvm/IR/Attributes.h" | |||
| 72 | #include "llvm/IR/BasicBlock.h" | |||
| 73 | #include "llvm/IR/CFG.h" | |||
| 74 | #include "llvm/IR/Comdat.h" | |||
| 75 | #include "llvm/IR/Constant.h" | |||
| 76 | #include "llvm/IR/Constants.h" | |||
| 77 | #include "llvm/IR/DiagnosticInfo.h" | |||
| 78 | #include "llvm/IR/Dominators.h" | |||
| 79 | #include "llvm/IR/Function.h" | |||
| 80 | #include "llvm/IR/GlobalAlias.h" | |||
| 81 | #include "llvm/IR/GlobalValue.h" | |||
| 82 | #include "llvm/IR/GlobalVariable.h" | |||
| 83 | #include "llvm/IR/IRBuilder.h" | |||
| 84 | #include "llvm/IR/InstVisitor.h" | |||
| 85 | #include "llvm/IR/InstrTypes.h" | |||
| 86 | #include "llvm/IR/Instruction.h" | |||
| 87 | #include "llvm/IR/Instructions.h" | |||
| 88 | #include "llvm/IR/IntrinsicInst.h" | |||
| 89 | #include "llvm/IR/Intrinsics.h" | |||
| 90 | #include "llvm/IR/LLVMContext.h" | |||
| 91 | #include "llvm/IR/MDBuilder.h" | |||
| 92 | #include "llvm/IR/Module.h" | |||
| 93 | #include "llvm/IR/PassManager.h" | |||
| 94 | #include "llvm/IR/ProfileSummary.h" | |||
| 95 | #include "llvm/IR/Type.h" | |||
| 96 | #include "llvm/IR/Value.h" | |||
| 97 | #include "llvm/InitializePasses.h" | |||
| 98 | #include "llvm/Pass.h" | |||
| 99 | #include "llvm/ProfileData/InstrProf.h" | |||
| 100 | #include "llvm/ProfileData/InstrProfReader.h" | |||
| 101 | #include "llvm/Support/BranchProbability.h" | |||
| 102 | #include "llvm/Support/CRC.h" | |||
| 103 | #include "llvm/Support/Casting.h" | |||
| 104 | #include "llvm/Support/CommandLine.h" | |||
| 105 | #include "llvm/Support/DOTGraphTraits.h" | |||
| 106 | #include "llvm/Support/Debug.h" | |||
| 107 | #include "llvm/Support/Error.h" | |||
| 108 | #include "llvm/Support/ErrorHandling.h" | |||
| 109 | #include "llvm/Support/GraphWriter.h" | |||
| 110 | #include "llvm/Support/raw_ostream.h" | |||
| 111 | #include "llvm/Transforms/Instrumentation.h" | |||
| 112 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | |||
| 113 | #include <algorithm> | |||
| 114 | #include <cassert> | |||
| 115 | #include <cstdint> | |||
| 116 | #include <memory> | |||
| 117 | #include <numeric> | |||
| 118 | #include <string> | |||
| 119 | #include <unordered_map> | |||
| 120 | #include <utility> | |||
| 121 | #include <vector> | |||
| 122 | ||||
| 123 | using namespace llvm; | |||
| 124 | using ProfileCount = Function::ProfileCount; | |||
| 125 | using VPCandidateInfo = ValueProfileCollector::CandidateInfo; | |||
| 126 | ||||
| 127 | #define DEBUG_TYPE"pgo-instrumentation" "pgo-instrumentation" | |||
| 128 | ||||
| 129 | STATISTIC(NumOfPGOInstrument, "Number of edges instrumented.")static llvm::Statistic NumOfPGOInstrument = {"pgo-instrumentation" , "NumOfPGOInstrument", "Number of edges instrumented."}; | |||
| 130 | STATISTIC(NumOfPGOSelectInsts, "Number of select instruction instrumented.")static llvm::Statistic NumOfPGOSelectInsts = {"pgo-instrumentation" , "NumOfPGOSelectInsts", "Number of select instruction instrumented." }; | |||
| 131 | STATISTIC(NumOfPGOMemIntrinsics, "Number of mem intrinsics instrumented.")static llvm::Statistic NumOfPGOMemIntrinsics = {"pgo-instrumentation" , "NumOfPGOMemIntrinsics", "Number of mem intrinsics instrumented." }; | |||
| 132 | STATISTIC(NumOfPGOEdge, "Number of edges.")static llvm::Statistic NumOfPGOEdge = {"pgo-instrumentation", "NumOfPGOEdge", "Number of edges."}; | |||
| 133 | STATISTIC(NumOfPGOBB, "Number of basic-blocks.")static llvm::Statistic NumOfPGOBB = {"pgo-instrumentation", "NumOfPGOBB" , "Number of basic-blocks."}; | |||
| 134 | STATISTIC(NumOfPGOSplit, "Number of critical edge splits.")static llvm::Statistic NumOfPGOSplit = {"pgo-instrumentation" , "NumOfPGOSplit", "Number of critical edge splits."}; | |||
| 135 | STATISTIC(NumOfPGOFunc, "Number of functions having valid profile counts.")static llvm::Statistic NumOfPGOFunc = {"pgo-instrumentation", "NumOfPGOFunc", "Number of functions having valid profile counts." }; | |||
| 136 | STATISTIC(NumOfPGOMismatch, "Number of functions having mismatch profile.")static llvm::Statistic NumOfPGOMismatch = {"pgo-instrumentation" , "NumOfPGOMismatch", "Number of functions having mismatch profile." }; | |||
| 137 | STATISTIC(NumOfPGOMissing, "Number of functions without profile.")static llvm::Statistic NumOfPGOMissing = {"pgo-instrumentation" , "NumOfPGOMissing", "Number of functions without profile."}; | |||
| 138 | STATISTIC(NumOfPGOICall, "Number of indirect call value instrumentations.")static llvm::Statistic NumOfPGOICall = {"pgo-instrumentation" , "NumOfPGOICall", "Number of indirect call value instrumentations." }; | |||
| 139 | STATISTIC(NumOfCSPGOInstrument, "Number of edges instrumented in CSPGO.")static llvm::Statistic NumOfCSPGOInstrument = {"pgo-instrumentation" , "NumOfCSPGOInstrument", "Number of edges instrumented in CSPGO." }; | |||
| 140 | STATISTIC(NumOfCSPGOSelectInsts,static llvm::Statistic NumOfCSPGOSelectInsts = {"pgo-instrumentation" , "NumOfCSPGOSelectInsts", "Number of select instruction instrumented in CSPGO." } | |||
| 141 | "Number of select instruction instrumented in CSPGO.")static llvm::Statistic NumOfCSPGOSelectInsts = {"pgo-instrumentation" , "NumOfCSPGOSelectInsts", "Number of select instruction instrumented in CSPGO." }; | |||
| 142 | STATISTIC(NumOfCSPGOMemIntrinsics,static llvm::Statistic NumOfCSPGOMemIntrinsics = {"pgo-instrumentation" , "NumOfCSPGOMemIntrinsics", "Number of mem intrinsics instrumented in CSPGO." } | |||
| 143 | "Number of mem intrinsics instrumented in CSPGO.")static llvm::Statistic NumOfCSPGOMemIntrinsics = {"pgo-instrumentation" , "NumOfCSPGOMemIntrinsics", "Number of mem intrinsics instrumented in CSPGO." }; | |||
| 144 | STATISTIC(NumOfCSPGOEdge, "Number of edges in CSPGO.")static llvm::Statistic NumOfCSPGOEdge = {"pgo-instrumentation" , "NumOfCSPGOEdge", "Number of edges in CSPGO."}; | |||
| 145 | STATISTIC(NumOfCSPGOBB, "Number of basic-blocks in CSPGO.")static llvm::Statistic NumOfCSPGOBB = {"pgo-instrumentation", "NumOfCSPGOBB", "Number of basic-blocks in CSPGO."}; | |||
| 146 | STATISTIC(NumOfCSPGOSplit, "Number of critical edge splits in CSPGO.")static llvm::Statistic NumOfCSPGOSplit = {"pgo-instrumentation" , "NumOfCSPGOSplit", "Number of critical edge splits in CSPGO." }; | |||
| 147 | STATISTIC(NumOfCSPGOFunc,static llvm::Statistic NumOfCSPGOFunc = {"pgo-instrumentation" , "NumOfCSPGOFunc", "Number of functions having valid profile counts in CSPGO." } | |||
| 148 | "Number of functions having valid profile counts in CSPGO.")static llvm::Statistic NumOfCSPGOFunc = {"pgo-instrumentation" , "NumOfCSPGOFunc", "Number of functions having valid profile counts in CSPGO." }; | |||
| 149 | STATISTIC(NumOfCSPGOMismatch,static llvm::Statistic NumOfCSPGOMismatch = {"pgo-instrumentation" , "NumOfCSPGOMismatch", "Number of functions having mismatch profile in CSPGO." } | |||
| 150 | "Number of functions having mismatch profile in CSPGO.")static llvm::Statistic NumOfCSPGOMismatch = {"pgo-instrumentation" , "NumOfCSPGOMismatch", "Number of functions having mismatch profile in CSPGO." }; | |||
| 151 | STATISTIC(NumOfCSPGOMissing, "Number of functions without profile in CSPGO.")static llvm::Statistic NumOfCSPGOMissing = {"pgo-instrumentation" , "NumOfCSPGOMissing", "Number of functions without profile in CSPGO." }; | |||
| 152 | ||||
| 153 | // Command line option to specify the file to read profile from. This is | |||
| 154 | // mainly used for testing. | |||
| 155 | static cl::opt<std::string> | |||
| 156 | PGOTestProfileFile("pgo-test-profile-file", cl::init(""), cl::Hidden, | |||
| 157 | cl::value_desc("filename"), | |||
| 158 | cl::desc("Specify the path of profile data file. This is" | |||
| 159 | "mainly for test purpose.")); | |||
| 160 | static cl::opt<std::string> PGOTestProfileRemappingFile( | |||
| 161 | "pgo-test-profile-remapping-file", cl::init(""), cl::Hidden, | |||
| 162 | cl::value_desc("filename"), | |||
| 163 | cl::desc("Specify the path of profile remapping file. This is mainly for " | |||
| 164 | "test purpose.")); | |||
| 165 | ||||
| 166 | // Command line option to disable value profiling. The default is false: | |||
| 167 | // i.e. value profiling is enabled by default. This is for debug purpose. | |||
| 168 | static cl::opt<bool> DisableValueProfiling("disable-vp", cl::init(false), | |||
| 169 | cl::Hidden, | |||
| 170 | cl::desc("Disable Value Profiling")); | |||
| 171 | ||||
| 172 | // Command line option to set the maximum number of VP annotations to write to | |||
| 173 | // the metadata for a single indirect call callsite. | |||
| 174 | static cl::opt<unsigned> MaxNumAnnotations( | |||
| 175 | "icp-max-annotations", cl::init(3), cl::Hidden, cl::ZeroOrMore, | |||
| 176 | cl::desc("Max number of annotations for a single indirect " | |||
| 177 | "call callsite")); | |||
| 178 | ||||
| 179 | // Command line option to set the maximum number of value annotations | |||
| 180 | // to write to the metadata for a single memop intrinsic. | |||
| 181 | static cl::opt<unsigned> MaxNumMemOPAnnotations( | |||
| 182 | "memop-max-annotations", cl::init(4), cl::Hidden, cl::ZeroOrMore, | |||
| 183 | cl::desc("Max number of preicise value annotations for a single memop" | |||
| 184 | "intrinsic")); | |||
| 185 | ||||
| 186 | // Command line option to control appending FunctionHash to the name of a COMDAT | |||
| 187 | // function. This is to avoid the hash mismatch caused by the preinliner. | |||
| 188 | static cl::opt<bool> DoComdatRenaming( | |||
| 189 | "do-comdat-renaming", cl::init(false), cl::Hidden, | |||
| 190 | cl::desc("Append function hash to the name of COMDAT function to avoid " | |||
| 191 | "function hash mismatch due to the preinliner")); | |||
| 192 | ||||
| 193 | // Command line option to enable/disable the warning about missing profile | |||
| 194 | // information. | |||
| 195 | static cl::opt<bool> | |||
| 196 | PGOWarnMissing("pgo-warn-missing-function", cl::init(false), cl::Hidden, | |||
| 197 | cl::desc("Use this option to turn on/off " | |||
| 198 | "warnings about missing profile data for " | |||
| 199 | "functions.")); | |||
| 200 | ||||
| 201 | // Command line option to enable/disable the warning about a hash mismatch in | |||
| 202 | // the profile data. | |||
| 203 | static cl::opt<bool> | |||
| 204 | NoPGOWarnMismatch("no-pgo-warn-mismatch", cl::init(false), cl::Hidden, | |||
| 205 | cl::desc("Use this option to turn off/on " | |||
| 206 | "warnings about profile cfg mismatch.")); | |||
| 207 | ||||
| 208 | // Command line option to enable/disable the warning about a hash mismatch in | |||
| 209 | // the profile data for Comdat functions, which often turns out to be false | |||
| 210 | // positive due to the pre-instrumentation inline. | |||
| 211 | static cl::opt<bool> | |||
| 212 | NoPGOWarnMismatchComdat("no-pgo-warn-mismatch-comdat", cl::init(true), | |||
| 213 | cl::Hidden, | |||
| 214 | cl::desc("The option is used to turn on/off " | |||
| 215 | "warnings about hash mismatch for comdat " | |||
| 216 | "functions.")); | |||
| 217 | ||||
| 218 | // Command line option to enable/disable select instruction instrumentation. | |||
| 219 | static cl::opt<bool> | |||
| 220 | PGOInstrSelect("pgo-instr-select", cl::init(true), cl::Hidden, | |||
| 221 | cl::desc("Use this option to turn on/off SELECT " | |||
| 222 | "instruction instrumentation. ")); | |||
| 223 | ||||
| 224 | // Command line option to turn on CFG dot or text dump of raw profile counts | |||
| 225 | static cl::opt<PGOViewCountsType> PGOViewRawCounts( | |||
| 226 | "pgo-view-raw-counts", cl::Hidden, | |||
| 227 | cl::desc("A boolean option to show CFG dag or text " | |||
| 228 | "with raw profile counts from " | |||
| 229 | "profile data. See also option " | |||
| 230 | "-pgo-view-counts. To limit graph " | |||
| 231 | "display to only one function, use " | |||
| 232 | "filtering option -view-bfi-func-name."), | |||
| 233 | cl::values(clEnumValN(PGOVCT_None, "none", "do not show.")llvm::cl::OptionEnumValue { "none", int(PGOVCT_None), "do not show." }, | |||
| 234 | clEnumValN(PGOVCT_Graph, "graph", "show a graph.")llvm::cl::OptionEnumValue { "graph", int(PGOVCT_Graph), "show a graph." }, | |||
| 235 | clEnumValN(PGOVCT_Text, "text", "show in text.")llvm::cl::OptionEnumValue { "text", int(PGOVCT_Text), "show in text." })); | |||
| 236 | ||||
| 237 | // Command line option to enable/disable memop intrinsic call.size profiling. | |||
| 238 | static cl::opt<bool> | |||
| 239 | PGOInstrMemOP("pgo-instr-memop", cl::init(true), cl::Hidden, | |||
| 240 | cl::desc("Use this option to turn on/off " | |||
| 241 | "memory intrinsic size profiling.")); | |||
| 242 | ||||
| 243 | // Emit branch probability as optimization remarks. | |||
| 244 | static cl::opt<bool> | |||
| 245 | EmitBranchProbability("pgo-emit-branch-prob", cl::init(false), cl::Hidden, | |||
| 246 | cl::desc("When this option is on, the annotated " | |||
| 247 | "branch probability will be emitted as " | |||
| 248 | "optimization remarks: -{Rpass|" | |||
| 249 | "pass-remarks}=pgo-instrumentation")); | |||
| 250 | ||||
| 251 | static cl::opt<bool> PGOInstrumentEntry( | |||
| 252 | "pgo-instrument-entry", cl::init(false), cl::Hidden, | |||
| 253 | cl::desc("Force to instrument function entry basicblock.")); | |||
| 254 | ||||
| 255 | static cl::opt<bool> | |||
| 256 | PGOFixEntryCount("pgo-fix-entry-count", cl::init(true), cl::Hidden, | |||
| 257 | cl::desc("Fix function entry count in profile use.")); | |||
| 258 | ||||
| 259 | static cl::opt<bool> PGOVerifyHotBFI( | |||
| 260 | "pgo-verify-hot-bfi", cl::init(false), cl::Hidden, | |||
| 261 | cl::desc("Print out the non-match BFI count if a hot raw profile count " | |||
| 262 | "becomes non-hot, or a cold raw profile count becomes hot. " | |||
| 263 | "The print is enabled under -Rpass-analysis=pgo, or " | |||
| 264 | "internal option -pass-remakrs-analysis=pgo.")); | |||
| 265 | ||||
| 266 | static cl::opt<bool> PGOVerifyBFI( | |||
| 267 | "pgo-verify-bfi", cl::init(false), cl::Hidden, | |||
| 268 | cl::desc("Print out mismatched BFI counts after setting profile metadata " | |||
| 269 | "The print is enabled under -Rpass-analysis=pgo, or " | |||
| 270 | "internal option -pass-remakrs-analysis=pgo.")); | |||
| 271 | ||||
| 272 | static cl::opt<unsigned> PGOVerifyBFIRatio( | |||
| 273 | "pgo-verify-bfi-ratio", cl::init(5), cl::Hidden, | |||
| 274 | cl::desc("Set the threshold for pgo-verify-big -- only print out " | |||
| 275 | "mismatched BFI if the difference percentage is greater than " | |||
| 276 | "this value (in percentage).")); | |||
| 277 | ||||
| 278 | static cl::opt<unsigned> PGOVerifyBFICutoff( | |||
| 279 | "pgo-verify-bfi-cutoff", cl::init(1), cl::Hidden, | |||
| 280 | cl::desc("Set the threshold for pgo-verify-bfi -- skip the counts whose " | |||
| 281 | "profile count value is below.")); | |||
| 282 | ||||
| 283 | namespace llvm { | |||
| 284 | // Command line option to turn on CFG dot dump after profile annotation. | |||
| 285 | // Defined in Analysis/BlockFrequencyInfo.cpp: -pgo-view-counts | |||
| 286 | extern cl::opt<PGOViewCountsType> PGOViewCounts; | |||
| 287 | ||||
| 288 | // Command line option to specify the name of the function for CFG dump | |||
| 289 | // Defined in Analysis/BlockFrequencyInfo.cpp: -view-bfi-func-name= | |||
| 290 | extern cl::opt<std::string> ViewBlockFreqFuncName; | |||
| 291 | } // namespace llvm | |||
| 292 | ||||
| 293 | static cl::opt<bool> | |||
| 294 | PGOOldCFGHashing("pgo-instr-old-cfg-hashing", cl::init(false), cl::Hidden, | |||
| 295 | cl::desc("Use the old CFG function hashing")); | |||
| 296 | ||||
| 297 | // Return a string describing the branch condition that can be | |||
| 298 | // used in static branch probability heuristics: | |||
| 299 | static std::string getBranchCondString(Instruction *TI) { | |||
| 300 | BranchInst *BI = dyn_cast<BranchInst>(TI); | |||
| 301 | if (!BI || !BI->isConditional()) | |||
| 302 | return std::string(); | |||
| 303 | ||||
| 304 | Value *Cond = BI->getCondition(); | |||
| 305 | ICmpInst *CI = dyn_cast<ICmpInst>(Cond); | |||
| 306 | if (!CI) | |||
| 307 | return std::string(); | |||
| 308 | ||||
| 309 | std::string result; | |||
| 310 | raw_string_ostream OS(result); | |||
| 311 | OS << CmpInst::getPredicateName(CI->getPredicate()) << "_"; | |||
| 312 | CI->getOperand(0)->getType()->print(OS, true); | |||
| 313 | ||||
| 314 | Value *RHS = CI->getOperand(1); | |||
| 315 | ConstantInt *CV = dyn_cast<ConstantInt>(RHS); | |||
| 316 | if (CV) { | |||
| 317 | if (CV->isZero()) | |||
| 318 | OS << "_Zero"; | |||
| 319 | else if (CV->isOne()) | |||
| 320 | OS << "_One"; | |||
| 321 | else if (CV->isMinusOne()) | |||
| 322 | OS << "_MinusOne"; | |||
| 323 | else | |||
| 324 | OS << "_Const"; | |||
| 325 | } | |||
| 326 | OS.flush(); | |||
| 327 | return result; | |||
| 328 | } | |||
| 329 | ||||
| 330 | static const char *ValueProfKindDescr[] = { | |||
| 331 | #define VALUE_PROF_KIND(Enumerator, Value, Descr) Descr, | |||
| 332 | #include "llvm/ProfileData/InstrProfData.inc" | |||
| 333 | }; | |||
| 334 | ||||
| 335 | namespace { | |||
| 336 | ||||
| 337 | /// The select instruction visitor plays three roles specified | |||
| 338 | /// by the mode. In \c VM_counting mode, it simply counts the number of | |||
| 339 | /// select instructions. In \c VM_instrument mode, it inserts code to count | |||
| 340 | /// the number times TrueValue of select is taken. In \c VM_annotate mode, | |||
| 341 | /// it reads the profile data and annotate the select instruction with metadata. | |||
| 342 | enum VisitMode { VM_counting, VM_instrument, VM_annotate }; | |||
| 343 | class PGOUseFunc; | |||
| 344 | ||||
| 345 | /// Instruction Visitor class to visit select instructions. | |||
| 346 | struct SelectInstVisitor : public InstVisitor<SelectInstVisitor> { | |||
| 347 | Function &F; | |||
| 348 | unsigned NSIs = 0; // Number of select instructions instrumented. | |||
| 349 | VisitMode Mode = VM_counting; // Visiting mode. | |||
| 350 | unsigned *CurCtrIdx = nullptr; // Pointer to current counter index. | |||
| 351 | unsigned TotalNumCtrs = 0; // Total number of counters | |||
| 352 | GlobalVariable *FuncNameVar = nullptr; | |||
| 353 | uint64_t FuncHash = 0; | |||
| 354 | PGOUseFunc *UseFunc = nullptr; | |||
| 355 | ||||
| 356 | SelectInstVisitor(Function &Func) : F(Func) {} | |||
| 357 | ||||
| 358 | void countSelects(Function &Func) { | |||
| 359 | NSIs = 0; | |||
| 360 | Mode = VM_counting; | |||
| 361 | visit(Func); | |||
| 362 | } | |||
| 363 | ||||
| 364 | // Visit the IR stream and instrument all select instructions. \p | |||
| 365 | // Ind is a pointer to the counter index variable; \p TotalNC | |||
| 366 | // is the total number of counters; \p FNV is the pointer to the | |||
| 367 | // PGO function name var; \p FHash is the function hash. | |||
| 368 | void instrumentSelects(Function &Func, unsigned *Ind, unsigned TotalNC, | |||
| 369 | GlobalVariable *FNV, uint64_t FHash) { | |||
| 370 | Mode = VM_instrument; | |||
| 371 | CurCtrIdx = Ind; | |||
| 372 | TotalNumCtrs = TotalNC; | |||
| 373 | FuncHash = FHash; | |||
| 374 | FuncNameVar = FNV; | |||
| 375 | visit(Func); | |||
| 376 | } | |||
| 377 | ||||
| 378 | // Visit the IR stream and annotate all select instructions. | |||
| 379 | void annotateSelects(Function &Func, PGOUseFunc *UF, unsigned *Ind) { | |||
| 380 | Mode = VM_annotate; | |||
| 381 | UseFunc = UF; | |||
| 382 | CurCtrIdx = Ind; | |||
| 383 | visit(Func); | |||
| 384 | } | |||
| 385 | ||||
| 386 | void instrumentOneSelectInst(SelectInst &SI); | |||
| 387 | void annotateOneSelectInst(SelectInst &SI); | |||
| 388 | ||||
| 389 | // Visit \p SI instruction and perform tasks according to visit mode. | |||
| 390 | void visitSelectInst(SelectInst &SI); | |||
| 391 | ||||
| 392 | // Return the number of select instructions. This needs be called after | |||
| 393 | // countSelects(). | |||
| 394 | unsigned getNumOfSelectInsts() const { return NSIs; } | |||
| 395 | }; | |||
| 396 | ||||
| 397 | ||||
| 398 | class PGOInstrumentationGenLegacyPass : public ModulePass { | |||
| 399 | public: | |||
| 400 | static char ID; | |||
| 401 | ||||
| 402 | PGOInstrumentationGenLegacyPass(bool IsCS = false) | |||
| 403 | : ModulePass(ID), IsCS(IsCS) { | |||
| 404 | initializePGOInstrumentationGenLegacyPassPass( | |||
| 405 | *PassRegistry::getPassRegistry()); | |||
| 406 | } | |||
| 407 | ||||
| 408 | StringRef getPassName() const override { return "PGOInstrumentationGenPass"; } | |||
| 409 | ||||
| 410 | private: | |||
| 411 | // Is this is context-sensitive instrumentation. | |||
| 412 | bool IsCS; | |||
| 413 | bool runOnModule(Module &M) override; | |||
| 414 | ||||
| 415 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
| 416 | AU.addRequired<BlockFrequencyInfoWrapperPass>(); | |||
| 417 | AU.addRequired<TargetLibraryInfoWrapperPass>(); | |||
| 418 | } | |||
| 419 | }; | |||
| 420 | ||||
| 421 | class PGOInstrumentationUseLegacyPass : public ModulePass { | |||
| 422 | public: | |||
| 423 | static char ID; | |||
| 424 | ||||
| 425 | // Provide the profile filename as the parameter. | |||
| 426 | PGOInstrumentationUseLegacyPass(std::string Filename = "", bool IsCS = false) | |||
| 427 | : ModulePass(ID), ProfileFileName(std::move(Filename)), IsCS(IsCS) { | |||
| 428 | if (!PGOTestProfileFile.empty()) | |||
| 429 | ProfileFileName = PGOTestProfileFile; | |||
| 430 | initializePGOInstrumentationUseLegacyPassPass( | |||
| 431 | *PassRegistry::getPassRegistry()); | |||
| 432 | } | |||
| 433 | ||||
| 434 | StringRef getPassName() const override { return "PGOInstrumentationUsePass"; } | |||
| 435 | ||||
| 436 | private: | |||
| 437 | std::string ProfileFileName; | |||
| 438 | // Is this is context-sensitive instrumentation use. | |||
| 439 | bool IsCS; | |||
| 440 | ||||
| 441 | bool runOnModule(Module &M) override; | |||
| 442 | ||||
| 443 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
| 444 | AU.addRequired<ProfileSummaryInfoWrapperPass>(); | |||
| 445 | AU.addRequired<BlockFrequencyInfoWrapperPass>(); | |||
| 446 | AU.addRequired<TargetLibraryInfoWrapperPass>(); | |||
| 447 | } | |||
| 448 | }; | |||
| 449 | ||||
| 450 | class PGOInstrumentationGenCreateVarLegacyPass : public ModulePass { | |||
| 451 | public: | |||
| 452 | static char ID; | |||
| 453 | StringRef getPassName() const override { | |||
| 454 | return "PGOInstrumentationGenCreateVarPass"; | |||
| 455 | } | |||
| 456 | PGOInstrumentationGenCreateVarLegacyPass(std::string CSInstrName = "") | |||
| 457 | : ModulePass(ID), InstrProfileOutput(CSInstrName) { | |||
| 458 | initializePGOInstrumentationGenCreateVarLegacyPassPass( | |||
| 459 | *PassRegistry::getPassRegistry()); | |||
| 460 | } | |||
| 461 | ||||
| 462 | private: | |||
| 463 | bool runOnModule(Module &M) override { | |||
| 464 | createProfileFileNameVar(M, InstrProfileOutput); | |||
| 465 | createIRLevelProfileFlagVar(M, /* IsCS */ true, PGOInstrumentEntry); | |||
| 466 | return false; | |||
| 467 | } | |||
| 468 | std::string InstrProfileOutput; | |||
| 469 | }; | |||
| 470 | ||||
| 471 | } // end anonymous namespace | |||
| 472 | ||||
| 473 | char PGOInstrumentationGenLegacyPass::ID = 0; | |||
| 474 | ||||
| 475 | INITIALIZE_PASS_BEGIN(PGOInstrumentationGenLegacyPass, "pgo-instr-gen",static void *initializePGOInstrumentationGenLegacyPassPassOnce (PassRegistry &Registry) { | |||
| 476 | "PGO instrumentation.", false, false)static void *initializePGOInstrumentationGenLegacyPassPassOnce (PassRegistry &Registry) { | |||
| 477 | INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)initializeBlockFrequencyInfoWrapperPassPass(Registry); | |||
| 478 | INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)initializeBranchProbabilityInfoWrapperPassPass(Registry); | |||
| 479 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); | |||
| 480 | INITIALIZE_PASS_END(PGOInstrumentationGenLegacyPass, "pgo-instr-gen",PassInfo *PI = new PassInfo( "PGO instrumentation.", "pgo-instr-gen" , &PGOInstrumentationGenLegacyPass::ID, PassInfo::NormalCtor_t (callDefaultCtor<PGOInstrumentationGenLegacyPass>), false , false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializePGOInstrumentationGenLegacyPassPassFlag ; void llvm::initializePGOInstrumentationGenLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializePGOInstrumentationGenLegacyPassPassFlag , initializePGOInstrumentationGenLegacyPassPassOnce, std::ref (Registry)); } | |||
| 481 | "PGO instrumentation.", false, false)PassInfo *PI = new PassInfo( "PGO instrumentation.", "pgo-instr-gen" , &PGOInstrumentationGenLegacyPass::ID, PassInfo::NormalCtor_t (callDefaultCtor<PGOInstrumentationGenLegacyPass>), false , false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializePGOInstrumentationGenLegacyPassPassFlag ; void llvm::initializePGOInstrumentationGenLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializePGOInstrumentationGenLegacyPassPassFlag , initializePGOInstrumentationGenLegacyPassPassOnce, std::ref (Registry)); } | |||
| 482 | ||||
| 483 | ModulePass *llvm::createPGOInstrumentationGenLegacyPass(bool IsCS) { | |||
| 484 | return new PGOInstrumentationGenLegacyPass(IsCS); | |||
| 485 | } | |||
| 486 | ||||
| 487 | char PGOInstrumentationUseLegacyPass::ID = 0; | |||
| 488 | ||||
| 489 | INITIALIZE_PASS_BEGIN(PGOInstrumentationUseLegacyPass, "pgo-instr-use",static void *initializePGOInstrumentationUseLegacyPassPassOnce (PassRegistry &Registry) { | |||
| 490 | "Read PGO instrumentation profile.", false, false)static void *initializePGOInstrumentationUseLegacyPassPassOnce (PassRegistry &Registry) { | |||
| 491 | INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)initializeBlockFrequencyInfoWrapperPassPass(Registry); | |||
| 492 | INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)initializeBranchProbabilityInfoWrapperPassPass(Registry); | |||
| 493 | INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)initializeProfileSummaryInfoWrapperPassPass(Registry); | |||
| 494 | INITIALIZE_PASS_END(PGOInstrumentationUseLegacyPass, "pgo-instr-use",PassInfo *PI = new PassInfo( "Read PGO instrumentation profile." , "pgo-instr-use", &PGOInstrumentationUseLegacyPass::ID, PassInfo ::NormalCtor_t(callDefaultCtor<PGOInstrumentationUseLegacyPass >), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializePGOInstrumentationUseLegacyPassPassFlag ; void llvm::initializePGOInstrumentationUseLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializePGOInstrumentationUseLegacyPassPassFlag , initializePGOInstrumentationUseLegacyPassPassOnce, std::ref (Registry)); } | |||
| 495 | "Read PGO instrumentation profile.", false, false)PassInfo *PI = new PassInfo( "Read PGO instrumentation profile." , "pgo-instr-use", &PGOInstrumentationUseLegacyPass::ID, PassInfo ::NormalCtor_t(callDefaultCtor<PGOInstrumentationUseLegacyPass >), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializePGOInstrumentationUseLegacyPassPassFlag ; void llvm::initializePGOInstrumentationUseLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializePGOInstrumentationUseLegacyPassPassFlag , initializePGOInstrumentationUseLegacyPassPassOnce, std::ref (Registry)); } | |||
| 496 | ||||
| 497 | ModulePass *llvm::createPGOInstrumentationUseLegacyPass(StringRef Filename, | |||
| 498 | bool IsCS) { | |||
| 499 | return new PGOInstrumentationUseLegacyPass(Filename.str(), IsCS); | |||
| 500 | } | |||
| 501 | ||||
| 502 | char PGOInstrumentationGenCreateVarLegacyPass::ID = 0; | |||
| 503 | ||||
| 504 | INITIALIZE_PASS(PGOInstrumentationGenCreateVarLegacyPass,static void *initializePGOInstrumentationGenCreateVarLegacyPassPassOnce (PassRegistry &Registry) { PassInfo *PI = new PassInfo( "Create PGO instrumentation version variable for CSPGO." , "pgo-instr-gen-create-var", &PGOInstrumentationGenCreateVarLegacyPass ::ID, PassInfo::NormalCtor_t(callDefaultCtor<PGOInstrumentationGenCreateVarLegacyPass >), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializePGOInstrumentationGenCreateVarLegacyPassPassFlag ; void llvm::initializePGOInstrumentationGenCreateVarLegacyPassPass (PassRegistry &Registry) { llvm::call_once(InitializePGOInstrumentationGenCreateVarLegacyPassPassFlag , initializePGOInstrumentationGenCreateVarLegacyPassPassOnce, std::ref(Registry)); } | |||
| 505 | "pgo-instr-gen-create-var",static void *initializePGOInstrumentationGenCreateVarLegacyPassPassOnce (PassRegistry &Registry) { PassInfo *PI = new PassInfo( "Create PGO instrumentation version variable for CSPGO." , "pgo-instr-gen-create-var", &PGOInstrumentationGenCreateVarLegacyPass ::ID, PassInfo::NormalCtor_t(callDefaultCtor<PGOInstrumentationGenCreateVarLegacyPass >), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializePGOInstrumentationGenCreateVarLegacyPassPassFlag ; void llvm::initializePGOInstrumentationGenCreateVarLegacyPassPass (PassRegistry &Registry) { llvm::call_once(InitializePGOInstrumentationGenCreateVarLegacyPassPassFlag , initializePGOInstrumentationGenCreateVarLegacyPassPassOnce, std::ref(Registry)); } | |||
| 506 | "Create PGO instrumentation version variable for CSPGO.", false,static void *initializePGOInstrumentationGenCreateVarLegacyPassPassOnce (PassRegistry &Registry) { PassInfo *PI = new PassInfo( "Create PGO instrumentation version variable for CSPGO." , "pgo-instr-gen-create-var", &PGOInstrumentationGenCreateVarLegacyPass ::ID, PassInfo::NormalCtor_t(callDefaultCtor<PGOInstrumentationGenCreateVarLegacyPass >), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializePGOInstrumentationGenCreateVarLegacyPassPassFlag ; void llvm::initializePGOInstrumentationGenCreateVarLegacyPassPass (PassRegistry &Registry) { llvm::call_once(InitializePGOInstrumentationGenCreateVarLegacyPassPassFlag , initializePGOInstrumentationGenCreateVarLegacyPassPassOnce, std::ref(Registry)); } | |||
| 507 | false)static void *initializePGOInstrumentationGenCreateVarLegacyPassPassOnce (PassRegistry &Registry) { PassInfo *PI = new PassInfo( "Create PGO instrumentation version variable for CSPGO." , "pgo-instr-gen-create-var", &PGOInstrumentationGenCreateVarLegacyPass ::ID, PassInfo::NormalCtor_t(callDefaultCtor<PGOInstrumentationGenCreateVarLegacyPass >), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializePGOInstrumentationGenCreateVarLegacyPassPassFlag ; void llvm::initializePGOInstrumentationGenCreateVarLegacyPassPass (PassRegistry &Registry) { llvm::call_once(InitializePGOInstrumentationGenCreateVarLegacyPassPassFlag , initializePGOInstrumentationGenCreateVarLegacyPassPassOnce, std::ref(Registry)); } | |||
| 508 | ||||
| 509 | ModulePass * | |||
| 510 | llvm::createPGOInstrumentationGenCreateVarLegacyPass(StringRef CSInstrName) { | |||
| 511 | return new PGOInstrumentationGenCreateVarLegacyPass(std::string(CSInstrName)); | |||
| 512 | } | |||
| 513 | ||||
| 514 | namespace { | |||
| 515 | ||||
| 516 | /// An MST based instrumentation for PGO | |||
| 517 | /// | |||
| 518 | /// Implements a Minimum Spanning Tree (MST) based instrumentation for PGO | |||
| 519 | /// in the function level. | |||
| 520 | struct PGOEdge { | |||
| 521 | // This class implements the CFG edges. Note the CFG can be a multi-graph. | |||
| 522 | // So there might be multiple edges with same SrcBB and DestBB. | |||
| 523 | const BasicBlock *SrcBB; | |||
| 524 | const BasicBlock *DestBB; | |||
| 525 | uint64_t Weight; | |||
| 526 | bool InMST = false; | |||
| 527 | bool Removed = false; | |||
| 528 | bool IsCritical = false; | |||
| 529 | ||||
| 530 | PGOEdge(const BasicBlock *Src, const BasicBlock *Dest, uint64_t W = 1) | |||
| 531 | : SrcBB(Src), DestBB(Dest), Weight(W) {} | |||
| 532 | ||||
| 533 | // Return the information string of an edge. | |||
| 534 | std::string infoString() const { | |||
| 535 | return (Twine(Removed ? "-" : " ") + (InMST ? " " : "*") + | |||
| 536 | (IsCritical ? "c" : " ") + " W=" + Twine(Weight)).str(); | |||
| 537 | } | |||
| 538 | }; | |||
| 539 | ||||
| 540 | // This class stores the auxiliary information for each BB. | |||
| 541 | struct BBInfo { | |||
| 542 | BBInfo *Group; | |||
| 543 | uint32_t Index; | |||
| 544 | uint32_t Rank = 0; | |||
| 545 | ||||
| 546 | BBInfo(unsigned IX) : Group(this), Index(IX) {} | |||
| 547 | ||||
| 548 | // Return the information string of this object. | |||
| 549 | std::string infoString() const { | |||
| 550 | return (Twine("Index=") + Twine(Index)).str(); | |||
| 551 | } | |||
| 552 | ||||
| 553 | // Empty function -- only applicable to UseBBInfo. | |||
| 554 | void addOutEdge(PGOEdge *E LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__))) {} | |||
| 555 | ||||
| 556 | // Empty function -- only applicable to UseBBInfo. | |||
| 557 | void addInEdge(PGOEdge *E LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__))) {} | |||
| 558 | }; | |||
| 559 | ||||
| 560 | // This class implements the CFG edges. Note the CFG can be a multi-graph. | |||
| 561 | template <class Edge, class BBInfo> class FuncPGOInstrumentation { | |||
| 562 | private: | |||
| 563 | Function &F; | |||
| 564 | ||||
| 565 | // Is this is context-sensitive instrumentation. | |||
| 566 | bool IsCS; | |||
| 567 | ||||
| 568 | // A map that stores the Comdat group in function F. | |||
| 569 | std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers; | |||
| 570 | ||||
| 571 | ValueProfileCollector VPC; | |||
| 572 | ||||
| 573 | void computeCFGHash(); | |||
| 574 | void renameComdatFunction(); | |||
| 575 | ||||
| 576 | public: | |||
| 577 | std::vector<std::vector<VPCandidateInfo>> ValueSites; | |||
| 578 | SelectInstVisitor SIVisitor; | |||
| 579 | std::string FuncName; | |||
| 580 | GlobalVariable *FuncNameVar; | |||
| 581 | ||||
| 582 | // CFG hash value for this function. | |||
| 583 | uint64_t FunctionHash = 0; | |||
| 584 | ||||
| 585 | // The Minimum Spanning Tree of function CFG. | |||
| 586 | CFGMST<Edge, BBInfo> MST; | |||
| 587 | ||||
| 588 | // Collect all the BBs that will be instrumented, and store them in | |||
| 589 | // InstrumentBBs. | |||
| 590 | void getInstrumentBBs(std::vector<BasicBlock *> &InstrumentBBs); | |||
| 591 | ||||
| 592 | // Give an edge, find the BB that will be instrumented. | |||
| 593 | // Return nullptr if there is no BB to be instrumented. | |||
| 594 | BasicBlock *getInstrBB(Edge *E); | |||
| 595 | ||||
| 596 | // Return the auxiliary BB information. | |||
| 597 | BBInfo &getBBInfo(const BasicBlock *BB) const { return MST.getBBInfo(BB); } | |||
| 598 | ||||
| 599 | // Return the auxiliary BB information if available. | |||
| 600 | BBInfo *findBBInfo(const BasicBlock *BB) const { return MST.findBBInfo(BB); } | |||
| 601 | ||||
| 602 | // Dump edges and BB information. | |||
| 603 | void dumpInfo(std::string Str = "") const { | |||
| 604 | MST.dumpEdges(dbgs(), Twine("Dump Function ") + FuncName + " Hash: " + | |||
| 605 | Twine(FunctionHash) + "\t" + Str); | |||
| 606 | } | |||
| 607 | ||||
| 608 | FuncPGOInstrumentation( | |||
| 609 | Function &Func, TargetLibraryInfo &TLI, | |||
| 610 | std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers, | |||
| 611 | bool CreateGlobalVar = false, BranchProbabilityInfo *BPI = nullptr, | |||
| 612 | BlockFrequencyInfo *BFI = nullptr, bool IsCS = false, | |||
| 613 | bool InstrumentFuncEntry = true) | |||
| 614 | : F(Func), IsCS(IsCS), ComdatMembers(ComdatMembers), VPC(Func, TLI), | |||
| 615 | ValueSites(IPVK_Last + 1), SIVisitor(Func), | |||
| 616 | MST(F, InstrumentFuncEntry, BPI, BFI) { | |||
| 617 | // This should be done before CFG hash computation. | |||
| 618 | SIVisitor.countSelects(Func); | |||
| 619 | ValueSites[IPVK_MemOPSize] = VPC.get(IPVK_MemOPSize); | |||
| 620 | if (!IsCS) { | |||
| 621 | NumOfPGOSelectInsts += SIVisitor.getNumOfSelectInsts(); | |||
| 622 | NumOfPGOMemIntrinsics += ValueSites[IPVK_MemOPSize].size(); | |||
| 623 | NumOfPGOBB += MST.BBInfos.size(); | |||
| 624 | ValueSites[IPVK_IndirectCallTarget] = VPC.get(IPVK_IndirectCallTarget); | |||
| 625 | } else { | |||
| 626 | NumOfCSPGOSelectInsts += SIVisitor.getNumOfSelectInsts(); | |||
| 627 | NumOfCSPGOMemIntrinsics += ValueSites[IPVK_MemOPSize].size(); | |||
| 628 | NumOfCSPGOBB += MST.BBInfos.size(); | |||
| 629 | } | |||
| 630 | ||||
| 631 | FuncName = getPGOFuncName(F); | |||
| 632 | computeCFGHash(); | |||
| 633 | if (!ComdatMembers.empty()) | |||
| 634 | renameComdatFunction(); | |||
| 635 | LLVM_DEBUG(dumpInfo("after CFGMST"))do { } while (false); | |||
| 636 | ||||
| 637 | for (auto &E : MST.AllEdges) { | |||
| 638 | if (E->Removed) | |||
| 639 | continue; | |||
| 640 | IsCS ? NumOfCSPGOEdge++ : NumOfPGOEdge++; | |||
| 641 | if (!E->InMST) | |||
| 642 | IsCS ? NumOfCSPGOInstrument++ : NumOfPGOInstrument++; | |||
| 643 | } | |||
| 644 | ||||
| 645 | if (CreateGlobalVar) | |||
| 646 | FuncNameVar = createPGOFuncNameVar(F, FuncName); | |||
| 647 | } | |||
| 648 | }; | |||
| 649 | ||||
| 650 | } // end anonymous namespace | |||
| 651 | ||||
| 652 | // Compute Hash value for the CFG: the lower 32 bits are CRC32 of the index | |||
| 653 | // value of each BB in the CFG. The higher 32 bits are the CRC32 of the numbers | |||
| 654 | // of selects, indirect calls, mem ops and edges. | |||
| 655 | template <class Edge, class BBInfo> | |||
| 656 | void FuncPGOInstrumentation<Edge, BBInfo>::computeCFGHash() { | |||
| 657 | std::vector<uint8_t> Indexes; | |||
| 658 | JamCRC JC; | |||
| 659 | for (auto &BB : F) { | |||
| 660 | const Instruction *TI = BB.getTerminator(); | |||
| 661 | for (unsigned I = 0, E = TI->getNumSuccessors(); I != E; ++I) { | |||
| 662 | BasicBlock *Succ = TI->getSuccessor(I); | |||
| 663 | auto BI = findBBInfo(Succ); | |||
| 664 | if (BI == nullptr) | |||
| 665 | continue; | |||
| 666 | uint32_t Index = BI->Index; | |||
| 667 | for (int J = 0; J < 4; J++) | |||
| 668 | Indexes.push_back((uint8_t)(Index >> (J * 8))); | |||
| 669 | } | |||
| 670 | } | |||
| 671 | JC.update(Indexes); | |||
| 672 | ||||
| 673 | JamCRC JCH; | |||
| 674 | if (PGOOldCFGHashing) { | |||
| 675 | // Hash format for context sensitive profile. Reserve 4 bits for other | |||
| 676 | // information. | |||
| 677 | FunctionHash = (uint64_t)SIVisitor.getNumOfSelectInsts() << 56 | | |||
| 678 | (uint64_t)ValueSites[IPVK_IndirectCallTarget].size() << 48 | | |||
| 679 | //(uint64_t)ValueSites[IPVK_MemOPSize].size() << 40 | | |||
| 680 | (uint64_t)MST.AllEdges.size() << 32 | JC.getCRC(); | |||
| 681 | } else { | |||
| 682 | // The higher 32 bits. | |||
| 683 | auto updateJCH = [&JCH](uint64_t Num) { | |||
| 684 | uint8_t Data[8]; | |||
| 685 | support::endian::write64le(Data, Num); | |||
| 686 | JCH.update(Data); | |||
| 687 | }; | |||
| 688 | updateJCH((uint64_t)SIVisitor.getNumOfSelectInsts()); | |||
| 689 | updateJCH((uint64_t)ValueSites[IPVK_IndirectCallTarget].size()); | |||
| 690 | updateJCH((uint64_t)ValueSites[IPVK_MemOPSize].size()); | |||
| 691 | updateJCH((uint64_t)MST.AllEdges.size()); | |||
| 692 | ||||
| 693 | // Hash format for context sensitive profile. Reserve 4 bits for other | |||
| 694 | // information. | |||
| 695 | FunctionHash = (((uint64_t)JCH.getCRC()) << 28) + JC.getCRC(); | |||
| 696 | } | |||
| 697 | ||||
| 698 | // Reserve bit 60-63 for other information purpose. | |||
| 699 | FunctionHash &= 0x0FFFFFFFFFFFFFFF; | |||
| 700 | if (IsCS) | |||
| 701 | NamedInstrProfRecord::setCSFlagInHash(FunctionHash); | |||
| 702 | LLVM_DEBUG(dbgs() << "Function Hash Computation for " << F.getName() << ":\n"do { } while (false) | |||
| 703 | << " CRC = " << JC.getCRC()do { } while (false) | |||
| 704 | << ", Selects = " << SIVisitor.getNumOfSelectInsts()do { } while (false) | |||
| 705 | << ", Edges = " << MST.AllEdges.size() << ", ICSites = "do { } while (false) | |||
| 706 | << ValueSites[IPVK_IndirectCallTarget].size())do { } while (false); | |||
| 707 | if (!PGOOldCFGHashing) { | |||
| 708 | LLVM_DEBUG(dbgs() << ", Memops = " << ValueSites[IPVK_MemOPSize].size()do { } while (false) | |||
| 709 | << ", High32 CRC = " << JCH.getCRC())do { } while (false); | |||
| 710 | } | |||
| 711 | LLVM_DEBUG(dbgs() << ", Hash = " << FunctionHash << "\n";)do { } while (false); | |||
| 712 | } | |||
| 713 | ||||
| 714 | // Check if we can safely rename this Comdat function. | |||
| 715 | static bool canRenameComdat( | |||
| 716 | Function &F, | |||
| 717 | std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers) { | |||
| 718 | if (!DoComdatRenaming || !canRenameComdatFunc(F, true)) | |||
| 719 | return false; | |||
| 720 | ||||
| 721 | // FIXME: Current only handle those Comdat groups that only containing one | |||
| 722 | // function. | |||
| 723 | // (1) For a Comdat group containing multiple functions, we need to have a | |||
| 724 | // unique postfix based on the hashes for each function. There is a | |||
| 725 | // non-trivial code refactoring to do this efficiently. | |||
| 726 | // (2) Variables can not be renamed, so we can not rename Comdat function in a | |||
| 727 | // group including global vars. | |||
| 728 | Comdat *C = F.getComdat(); | |||
| 729 | for (auto &&CM : make_range(ComdatMembers.equal_range(C))) { | |||
| 730 | assert(!isa<GlobalAlias>(CM.second))((void)0); | |||
| 731 | Function *FM = dyn_cast<Function>(CM.second); | |||
| 732 | if (FM != &F) | |||
| 733 | return false; | |||
| 734 | } | |||
| 735 | return true; | |||
| 736 | } | |||
| 737 | ||||
| 738 | // Append the CFGHash to the Comdat function name. | |||
| 739 | template <class Edge, class BBInfo> | |||
| 740 | void FuncPGOInstrumentation<Edge, BBInfo>::renameComdatFunction() { | |||
| 741 | if (!canRenameComdat(F, ComdatMembers)) | |||
| 742 | return; | |||
| 743 | std::string OrigName = F.getName().str(); | |||
| 744 | std::string NewFuncName = | |||
| 745 | Twine(F.getName() + "." + Twine(FunctionHash)).str(); | |||
| 746 | F.setName(Twine(NewFuncName)); | |||
| 747 | GlobalAlias::create(GlobalValue::WeakAnyLinkage, OrigName, &F); | |||
| 748 | FuncName = Twine(FuncName + "." + Twine(FunctionHash)).str(); | |||
| 749 | Comdat *NewComdat; | |||
| 750 | Module *M = F.getParent(); | |||
| 751 | // For AvailableExternallyLinkage functions, change the linkage to | |||
| 752 | // LinkOnceODR and put them into comdat. This is because after renaming, there | |||
| 753 | // is no backup external copy available for the function. | |||
| 754 | if (!F.hasComdat()) { | |||
| 755 | assert(F.getLinkage() == GlobalValue::AvailableExternallyLinkage)((void)0); | |||
| 756 | NewComdat = M->getOrInsertComdat(StringRef(NewFuncName)); | |||
| 757 | F.setLinkage(GlobalValue::LinkOnceODRLinkage); | |||
| 758 | F.setComdat(NewComdat); | |||
| 759 | return; | |||
| 760 | } | |||
| 761 | ||||
| 762 | // This function belongs to a single function Comdat group. | |||
| 763 | Comdat *OrigComdat = F.getComdat(); | |||
| 764 | std::string NewComdatName = | |||
| 765 | Twine(OrigComdat->getName() + "." + Twine(FunctionHash)).str(); | |||
| 766 | NewComdat = M->getOrInsertComdat(StringRef(NewComdatName)); | |||
| 767 | NewComdat->setSelectionKind(OrigComdat->getSelectionKind()); | |||
| 768 | ||||
| 769 | for (auto &&CM : make_range(ComdatMembers.equal_range(OrigComdat))) { | |||
| 770 | // Must be a function. | |||
| 771 | cast<Function>(CM.second)->setComdat(NewComdat); | |||
| 772 | } | |||
| 773 | } | |||
| 774 | ||||
| 775 | // Collect all the BBs that will be instruments and return them in | |||
| 776 | // InstrumentBBs and setup InEdges/OutEdge for UseBBInfo. | |||
| 777 | template <class Edge, class BBInfo> | |||
| 778 | void FuncPGOInstrumentation<Edge, BBInfo>::getInstrumentBBs( | |||
| 779 | std::vector<BasicBlock *> &InstrumentBBs) { | |||
| 780 | // Use a worklist as we will update the vector during the iteration. | |||
| 781 | std::vector<Edge *> EdgeList; | |||
| 782 | EdgeList.reserve(MST.AllEdges.size()); | |||
| 783 | for (auto &E : MST.AllEdges) | |||
| 784 | EdgeList.push_back(E.get()); | |||
| 785 | ||||
| 786 | for (auto &E : EdgeList) { | |||
| 787 | BasicBlock *InstrBB = getInstrBB(E); | |||
| 788 | if (InstrBB) | |||
| 789 | InstrumentBBs.push_back(InstrBB); | |||
| 790 | } | |||
| 791 | ||||
| 792 | // Set up InEdges/OutEdges for all BBs. | |||
| 793 | for (auto &E : MST.AllEdges) { | |||
| 794 | if (E->Removed) | |||
| 795 | continue; | |||
| 796 | const BasicBlock *SrcBB = E->SrcBB; | |||
| 797 | const BasicBlock *DestBB = E->DestBB; | |||
| 798 | BBInfo &SrcInfo = getBBInfo(SrcBB); | |||
| 799 | BBInfo &DestInfo = getBBInfo(DestBB); | |||
| 800 | SrcInfo.addOutEdge(E.get()); | |||
| 801 | DestInfo.addInEdge(E.get()); | |||
| 802 | } | |||
| 803 | } | |||
| 804 | ||||
| 805 | // Given a CFG E to be instrumented, find which BB to place the instrumented | |||
| 806 | // code. The function will split the critical edge if necessary. | |||
| 807 | template <class Edge, class BBInfo> | |||
| 808 | BasicBlock *FuncPGOInstrumentation<Edge, BBInfo>::getInstrBB(Edge *E) { | |||
| 809 | if (E->InMST || E->Removed) | |||
| 810 | return nullptr; | |||
| 811 | ||||
| 812 | BasicBlock *SrcBB = const_cast<BasicBlock *>(E->SrcBB); | |||
| 813 | BasicBlock *DestBB = const_cast<BasicBlock *>(E->DestBB); | |||
| 814 | // For a fake edge, instrument the real BB. | |||
| 815 | if (SrcBB == nullptr) | |||
| 816 | return DestBB; | |||
| 817 | if (DestBB == nullptr) | |||
| 818 | return SrcBB; | |||
| 819 | ||||
| 820 | auto canInstrument = [](BasicBlock *BB) -> BasicBlock * { | |||
| 821 | // There are basic blocks (such as catchswitch) cannot be instrumented. | |||
| 822 | // If the returned first insertion point is the end of BB, skip this BB. | |||
| 823 | if (BB->getFirstInsertionPt() == BB->end()) | |||
| 824 | return nullptr; | |||
| 825 | return BB; | |||
| 826 | }; | |||
| 827 | ||||
| 828 | // Instrument the SrcBB if it has a single successor, | |||
| 829 | // otherwise, the DestBB if this is not a critical edge. | |||
| 830 | Instruction *TI = SrcBB->getTerminator(); | |||
| 831 | if (TI->getNumSuccessors() <= 1) | |||
| 832 | return canInstrument(SrcBB); | |||
| 833 | if (!E->IsCritical) | |||
| 834 | return canInstrument(DestBB); | |||
| 835 | ||||
| 836 | // Some IndirectBr critical edges cannot be split by the previous | |||
| 837 | // SplitIndirectBrCriticalEdges call. Bail out. | |||
| 838 | unsigned SuccNum = GetSuccessorNumber(SrcBB, DestBB); | |||
| 839 | BasicBlock *InstrBB = | |||
| 840 | isa<IndirectBrInst>(TI) ? nullptr : SplitCriticalEdge(TI, SuccNum); | |||
| 841 | if (!InstrBB) { | |||
| 842 | LLVM_DEBUG(do { } while (false) | |||
| 843 | dbgs() << "Fail to split critical edge: not instrument this edge.\n")do { } while (false); | |||
| 844 | return nullptr; | |||
| 845 | } | |||
| 846 | // For a critical edge, we have to split. Instrument the newly | |||
| 847 | // created BB. | |||
| 848 | IsCS ? NumOfCSPGOSplit++ : NumOfPGOSplit++; | |||
| 849 | LLVM_DEBUG(dbgs() << "Split critical edge: " << getBBInfo(SrcBB).Indexdo { } while (false) | |||
| 850 | << " --> " << getBBInfo(DestBB).Index << "\n")do { } while (false); | |||
| 851 | // Need to add two new edges. First one: Add new edge of SrcBB->InstrBB. | |||
| 852 | MST.addEdge(SrcBB, InstrBB, 0); | |||
| 853 | // Second one: Add new edge of InstrBB->DestBB. | |||
| 854 | Edge &NewEdge1 = MST.addEdge(InstrBB, DestBB, 0); | |||
| 855 | NewEdge1.InMST = true; | |||
| 856 | E->Removed = true; | |||
| 857 | ||||
| 858 | return canInstrument(InstrBB); | |||
| 859 | } | |||
| 860 | ||||
| 861 | // When generating value profiling calls on Windows routines that make use of | |||
| 862 | // handler funclets for exception processing an operand bundle needs to attached | |||
| 863 | // to the called function. This routine will set \p OpBundles to contain the | |||
| 864 | // funclet information, if any is needed, that should be placed on the generated | |||
| 865 | // value profiling call for the value profile candidate call. | |||
| 866 | static void | |||
| 867 | populateEHOperandBundle(VPCandidateInfo &Cand, | |||
| 868 | DenseMap<BasicBlock *, ColorVector> &BlockColors, | |||
| 869 | SmallVectorImpl<OperandBundleDef> &OpBundles) { | |||
| 870 | auto *OrigCall = dyn_cast<CallBase>(Cand.AnnotatedInst); | |||
| 871 | if (OrigCall
| |||
| 872 | // The instrumentation call should belong to the same funclet as a | |||
| 873 | // non-intrinsic call, so just copy the operand bundle, if any exists. | |||
| 874 | Optional<OperandBundleUse> ParentFunclet = | |||
| 875 | OrigCall->getOperandBundle(LLVMContext::OB_funclet); | |||
| 876 | if (ParentFunclet) | |||
| 877 | OpBundles.emplace_back(OperandBundleDef(*ParentFunclet)); | |||
| 878 | } else { | |||
| 879 | // Intrinsics or other instructions do not get funclet information from the | |||
| 880 | // front-end. Need to use the BlockColors that was computed by the routine | |||
| 881 | // colorEHFunclets to determine whether a funclet is needed. | |||
| 882 | if (!BlockColors.empty()) { | |||
| 883 | const ColorVector &CV = BlockColors.find(OrigCall->getParent())->second; | |||
| ||||
| 884 | assert(CV.size() == 1 && "non-unique color for block!")((void)0); | |||
| 885 | Instruction *EHPad = CV.front()->getFirstNonPHI(); | |||
| 886 | if (EHPad->isEHPad()) | |||
| 887 | OpBundles.emplace_back("funclet", EHPad); | |||
| 888 | } | |||
| 889 | } | |||
| 890 | } | |||
| 891 | ||||
| 892 | // Visit all edge and instrument the edges not in MST, and do value profiling. | |||
| 893 | // Critical edges will be split. | |||
| 894 | static void instrumentOneFunc( | |||
| 895 | Function &F, Module *M, TargetLibraryInfo &TLI, BranchProbabilityInfo *BPI, | |||
| 896 | BlockFrequencyInfo *BFI, | |||
| 897 | std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers, | |||
| 898 | bool IsCS) { | |||
| 899 | // Split indirectbr critical edges here before computing the MST rather than | |||
| 900 | // later in getInstrBB() to avoid invalidating it. | |||
| 901 | SplitIndirectBrCriticalEdges(F, BPI, BFI); | |||
| 902 | ||||
| 903 | FuncPGOInstrumentation<PGOEdge, BBInfo> FuncInfo( | |||
| 904 | F, TLI, ComdatMembers, true, BPI, BFI, IsCS, PGOInstrumentEntry); | |||
| 905 | std::vector<BasicBlock *> InstrumentBBs; | |||
| 906 | FuncInfo.getInstrumentBBs(InstrumentBBs); | |||
| 907 | unsigned NumCounters = | |||
| 908 | InstrumentBBs.size() + FuncInfo.SIVisitor.getNumOfSelectInsts(); | |||
| 909 | ||||
| 910 | uint32_t I = 0; | |||
| 911 | Type *I8PtrTy = Type::getInt8PtrTy(M->getContext()); | |||
| 912 | for (auto *InstrBB : InstrumentBBs) { | |||
| 913 | IRBuilder<> Builder(InstrBB, InstrBB->getFirstInsertionPt()); | |||
| 914 | assert(Builder.GetInsertPoint() != InstrBB->end() &&((void)0) | |||
| 915 | "Cannot get the Instrumentation point")((void)0); | |||
| 916 | Builder.CreateCall( | |||
| 917 | Intrinsic::getDeclaration(M, Intrinsic::instrprof_increment), | |||
| 918 | {ConstantExpr::getBitCast(FuncInfo.FuncNameVar, I8PtrTy), | |||
| 919 | Builder.getInt64(FuncInfo.FunctionHash), Builder.getInt32(NumCounters), | |||
| 920 | Builder.getInt32(I++)}); | |||
| 921 | } | |||
| 922 | ||||
| 923 | // Now instrument select instructions: | |||
| 924 | FuncInfo.SIVisitor.instrumentSelects(F, &I, NumCounters, FuncInfo.FuncNameVar, | |||
| 925 | FuncInfo.FunctionHash); | |||
| 926 | assert(I == NumCounters)((void)0); | |||
| 927 | ||||
| 928 | if (DisableValueProfiling) | |||
| 929 | return; | |||
| 930 | ||||
| 931 | NumOfPGOICall += FuncInfo.ValueSites[IPVK_IndirectCallTarget].size(); | |||
| 932 | ||||
| 933 | // Intrinsic function calls do not have funclet operand bundles needed for | |||
| 934 | // Windows exception handling attached to them. However, if value profiling is | |||
| 935 | // inserted for one of these calls, then a funclet value will need to be set | |||
| 936 | // on the instrumentation call based on the funclet coloring. | |||
| 937 | DenseMap<BasicBlock *, ColorVector> BlockColors; | |||
| 938 | if (F.hasPersonalityFn() && | |||
| 939 | isFuncletEHPersonality(classifyEHPersonality(F.getPersonalityFn()))) | |||
| 940 | BlockColors = colorEHFunclets(F); | |||
| 941 | ||||
| 942 | // For each VP Kind, walk the VP candidates and instrument each one. | |||
| 943 | for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) { | |||
| 944 | unsigned SiteIndex = 0; | |||
| 945 | if (Kind
| |||
| 946 | continue; | |||
| 947 | ||||
| 948 | for (VPCandidateInfo Cand : FuncInfo.ValueSites[Kind]) { | |||
| 949 | LLVM_DEBUG(dbgs() << "Instrument one VP " << ValueProfKindDescr[Kind]do { } while (false) | |||
| 950 | << " site: CallSite Index = " << SiteIndex << "\n")do { } while (false); | |||
| 951 | ||||
| 952 | IRBuilder<> Builder(Cand.InsertPt); | |||
| 953 | assert(Builder.GetInsertPoint() != Cand.InsertPt->getParent()->end() &&((void)0) | |||
| 954 | "Cannot get the Instrumentation point")((void)0); | |||
| 955 | ||||
| 956 | Value *ToProfile = nullptr; | |||
| 957 | if (Cand.V->getType()->isIntegerTy()) | |||
| 958 | ToProfile = Builder.CreateZExtOrTrunc(Cand.V, Builder.getInt64Ty()); | |||
| 959 | else if (Cand.V->getType()->isPointerTy()) | |||
| 960 | ToProfile = Builder.CreatePtrToInt(Cand.V, Builder.getInt64Ty()); | |||
| 961 | assert(ToProfile && "value profiling Value is of unexpected type")((void)0); | |||
| 962 | ||||
| 963 | SmallVector<OperandBundleDef, 1> OpBundles; | |||
| 964 | populateEHOperandBundle(Cand, BlockColors, OpBundles); | |||
| 965 | Builder.CreateCall( | |||
| 966 | Intrinsic::getDeclaration(M, Intrinsic::instrprof_value_profile), | |||
| 967 | {ConstantExpr::getBitCast(FuncInfo.FuncNameVar, I8PtrTy), | |||
| 968 | Builder.getInt64(FuncInfo.FunctionHash), ToProfile, | |||
| 969 | Builder.getInt32(Kind), Builder.getInt32(SiteIndex++)}, | |||
| 970 | OpBundles); | |||
| 971 | } | |||
| 972 | } // IPVK_First <= Kind <= IPVK_Last | |||
| 973 | } | |||
| 974 | ||||
| 975 | namespace { | |||
| 976 | ||||
| 977 | // This class represents a CFG edge in profile use compilation. | |||
| 978 | struct PGOUseEdge : public PGOEdge { | |||
| 979 | bool CountValid = false; | |||
| 980 | uint64_t CountValue = 0; | |||
| 981 | ||||
| 982 | PGOUseEdge(const BasicBlock *Src, const BasicBlock *Dest, uint64_t W = 1) | |||
| 983 | : PGOEdge(Src, Dest, W) {} | |||
| 984 | ||||
| 985 | // Set edge count value | |||
| 986 | void setEdgeCount(uint64_t Value) { | |||
| 987 | CountValue = Value; | |||
| 988 | CountValid = true; | |||
| 989 | } | |||
| 990 | ||||
| 991 | // Return the information string for this object. | |||
| 992 | std::string infoString() const { | |||
| 993 | if (!CountValid) | |||
| 994 | return PGOEdge::infoString(); | |||
| 995 | return (Twine(PGOEdge::infoString()) + " Count=" + Twine(CountValue)) | |||
| 996 | .str(); | |||
| 997 | } | |||
| 998 | }; | |||
| 999 | ||||
| 1000 | using DirectEdges = SmallVector<PGOUseEdge *, 2>; | |||
| 1001 | ||||
| 1002 | // This class stores the auxiliary information for each BB. | |||
| 1003 | struct UseBBInfo : public BBInfo { | |||
| 1004 | uint64_t CountValue = 0; | |||
| 1005 | bool CountValid; | |||
| 1006 | int32_t UnknownCountInEdge = 0; | |||
| 1007 | int32_t UnknownCountOutEdge = 0; | |||
| 1008 | DirectEdges InEdges; | |||
| 1009 | DirectEdges OutEdges; | |||
| 1010 | ||||
| 1011 | UseBBInfo(unsigned IX) : BBInfo(IX), CountValid(false) {} | |||
| 1012 | ||||
| 1013 | UseBBInfo(unsigned IX, uint64_t C) | |||
| 1014 | : BBInfo(IX), CountValue(C), CountValid(true) {} | |||
| 1015 | ||||
| 1016 | // Set the profile count value for this BB. | |||
| 1017 | void setBBInfoCount(uint64_t Value) { | |||
| 1018 | CountValue = Value; | |||
| 1019 | CountValid = true; | |||
| 1020 | } | |||
| 1021 | ||||
| 1022 | // Return the information string of this object. | |||
| 1023 | std::string infoString() const { | |||
| 1024 | if (!CountValid) | |||
| 1025 | return BBInfo::infoString(); | |||
| 1026 | return (Twine(BBInfo::infoString()) + " Count=" + Twine(CountValue)).str(); | |||
| 1027 | } | |||
| 1028 | ||||
| 1029 | // Add an OutEdge and update the edge count. | |||
| 1030 | void addOutEdge(PGOUseEdge *E) { | |||
| 1031 | OutEdges.push_back(E); | |||
| 1032 | UnknownCountOutEdge++; | |||
| 1033 | } | |||
| 1034 | ||||
| 1035 | // Add an InEdge and update the edge count. | |||
| 1036 | void addInEdge(PGOUseEdge *E) { | |||
| 1037 | InEdges.push_back(E); | |||
| 1038 | UnknownCountInEdge++; | |||
| 1039 | } | |||
| 1040 | }; | |||
| 1041 | ||||
| 1042 | } // end anonymous namespace | |||
| 1043 | ||||
| 1044 | // Sum up the count values for all the edges. | |||
| 1045 | static uint64_t sumEdgeCount(const ArrayRef<PGOUseEdge *> Edges) { | |||
| 1046 | uint64_t Total = 0; | |||
| 1047 | for (auto &E : Edges) { | |||
| 1048 | if (E->Removed) | |||
| 1049 | continue; | |||
| 1050 | Total += E->CountValue; | |||
| 1051 | } | |||
| 1052 | return Total; | |||
| 1053 | } | |||
| 1054 | ||||
| 1055 | namespace { | |||
| 1056 | ||||
| 1057 | class PGOUseFunc { | |||
| 1058 | public: | |||
| 1059 | PGOUseFunc(Function &Func, Module *Modu, TargetLibraryInfo &TLI, | |||
| 1060 | std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers, | |||
| 1061 | BranchProbabilityInfo *BPI, BlockFrequencyInfo *BFIin, | |||
| 1062 | ProfileSummaryInfo *PSI, bool IsCS, bool InstrumentFuncEntry) | |||
| 1063 | : F(Func), M(Modu), BFI(BFIin), PSI(PSI), | |||
| 1064 | FuncInfo(Func, TLI, ComdatMembers, false, BPI, BFIin, IsCS, | |||
| 1065 | InstrumentFuncEntry), | |||
| 1066 | FreqAttr(FFA_Normal), IsCS(IsCS) {} | |||
| 1067 | ||||
| 1068 | // Read counts for the instrumented BB from profile. | |||
| 1069 | bool readCounters(IndexedInstrProfReader *PGOReader, bool &AllZeros, | |||
| 1070 | bool &AllMinusOnes); | |||
| 1071 | ||||
| 1072 | // Populate the counts for all BBs. | |||
| 1073 | void populateCounters(); | |||
| 1074 | ||||
| 1075 | // Set the branch weights based on the count values. | |||
| 1076 | void setBranchWeights(); | |||
| 1077 | ||||
| 1078 | // Annotate the value profile call sites for all value kind. | |||
| 1079 | void annotateValueSites(); | |||
| 1080 | ||||
| 1081 | // Annotate the value profile call sites for one value kind. | |||
| 1082 | void annotateValueSites(uint32_t Kind); | |||
| 1083 | ||||
| 1084 | // Annotate the irreducible loop header weights. | |||
| 1085 | void annotateIrrLoopHeaderWeights(); | |||
| 1086 | ||||
| 1087 | // The hotness of the function from the profile count. | |||
| 1088 | enum FuncFreqAttr { FFA_Normal, FFA_Cold, FFA_Hot }; | |||
| 1089 | ||||
| 1090 | // Return the function hotness from the profile. | |||
| 1091 | FuncFreqAttr getFuncFreqAttr() const { return FreqAttr; } | |||
| 1092 | ||||
| 1093 | // Return the function hash. | |||
| 1094 | uint64_t getFuncHash() const { return FuncInfo.FunctionHash; } | |||
| 1095 | ||||
| 1096 | // Return the profile record for this function; | |||
| 1097 | InstrProfRecord &getProfileRecord() { return ProfileRecord; } | |||
| 1098 | ||||
| 1099 | // Return the auxiliary BB information. | |||
| 1100 | UseBBInfo &getBBInfo(const BasicBlock *BB) const { | |||
| 1101 | return FuncInfo.getBBInfo(BB); | |||
| 1102 | } | |||
| 1103 | ||||
| 1104 | // Return the auxiliary BB information if available. | |||
| 1105 | UseBBInfo *findBBInfo(const BasicBlock *BB) const { | |||
| 1106 | return FuncInfo.findBBInfo(BB); | |||
| 1107 | } | |||
| 1108 | ||||
| 1109 | Function &getFunc() const { return F; } | |||
| 1110 | ||||
| 1111 | void dumpInfo(std::string Str = "") const { | |||
| 1112 | FuncInfo.dumpInfo(Str); | |||
| 1113 | } | |||
| 1114 | ||||
| 1115 | uint64_t getProgramMaxCount() const { return ProgramMaxCount; } | |||
| 1116 | private: | |||
| 1117 | Function &F; | |||
| 1118 | Module *M; | |||
| 1119 | BlockFrequencyInfo *BFI; | |||
| 1120 | ProfileSummaryInfo *PSI; | |||
| 1121 | ||||
| 1122 | // This member stores the shared information with class PGOGenFunc. | |||
| 1123 | FuncPGOInstrumentation<PGOUseEdge, UseBBInfo> FuncInfo; | |||
| 1124 | ||||
| 1125 | // The maximum count value in the profile. This is only used in PGO use | |||
| 1126 | // compilation. | |||
| 1127 | uint64_t ProgramMaxCount; | |||
| 1128 | ||||
| 1129 | // Position of counter that remains to be read. | |||
| 1130 | uint32_t CountPosition = 0; | |||
| 1131 | ||||
| 1132 | // Total size of the profile count for this function. | |||
| 1133 | uint32_t ProfileCountSize = 0; | |||
| 1134 | ||||
| 1135 | // ProfileRecord for this function. | |||
| 1136 | InstrProfRecord ProfileRecord; | |||
| 1137 | ||||
| 1138 | // Function hotness info derived from profile. | |||
| 1139 | FuncFreqAttr FreqAttr; | |||
| 1140 | ||||
| 1141 | // Is to use the context sensitive profile. | |||
| 1142 | bool IsCS; | |||
| 1143 | ||||
| 1144 | // Find the Instrumented BB and set the value. Return false on error. | |||
| 1145 | bool setInstrumentedCounts(const std::vector<uint64_t> &CountFromProfile); | |||
| 1146 | ||||
| 1147 | // Set the edge counter value for the unknown edge -- there should be only | |||
| 1148 | // one unknown edge. | |||
| 1149 | void setEdgeCount(DirectEdges &Edges, uint64_t Value); | |||
| 1150 | ||||
| 1151 | // Return FuncName string; | |||
| 1152 | std::string getFuncName() const { return FuncInfo.FuncName; } | |||
| 1153 | ||||
| 1154 | // Set the hot/cold inline hints based on the count values. | |||
| 1155 | // FIXME: This function should be removed once the functionality in | |||
| 1156 | // the inliner is implemented. | |||
| 1157 | void markFunctionAttributes(uint64_t EntryCount, uint64_t MaxCount) { | |||
| 1158 | if (PSI->isHotCount(EntryCount)) | |||
| 1159 | FreqAttr = FFA_Hot; | |||
| 1160 | else if (PSI->isColdCount(MaxCount)) | |||
| 1161 | FreqAttr = FFA_Cold; | |||
| 1162 | } | |||
| 1163 | }; | |||
| 1164 | ||||
| 1165 | } // end anonymous namespace | |||
| 1166 | ||||
| 1167 | // Visit all the edges and assign the count value for the instrumented | |||
| 1168 | // edges and the BB. Return false on error. | |||
| 1169 | bool PGOUseFunc::setInstrumentedCounts( | |||
| 1170 | const std::vector<uint64_t> &CountFromProfile) { | |||
| 1171 | ||||
| 1172 | std::vector<BasicBlock *> InstrumentBBs; | |||
| 1173 | FuncInfo.getInstrumentBBs(InstrumentBBs); | |||
| 1174 | unsigned NumCounters = | |||
| 1175 | InstrumentBBs.size() + FuncInfo.SIVisitor.getNumOfSelectInsts(); | |||
| 1176 | // The number of counters here should match the number of counters | |||
| 1177 | // in profile. Return if they mismatch. | |||
| 1178 | if (NumCounters != CountFromProfile.size()) { | |||
| 1179 | return false; | |||
| 1180 | } | |||
| 1181 | auto *FuncEntry = &*F.begin(); | |||
| 1182 | ||||
| 1183 | // Set the profile count to the Instrumented BBs. | |||
| 1184 | uint32_t I = 0; | |||
| 1185 | for (BasicBlock *InstrBB : InstrumentBBs) { | |||
| 1186 | uint64_t CountValue = CountFromProfile[I++]; | |||
| 1187 | UseBBInfo &Info = getBBInfo(InstrBB); | |||
| 1188 | // If we reach here, we know that we have some nonzero count | |||
| 1189 | // values in this function. The entry count should not be 0. | |||
| 1190 | // Fix it if necessary. | |||
| 1191 | if (InstrBB == FuncEntry && CountValue == 0) | |||
| 1192 | CountValue = 1; | |||
| 1193 | Info.setBBInfoCount(CountValue); | |||
| 1194 | } | |||
| 1195 | ProfileCountSize = CountFromProfile.size(); | |||
| 1196 | CountPosition = I; | |||
| 1197 | ||||
| 1198 | // Set the edge count and update the count of unknown edges for BBs. | |||
| 1199 | auto setEdgeCount = [this](PGOUseEdge *E, uint64_t Value) -> void { | |||
| 1200 | E->setEdgeCount(Value); | |||
| 1201 | this->getBBInfo(E->SrcBB).UnknownCountOutEdge--; | |||
| 1202 | this->getBBInfo(E->DestBB).UnknownCountInEdge--; | |||
| 1203 | }; | |||
| 1204 | ||||
| 1205 | // Set the profile count the Instrumented edges. There are BBs that not in | |||
| 1206 | // MST but not instrumented. Need to set the edge count value so that we can | |||
| 1207 | // populate the profile counts later. | |||
| 1208 | for (auto &E : FuncInfo.MST.AllEdges) { | |||
| 1209 | if (E->Removed || E->InMST) | |||
| 1210 | continue; | |||
| 1211 | const BasicBlock *SrcBB = E->SrcBB; | |||
| 1212 | UseBBInfo &SrcInfo = getBBInfo(SrcBB); | |||
| 1213 | ||||
| 1214 | // If only one out-edge, the edge profile count should be the same as BB | |||
| 1215 | // profile count. | |||
| 1216 | if (SrcInfo.CountValid && SrcInfo.OutEdges.size() == 1) | |||
| 1217 | setEdgeCount(E.get(), SrcInfo.CountValue); | |||
| 1218 | else { | |||
| 1219 | const BasicBlock *DestBB = E->DestBB; | |||
| 1220 | UseBBInfo &DestInfo = getBBInfo(DestBB); | |||
| 1221 | // If only one in-edge, the edge profile count should be the same as BB | |||
| 1222 | // profile count. | |||
| 1223 | if (DestInfo.CountValid && DestInfo.InEdges.size() == 1) | |||
| 1224 | setEdgeCount(E.get(), DestInfo.CountValue); | |||
| 1225 | } | |||
| 1226 | if (E->CountValid) | |||
| 1227 | continue; | |||
| 1228 | // E's count should have been set from profile. If not, this meenas E skips | |||
| 1229 | // the instrumentation. We set the count to 0. | |||
| 1230 | setEdgeCount(E.get(), 0); | |||
| 1231 | } | |||
| 1232 | return true; | |||
| 1233 | } | |||
| 1234 | ||||
| 1235 | // Set the count value for the unknown edge. There should be one and only one | |||
| 1236 | // unknown edge in Edges vector. | |||
| 1237 | void PGOUseFunc::setEdgeCount(DirectEdges &Edges, uint64_t Value) { | |||
| 1238 | for (auto &E : Edges) { | |||
| 1239 | if (E->CountValid) | |||
| 1240 | continue; | |||
| 1241 | E->setEdgeCount(Value); | |||
| 1242 | ||||
| 1243 | getBBInfo(E->SrcBB).UnknownCountOutEdge--; | |||
| 1244 | getBBInfo(E->DestBB).UnknownCountInEdge--; | |||
| 1245 | return; | |||
| 1246 | } | |||
| 1247 | llvm_unreachable("Cannot find the unknown count edge")__builtin_unreachable(); | |||
| 1248 | } | |||
| 1249 | ||||
| 1250 | // Emit function metadata indicating PGO profile mismatch. | |||
| 1251 | static void annotateFunctionWithHashMismatch(Function &F, | |||
| 1252 | LLVMContext &ctx) { | |||
| 1253 | const char MetadataName[] = "instr_prof_hash_mismatch"; | |||
| 1254 | SmallVector<Metadata *, 2> Names; | |||
| 1255 | // If this metadata already exists, ignore. | |||
| 1256 | auto *Existing = F.getMetadata(LLVMContext::MD_annotation); | |||
| 1257 | if (Existing) { | |||
| 1258 | MDTuple *Tuple = cast<MDTuple>(Existing); | |||
| 1259 | for (auto &N : Tuple->operands()) { | |||
| 1260 | if (cast<MDString>(N.get())->getString() == MetadataName) | |||
| 1261 | return; | |||
| 1262 | Names.push_back(N.get()); | |||
| 1263 | } | |||
| 1264 | } | |||
| 1265 | ||||
| 1266 | MDBuilder MDB(ctx); | |||
| 1267 | Names.push_back(MDB.createString(MetadataName)); | |||
| 1268 | MDNode *MD = MDTuple::get(ctx, Names); | |||
| 1269 | F.setMetadata(LLVMContext::MD_annotation, MD); | |||
| 1270 | } | |||
| 1271 | ||||
| 1272 | // Read the profile from ProfileFileName and assign the value to the | |||
| 1273 | // instrumented BB and the edges. This function also updates ProgramMaxCount. | |||
| 1274 | // Return true if the profile are successfully read, and false on errors. | |||
| 1275 | bool PGOUseFunc::readCounters(IndexedInstrProfReader *PGOReader, bool &AllZeros, | |||
| 1276 | bool &AllMinusOnes) { | |||
| 1277 | auto &Ctx = M->getContext(); | |||
| 1278 | Expected<InstrProfRecord> Result = | |||
| 1279 | PGOReader->getInstrProfRecord(FuncInfo.FuncName, FuncInfo.FunctionHash); | |||
| 1280 | if (Error E = Result.takeError()) { | |||
| 1281 | handleAllErrors(std::move(E), [&](const InstrProfError &IPE) { | |||
| 1282 | auto Err = IPE.get(); | |||
| 1283 | bool SkipWarning = false; | |||
| 1284 | LLVM_DEBUG(dbgs() << "Error in reading profile for Func "do { } while (false) | |||
| 1285 | << FuncInfo.FuncName << ": ")do { } while (false); | |||
| 1286 | if (Err == instrprof_error::unknown_function) { | |||
| 1287 | IsCS ? NumOfCSPGOMissing++ : NumOfPGOMissing++; | |||
| 1288 | SkipWarning = !PGOWarnMissing; | |||
| 1289 | LLVM_DEBUG(dbgs() << "unknown function")do { } while (false); | |||
| 1290 | } else if (Err == instrprof_error::hash_mismatch || | |||
| 1291 | Err == instrprof_error::malformed) { | |||
| 1292 | IsCS ? NumOfCSPGOMismatch++ : NumOfPGOMismatch++; | |||
| 1293 | SkipWarning = | |||
| 1294 | NoPGOWarnMismatch || | |||
| 1295 | (NoPGOWarnMismatchComdat && | |||
| 1296 | (F.hasComdat() || | |||
| 1297 | F.getLinkage() == GlobalValue::AvailableExternallyLinkage)); | |||
| 1298 | LLVM_DEBUG(dbgs() << "hash mismatch (skip=" << SkipWarning << ")")do { } while (false); | |||
| 1299 | // Emit function metadata indicating PGO profile mismatch. | |||
| 1300 | annotateFunctionWithHashMismatch(F, M->getContext()); | |||
| 1301 | } | |||
| 1302 | ||||
| 1303 | LLVM_DEBUG(dbgs() << " IsCS=" << IsCS << "\n")do { } while (false); | |||
| 1304 | if (SkipWarning) | |||
| 1305 | return; | |||
| 1306 | ||||
| 1307 | std::string Msg = IPE.message() + std::string(" ") + F.getName().str() + | |||
| 1308 | std::string(" Hash = ") + | |||
| 1309 | std::to_string(FuncInfo.FunctionHash); | |||
| 1310 | ||||
| 1311 | Ctx.diagnose( | |||
| 1312 | DiagnosticInfoPGOProfile(M->getName().data(), Msg, DS_Warning)); | |||
| 1313 | }); | |||
| 1314 | return false; | |||
| 1315 | } | |||
| 1316 | ProfileRecord = std::move(Result.get()); | |||
| 1317 | std::vector<uint64_t> &CountFromProfile = ProfileRecord.Counts; | |||
| 1318 | ||||
| 1319 | IsCS ? NumOfCSPGOFunc++ : NumOfPGOFunc++; | |||
| 1320 | LLVM_DEBUG(dbgs() << CountFromProfile.size() << " counts\n")do { } while (false); | |||
| 1321 | AllMinusOnes = (CountFromProfile.size() > 0); | |||
| 1322 | uint64_t ValueSum = 0; | |||
| 1323 | for (unsigned I = 0, S = CountFromProfile.size(); I < S; I++) { | |||
| 1324 | LLVM_DEBUG(dbgs() << " " << I << ": " << CountFromProfile[I] << "\n")do { } while (false); | |||
| 1325 | ValueSum += CountFromProfile[I]; | |||
| 1326 | if (CountFromProfile[I] != (uint64_t)-1) | |||
| 1327 | AllMinusOnes = false; | |||
| 1328 | } | |||
| 1329 | AllZeros = (ValueSum == 0); | |||
| 1330 | ||||
| 1331 | LLVM_DEBUG(dbgs() << "SUM = " << ValueSum << "\n")do { } while (false); | |||
| 1332 | ||||
| 1333 | getBBInfo(nullptr).UnknownCountOutEdge = 2; | |||
| 1334 | getBBInfo(nullptr).UnknownCountInEdge = 2; | |||
| 1335 | ||||
| 1336 | if (!setInstrumentedCounts(CountFromProfile)) { | |||
| 1337 | LLVM_DEBUG(do { } while (false) | |||
| 1338 | dbgs() << "Inconsistent number of counts, skipping this function")do { } while (false); | |||
| 1339 | Ctx.diagnose(DiagnosticInfoPGOProfile( | |||
| 1340 | M->getName().data(), | |||
| 1341 | Twine("Inconsistent number of counts in ") + F.getName().str() | |||
| 1342 | + Twine(": the profile may be stale or there is a function name collision."), | |||
| 1343 | DS_Warning)); | |||
| 1344 | return false; | |||
| 1345 | } | |||
| 1346 | ProgramMaxCount = PGOReader->getMaximumFunctionCount(IsCS); | |||
| 1347 | return true; | |||
| 1348 | } | |||
| 1349 | ||||
| 1350 | // Populate the counters from instrumented BBs to all BBs. | |||
| 1351 | // In the end of this operation, all BBs should have a valid count value. | |||
| 1352 | void PGOUseFunc::populateCounters() { | |||
| 1353 | bool Changes = true; | |||
| 1354 | unsigned NumPasses = 0; | |||
| 1355 | while (Changes) { | |||
| 1356 | NumPasses++; | |||
| 1357 | Changes = false; | |||
| 1358 | ||||
| 1359 | // For efficient traversal, it's better to start from the end as most | |||
| 1360 | // of the instrumented edges are at the end. | |||
| 1361 | for (auto &BB : reverse(F)) { | |||
| 1362 | UseBBInfo *Count = findBBInfo(&BB); | |||
| 1363 | if (Count == nullptr) | |||
| 1364 | continue; | |||
| 1365 | if (!Count->CountValid) { | |||
| 1366 | if (Count->UnknownCountOutEdge == 0) { | |||
| 1367 | Count->CountValue = sumEdgeCount(Count->OutEdges); | |||
| 1368 | Count->CountValid = true; | |||
| 1369 | Changes = true; | |||
| 1370 | } else if (Count->UnknownCountInEdge == 0) { | |||
| 1371 | Count->CountValue = sumEdgeCount(Count->InEdges); | |||
| 1372 | Count->CountValid = true; | |||
| 1373 | Changes = true; | |||
| 1374 | } | |||
| 1375 | } | |||
| 1376 | if (Count->CountValid) { | |||
| 1377 | if (Count->UnknownCountOutEdge == 1) { | |||
| 1378 | uint64_t Total = 0; | |||
| 1379 | uint64_t OutSum = sumEdgeCount(Count->OutEdges); | |||
| 1380 | // If the one of the successor block can early terminate (no-return), | |||
| 1381 | // we can end up with situation where out edge sum count is larger as | |||
| 1382 | // the source BB's count is collected by a post-dominated block. | |||
| 1383 | if (Count->CountValue > OutSum) | |||
| 1384 | Total = Count->CountValue - OutSum; | |||
| 1385 | setEdgeCount(Count->OutEdges, Total); | |||
| 1386 | Changes = true; | |||
| 1387 | } | |||
| 1388 | if (Count->UnknownCountInEdge == 1) { | |||
| 1389 | uint64_t Total = 0; | |||
| 1390 | uint64_t InSum = sumEdgeCount(Count->InEdges); | |||
| 1391 | if (Count->CountValue > InSum) | |||
| 1392 | Total = Count->CountValue - InSum; | |||
| 1393 | setEdgeCount(Count->InEdges, Total); | |||
| 1394 | Changes = true; | |||
| 1395 | } | |||
| 1396 | } | |||
| 1397 | } | |||
| 1398 | } | |||
| 1399 | ||||
| 1400 | LLVM_DEBUG(dbgs() << "Populate counts in " << NumPasses << " passes.\n")do { } while (false); | |||
| 1401 | #ifndef NDEBUG1 | |||
| 1402 | // Assert every BB has a valid counter. | |||
| 1403 | for (auto &BB : F) { | |||
| 1404 | auto BI = findBBInfo(&BB); | |||
| 1405 | if (BI == nullptr) | |||
| 1406 | continue; | |||
| 1407 | assert(BI->CountValid && "BB count is not valid")((void)0); | |||
| 1408 | } | |||
| 1409 | #endif | |||
| 1410 | uint64_t FuncEntryCount = getBBInfo(&*F.begin()).CountValue; | |||
| 1411 | uint64_t FuncMaxCount = FuncEntryCount; | |||
| 1412 | for (auto &BB : F) { | |||
| 1413 | auto BI = findBBInfo(&BB); | |||
| 1414 | if (BI == nullptr) | |||
| 1415 | continue; | |||
| 1416 | FuncMaxCount = std::max(FuncMaxCount, BI->CountValue); | |||
| 1417 | } | |||
| 1418 | ||||
| 1419 | // Fix the obviously inconsistent entry count. | |||
| 1420 | if (FuncMaxCount > 0 && FuncEntryCount == 0) | |||
| 1421 | FuncEntryCount = 1; | |||
| 1422 | F.setEntryCount(ProfileCount(FuncEntryCount, Function::PCT_Real)); | |||
| 1423 | markFunctionAttributes(FuncEntryCount, FuncMaxCount); | |||
| 1424 | ||||
| 1425 | // Now annotate select instructions | |||
| 1426 | FuncInfo.SIVisitor.annotateSelects(F, this, &CountPosition); | |||
| 1427 | assert(CountPosition == ProfileCountSize)((void)0); | |||
| 1428 | ||||
| 1429 | LLVM_DEBUG(FuncInfo.dumpInfo("after reading profile."))do { } while (false); | |||
| 1430 | } | |||
| 1431 | ||||
| 1432 | // Assign the scaled count values to the BB with multiple out edges. | |||
| 1433 | void PGOUseFunc::setBranchWeights() { | |||
| 1434 | // Generate MD_prof metadata for every branch instruction. | |||
| 1435 | LLVM_DEBUG(dbgs() << "\nSetting branch weights for func " << F.getName()do { } while (false) | |||
| 1436 | << " IsCS=" << IsCS << "\n")do { } while (false); | |||
| 1437 | for (auto &BB : F) { | |||
| 1438 | Instruction *TI = BB.getTerminator(); | |||
| 1439 | if (TI->getNumSuccessors() < 2) | |||
| 1440 | continue; | |||
| 1441 | if (!(isa<BranchInst>(TI) || isa<SwitchInst>(TI) || | |||
| 1442 | isa<IndirectBrInst>(TI) || isa<InvokeInst>(TI))) | |||
| 1443 | continue; | |||
| 1444 | ||||
| 1445 | if (getBBInfo(&BB).CountValue == 0) | |||
| 1446 | continue; | |||
| 1447 | ||||
| 1448 | // We have a non-zero Branch BB. | |||
| 1449 | const UseBBInfo &BBCountInfo = getBBInfo(&BB); | |||
| 1450 | unsigned Size = BBCountInfo.OutEdges.size(); | |||
| 1451 | SmallVector<uint64_t, 2> EdgeCounts(Size, 0); | |||
| 1452 | uint64_t MaxCount = 0; | |||
| 1453 | for (unsigned s = 0; s < Size; s++) { | |||
| 1454 | const PGOUseEdge *E = BBCountInfo.OutEdges[s]; | |||
| 1455 | const BasicBlock *SrcBB = E->SrcBB; | |||
| 1456 | const BasicBlock *DestBB = E->DestBB; | |||
| 1457 | if (DestBB == nullptr) | |||
| 1458 | continue; | |||
| 1459 | unsigned SuccNum = GetSuccessorNumber(SrcBB, DestBB); | |||
| 1460 | uint64_t EdgeCount = E->CountValue; | |||
| 1461 | if (EdgeCount > MaxCount) | |||
| 1462 | MaxCount = EdgeCount; | |||
| 1463 | EdgeCounts[SuccNum] = EdgeCount; | |||
| 1464 | } | |||
| 1465 | setProfMetadata(M, TI, EdgeCounts, MaxCount); | |||
| 1466 | } | |||
| 1467 | } | |||
| 1468 | ||||
| 1469 | static bool isIndirectBrTarget(BasicBlock *BB) { | |||
| 1470 | for (BasicBlock *Pred : predecessors(BB)) { | |||
| 1471 | if (isa<IndirectBrInst>(Pred->getTerminator())) | |||
| 1472 | return true; | |||
| 1473 | } | |||
| 1474 | return false; | |||
| 1475 | } | |||
| 1476 | ||||
| 1477 | void PGOUseFunc::annotateIrrLoopHeaderWeights() { | |||
| 1478 | LLVM_DEBUG(dbgs() << "\nAnnotating irreducible loop header weights.\n")do { } while (false); | |||
| 1479 | // Find irr loop headers | |||
| 1480 | for (auto &BB : F) { | |||
| 1481 | // As a heuristic also annotate indrectbr targets as they have a high chance | |||
| 1482 | // to become an irreducible loop header after the indirectbr tail | |||
| 1483 | // duplication. | |||
| 1484 | if (BFI->isIrrLoopHeader(&BB) || isIndirectBrTarget(&BB)) { | |||
| 1485 | Instruction *TI = BB.getTerminator(); | |||
| 1486 | const UseBBInfo &BBCountInfo = getBBInfo(&BB); | |||
| 1487 | setIrrLoopHeaderMetadata(M, TI, BBCountInfo.CountValue); | |||
| 1488 | } | |||
| 1489 | } | |||
| 1490 | } | |||
| 1491 | ||||
| 1492 | void SelectInstVisitor::instrumentOneSelectInst(SelectInst &SI) { | |||
| 1493 | Module *M = F.getParent(); | |||
| 1494 | IRBuilder<> Builder(&SI); | |||
| 1495 | Type *Int64Ty = Builder.getInt64Ty(); | |||
| 1496 | Type *I8PtrTy = Builder.getInt8PtrTy(); | |||
| 1497 | auto *Step = Builder.CreateZExt(SI.getCondition(), Int64Ty); | |||
| 1498 | Builder.CreateCall( | |||
| 1499 | Intrinsic::getDeclaration(M, Intrinsic::instrprof_increment_step), | |||
| 1500 | {ConstantExpr::getBitCast(FuncNameVar, I8PtrTy), | |||
| 1501 | Builder.getInt64(FuncHash), Builder.getInt32(TotalNumCtrs), | |||
| 1502 | Builder.getInt32(*CurCtrIdx), Step}); | |||
| 1503 | ++(*CurCtrIdx); | |||
| 1504 | } | |||
| 1505 | ||||
| 1506 | void SelectInstVisitor::annotateOneSelectInst(SelectInst &SI) { | |||
| 1507 | std::vector<uint64_t> &CountFromProfile = UseFunc->getProfileRecord().Counts; | |||
| 1508 | assert(*CurCtrIdx < CountFromProfile.size() &&((void)0) | |||
| 1509 | "Out of bound access of counters")((void)0); | |||
| 1510 | uint64_t SCounts[2]; | |||
| 1511 | SCounts[0] = CountFromProfile[*CurCtrIdx]; // True count | |||
| 1512 | ++(*CurCtrIdx); | |||
| 1513 | uint64_t TotalCount = 0; | |||
| 1514 | auto BI = UseFunc->findBBInfo(SI.getParent()); | |||
| 1515 | if (BI != nullptr) | |||
| 1516 | TotalCount = BI->CountValue; | |||
| 1517 | // False Count | |||
| 1518 | SCounts[1] = (TotalCount > SCounts[0] ? TotalCount - SCounts[0] : 0); | |||
| 1519 | uint64_t MaxCount = std::max(SCounts[0], SCounts[1]); | |||
| 1520 | if (MaxCount) | |||
| 1521 | setProfMetadata(F.getParent(), &SI, SCounts, MaxCount); | |||
| 1522 | } | |||
| 1523 | ||||
| 1524 | void SelectInstVisitor::visitSelectInst(SelectInst &SI) { | |||
| 1525 | if (!PGOInstrSelect) | |||
| 1526 | return; | |||
| 1527 | // FIXME: do not handle this yet. | |||
| 1528 | if (SI.getCondition()->getType()->isVectorTy()) | |||
| 1529 | return; | |||
| 1530 | ||||
| 1531 | switch (Mode) { | |||
| 1532 | case VM_counting: | |||
| 1533 | NSIs++; | |||
| 1534 | return; | |||
| 1535 | case VM_instrument: | |||
| 1536 | instrumentOneSelectInst(SI); | |||
| 1537 | return; | |||
| 1538 | case VM_annotate: | |||
| 1539 | annotateOneSelectInst(SI); | |||
| 1540 | return; | |||
| 1541 | } | |||
| 1542 | ||||
| 1543 | llvm_unreachable("Unknown visiting mode")__builtin_unreachable(); | |||
| 1544 | } | |||
| 1545 | ||||
| 1546 | // Traverse all valuesites and annotate the instructions for all value kind. | |||
| 1547 | void PGOUseFunc::annotateValueSites() { | |||
| 1548 | if (DisableValueProfiling) | |||
| 1549 | return; | |||
| 1550 | ||||
| 1551 | // Create the PGOFuncName meta data. | |||
| 1552 | createPGOFuncNameMetadata(F, FuncInfo.FuncName); | |||
| 1553 | ||||
| 1554 | for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) | |||
| 1555 | annotateValueSites(Kind); | |||
| 1556 | } | |||
| 1557 | ||||
| 1558 | // Annotate the instructions for a specific value kind. | |||
| 1559 | void PGOUseFunc::annotateValueSites(uint32_t Kind) { | |||
| 1560 | assert(Kind <= IPVK_Last)((void)0); | |||
| 1561 | unsigned ValueSiteIndex = 0; | |||
| 1562 | auto &ValueSites = FuncInfo.ValueSites[Kind]; | |||
| 1563 | unsigned NumValueSites = ProfileRecord.getNumValueSites(Kind); | |||
| 1564 | if (NumValueSites != ValueSites.size()) { | |||
| 1565 | auto &Ctx = M->getContext(); | |||
| 1566 | Ctx.diagnose(DiagnosticInfoPGOProfile( | |||
| 1567 | M->getName().data(), | |||
| 1568 | Twine("Inconsistent number of value sites for ") + | |||
| 1569 | Twine(ValueProfKindDescr[Kind]) + | |||
| 1570 | Twine(" profiling in \"") + F.getName().str() + | |||
| 1571 | Twine("\", possibly due to the use of a stale profile."), | |||
| 1572 | DS_Warning)); | |||
| 1573 | return; | |||
| 1574 | } | |||
| 1575 | ||||
| 1576 | for (VPCandidateInfo &I : ValueSites) { | |||
| 1577 | LLVM_DEBUG(dbgs() << "Read one value site profile (kind = " << Kinddo { } while (false) | |||
| 1578 | << "): Index = " << ValueSiteIndex << " out of "do { } while (false) | |||
| 1579 | << NumValueSites << "\n")do { } while (false); | |||
| 1580 | annotateValueSite(*M, *I.AnnotatedInst, ProfileRecord, | |||
| 1581 | static_cast<InstrProfValueKind>(Kind), ValueSiteIndex, | |||
| 1582 | Kind == IPVK_MemOPSize ? MaxNumMemOPAnnotations | |||
| 1583 | : MaxNumAnnotations); | |||
| 1584 | ValueSiteIndex++; | |||
| 1585 | } | |||
| 1586 | } | |||
| 1587 | ||||
| 1588 | // Collect the set of members for each Comdat in module M and store | |||
| 1589 | // in ComdatMembers. | |||
| 1590 | static void collectComdatMembers( | |||
| 1591 | Module &M, | |||
| 1592 | std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers) { | |||
| 1593 | if (!DoComdatRenaming) | |||
| 1594 | return; | |||
| 1595 | for (Function &F : M) | |||
| 1596 | if (Comdat *C = F.getComdat()) | |||
| 1597 | ComdatMembers.insert(std::make_pair(C, &F)); | |||
| 1598 | for (GlobalVariable &GV : M.globals()) | |||
| 1599 | if (Comdat *C = GV.getComdat()) | |||
| 1600 | ComdatMembers.insert(std::make_pair(C, &GV)); | |||
| 1601 | for (GlobalAlias &GA : M.aliases()) | |||
| 1602 | if (Comdat *C = GA.getComdat()) | |||
| 1603 | ComdatMembers.insert(std::make_pair(C, &GA)); | |||
| 1604 | } | |||
| 1605 | ||||
| 1606 | static bool InstrumentAllFunctions( | |||
| 1607 | Module &M, function_ref<TargetLibraryInfo &(Function &)> LookupTLI, | |||
| 1608 | function_ref<BranchProbabilityInfo *(Function &)> LookupBPI, | |||
| 1609 | function_ref<BlockFrequencyInfo *(Function &)> LookupBFI, bool IsCS) { | |||
| 1610 | // For the context-sensitve instrumentation, we should have a separated pass | |||
| 1611 | // (before LTO/ThinLTO linking) to create these variables. | |||
| 1612 | if (!IsCS) | |||
| 1613 | createIRLevelProfileFlagVar(M, /* IsCS */ false, PGOInstrumentEntry); | |||
| 1614 | std::unordered_multimap<Comdat *, GlobalValue *> ComdatMembers; | |||
| 1615 | collectComdatMembers(M, ComdatMembers); | |||
| 1616 | ||||
| 1617 | for (auto &F : M) { | |||
| 1618 | if (F.isDeclaration()) | |||
| 1619 | continue; | |||
| 1620 | if (F.hasFnAttribute(llvm::Attribute::NoProfile)) | |||
| 1621 | continue; | |||
| 1622 | auto &TLI = LookupTLI(F); | |||
| 1623 | auto *BPI = LookupBPI(F); | |||
| 1624 | auto *BFI = LookupBFI(F); | |||
| 1625 | instrumentOneFunc(F, &M, TLI, BPI, BFI, ComdatMembers, IsCS); | |||
| 1626 | } | |||
| 1627 | return true; | |||
| 1628 | } | |||
| 1629 | ||||
| 1630 | PreservedAnalyses | |||
| 1631 | PGOInstrumentationGenCreateVar::run(Module &M, ModuleAnalysisManager &AM) { | |||
| 1632 | createProfileFileNameVar(M, CSInstrName); | |||
| 1633 | createIRLevelProfileFlagVar(M, /* IsCS */ true, PGOInstrumentEntry); | |||
| 1634 | return PreservedAnalyses::all(); | |||
| 1635 | } | |||
| 1636 | ||||
| 1637 | bool PGOInstrumentationGenLegacyPass::runOnModule(Module &M) { | |||
| 1638 | if (skipModule(M)) | |||
| 1639 | return false; | |||
| 1640 | ||||
| 1641 | auto LookupTLI = [this](Function &F) -> TargetLibraryInfo & { | |||
| 1642 | return this->getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); | |||
| 1643 | }; | |||
| 1644 | auto LookupBPI = [this](Function &F) { | |||
| 1645 | return &this->getAnalysis<BranchProbabilityInfoWrapperPass>(F).getBPI(); | |||
| 1646 | }; | |||
| 1647 | auto LookupBFI = [this](Function &F) { | |||
| 1648 | return &this->getAnalysis<BlockFrequencyInfoWrapperPass>(F).getBFI(); | |||
| 1649 | }; | |||
| 1650 | return InstrumentAllFunctions(M, LookupTLI, LookupBPI, LookupBFI, IsCS); | |||
| 1651 | } | |||
| 1652 | ||||
| 1653 | PreservedAnalyses PGOInstrumentationGen::run(Module &M, | |||
| 1654 | ModuleAnalysisManager &AM) { | |||
| 1655 | auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); | |||
| 1656 | auto LookupTLI = [&FAM](Function &F) -> TargetLibraryInfo & { | |||
| 1657 | return FAM.getResult<TargetLibraryAnalysis>(F); | |||
| 1658 | }; | |||
| 1659 | auto LookupBPI = [&FAM](Function &F) { | |||
| 1660 | return &FAM.getResult<BranchProbabilityAnalysis>(F); | |||
| 1661 | }; | |||
| 1662 | auto LookupBFI = [&FAM](Function &F) { | |||
| 1663 | return &FAM.getResult<BlockFrequencyAnalysis>(F); | |||
| 1664 | }; | |||
| 1665 | ||||
| 1666 | if (!InstrumentAllFunctions(M, LookupTLI, LookupBPI, LookupBFI, IsCS)) | |||
| ||||
| 1667 | return PreservedAnalyses::all(); | |||
| 1668 | ||||
| 1669 | return PreservedAnalyses::none(); | |||
| 1670 | } | |||
| 1671 | ||||
| 1672 | // Using the ratio b/w sums of profile count values and BFI count values to | |||
| 1673 | // adjust the func entry count. | |||
| 1674 | static void fixFuncEntryCount(PGOUseFunc &Func, LoopInfo &LI, | |||
| 1675 | BranchProbabilityInfo &NBPI) { | |||
| 1676 | Function &F = Func.getFunc(); | |||
| 1677 | BlockFrequencyInfo NBFI(F, NBPI, LI); | |||
| 1678 | #ifndef NDEBUG1 | |||
| 1679 | auto BFIEntryCount = F.getEntryCount(); | |||
| 1680 | assert(BFIEntryCount.hasValue() && (BFIEntryCount.getCount() > 0) &&((void)0) | |||
| 1681 | "Invalid BFI Entrycount")((void)0); | |||
| 1682 | #endif | |||
| 1683 | auto SumCount = APFloat::getZero(APFloat::IEEEdouble()); | |||
| 1684 | auto SumBFICount = APFloat::getZero(APFloat::IEEEdouble()); | |||
| 1685 | for (auto &BBI : F) { | |||
| 1686 | uint64_t CountValue = 0; | |||
| 1687 | uint64_t BFICountValue = 0; | |||
| 1688 | if (!Func.findBBInfo(&BBI)) | |||
| 1689 | continue; | |||
| 1690 | auto BFICount = NBFI.getBlockProfileCount(&BBI); | |||
| 1691 | CountValue = Func.getBBInfo(&BBI).CountValue; | |||
| 1692 | BFICountValue = BFICount.getValue(); | |||
| 1693 | SumCount.add(APFloat(CountValue * 1.0), APFloat::rmNearestTiesToEven); | |||
| 1694 | SumBFICount.add(APFloat(BFICountValue * 1.0), APFloat::rmNearestTiesToEven); | |||
| 1695 | } | |||
| 1696 | if (SumCount.isZero()) | |||
| 1697 | return; | |||
| 1698 | ||||
| 1699 | assert(SumBFICount.compare(APFloat(0.0)) == APFloat::cmpGreaterThan &&((void)0) | |||
| 1700 | "Incorrect sum of BFI counts")((void)0); | |||
| 1701 | if (SumBFICount.compare(SumCount) == APFloat::cmpEqual) | |||
| 1702 | return; | |||
| 1703 | double Scale = (SumCount / SumBFICount).convertToDouble(); | |||
| 1704 | if (Scale < 1.001 && Scale > 0.999) | |||
| 1705 | return; | |||
| 1706 | ||||
| 1707 | uint64_t FuncEntryCount = Func.getBBInfo(&*F.begin()).CountValue; | |||
| 1708 | uint64_t NewEntryCount = 0.5 + FuncEntryCount * Scale; | |||
| 1709 | if (NewEntryCount == 0) | |||
| 1710 | NewEntryCount = 1; | |||
| 1711 | if (NewEntryCount != FuncEntryCount) { | |||
| 1712 | F.setEntryCount(ProfileCount(NewEntryCount, Function::PCT_Real)); | |||
| 1713 | LLVM_DEBUG(dbgs() << "FixFuncEntryCount: in " << F.getName()do { } while (false) | |||
| 1714 | << ", entry_count " << FuncEntryCount << " --> "do { } while (false) | |||
| 1715 | << NewEntryCount << "\n")do { } while (false); | |||
| 1716 | } | |||
| 1717 | } | |||
| 1718 | ||||
| 1719 | // Compare the profile count values with BFI count values, and print out | |||
| 1720 | // the non-matching ones. | |||
| 1721 | static void verifyFuncBFI(PGOUseFunc &Func, LoopInfo &LI, | |||
| 1722 | BranchProbabilityInfo &NBPI, | |||
| 1723 | uint64_t HotCountThreshold, | |||
| 1724 | uint64_t ColdCountThreshold) { | |||
| 1725 | Function &F = Func.getFunc(); | |||
| 1726 | BlockFrequencyInfo NBFI(F, NBPI, LI); | |||
| 1727 | // bool PrintFunc = false; | |||
| 1728 | bool HotBBOnly = PGOVerifyHotBFI; | |||
| 1729 | std::string Msg; | |||
| 1730 | OptimizationRemarkEmitter ORE(&F); | |||
| 1731 | ||||
| 1732 | unsigned BBNum = 0, BBMisMatchNum = 0, NonZeroBBNum = 0; | |||
| 1733 | for (auto &BBI : F) { | |||
| 1734 | uint64_t CountValue = 0; | |||
| 1735 | uint64_t BFICountValue = 0; | |||
| 1736 | ||||
| 1737 | if (Func.getBBInfo(&BBI).CountValid) | |||
| 1738 | CountValue = Func.getBBInfo(&BBI).CountValue; | |||
| 1739 | ||||
| 1740 | BBNum++; | |||
| 1741 | if (CountValue) | |||
| 1742 | NonZeroBBNum++; | |||
| 1743 | auto BFICount = NBFI.getBlockProfileCount(&BBI); | |||
| 1744 | if (BFICount) | |||
| 1745 | BFICountValue = BFICount.getValue(); | |||
| 1746 | ||||
| 1747 | if (HotBBOnly) { | |||
| 1748 | bool rawIsHot = CountValue >= HotCountThreshold; | |||
| 1749 | bool BFIIsHot = BFICountValue >= HotCountThreshold; | |||
| 1750 | bool rawIsCold = CountValue <= ColdCountThreshold; | |||
| 1751 | bool ShowCount = false; | |||
| 1752 | if (rawIsHot && !BFIIsHot) { | |||
| 1753 | Msg = "raw-Hot to BFI-nonHot"; | |||
| 1754 | ShowCount = true; | |||
| 1755 | } else if (rawIsCold && BFIIsHot) { | |||
| 1756 | Msg = "raw-Cold to BFI-Hot"; | |||
| 1757 | ShowCount = true; | |||
| 1758 | } | |||
| 1759 | if (!ShowCount) | |||
| 1760 | continue; | |||
| 1761 | } else { | |||
| 1762 | if ((CountValue < PGOVerifyBFICutoff) && | |||
| 1763 | (BFICountValue < PGOVerifyBFICutoff)) | |||
| 1764 | continue; | |||
| 1765 | uint64_t Diff = (BFICountValue >= CountValue) | |||
| 1766 | ? BFICountValue - CountValue | |||
| 1767 | : CountValue - BFICountValue; | |||
| 1768 | if (Diff < CountValue / 100 * PGOVerifyBFIRatio) | |||
| 1769 | continue; | |||
| 1770 | } | |||
| 1771 | BBMisMatchNum++; | |||
| 1772 | ||||
| 1773 | ORE.emit([&]() { | |||
| 1774 | OptimizationRemarkAnalysis Remark(DEBUG_TYPE"pgo-instrumentation", "bfi-verify", | |||
| 1775 | F.getSubprogram(), &BBI); | |||
| 1776 | Remark << "BB " << ore::NV("Block", BBI.getName()) | |||
| 1777 | << " Count=" << ore::NV("Count", CountValue) | |||
| 1778 | << " BFI_Count=" << ore::NV("Count", BFICountValue); | |||
| 1779 | if (!Msg.empty()) | |||
| 1780 | Remark << " (" << Msg << ")"; | |||
| 1781 | return Remark; | |||
| 1782 | }); | |||
| 1783 | } | |||
| 1784 | if (BBMisMatchNum) | |||
| 1785 | ORE.emit([&]() { | |||
| 1786 | return OptimizationRemarkAnalysis(DEBUG_TYPE"pgo-instrumentation", "bfi-verify", | |||
| 1787 | F.getSubprogram(), &F.getEntryBlock()) | |||
| 1788 | << "In Func " << ore::NV("Function", F.getName()) | |||
| 1789 | << ": Num_of_BB=" << ore::NV("Count", BBNum) | |||
| 1790 | << ", Num_of_non_zerovalue_BB=" << ore::NV("Count", NonZeroBBNum) | |||
| 1791 | << ", Num_of_mis_matching_BB=" << ore::NV("Count", BBMisMatchNum); | |||
| 1792 | }); | |||
| 1793 | } | |||
| 1794 | ||||
| 1795 | static bool annotateAllFunctions( | |||
| 1796 | Module &M, StringRef ProfileFileName, StringRef ProfileRemappingFileName, | |||
| 1797 | function_ref<TargetLibraryInfo &(Function &)> LookupTLI, | |||
| 1798 | function_ref<BranchProbabilityInfo *(Function &)> LookupBPI, | |||
| 1799 | function_ref<BlockFrequencyInfo *(Function &)> LookupBFI, | |||
| 1800 | ProfileSummaryInfo *PSI, bool IsCS) { | |||
| 1801 | LLVM_DEBUG(dbgs() << "Read in profile counters: ")do { } while (false); | |||
| 1802 | auto &Ctx = M.getContext(); | |||
| 1803 | // Read the counter array from file. | |||
| 1804 | auto ReaderOrErr = | |||
| 1805 | IndexedInstrProfReader::create(ProfileFileName, ProfileRemappingFileName); | |||
| 1806 | if (Error E = ReaderOrErr.takeError()) { | |||
| 1807 | handleAllErrors(std::move(E), [&](const ErrorInfoBase &EI) { | |||
| 1808 | Ctx.diagnose( | |||
| 1809 | DiagnosticInfoPGOProfile(ProfileFileName.data(), EI.message())); | |||
| 1810 | }); | |||
| 1811 | return false; | |||
| 1812 | } | |||
| 1813 | ||||
| 1814 | std::unique_ptr<IndexedInstrProfReader> PGOReader = | |||
| 1815 | std::move(ReaderOrErr.get()); | |||
| 1816 | if (!PGOReader) { | |||
| 1817 | Ctx.diagnose(DiagnosticInfoPGOProfile(ProfileFileName.data(), | |||
| 1818 | StringRef("Cannot get PGOReader"))); | |||
| 1819 | return false; | |||
| 1820 | } | |||
| 1821 | if (!PGOReader->hasCSIRLevelProfile() && IsCS) | |||
| 1822 | return false; | |||
| 1823 | ||||
| 1824 | // TODO: might need to change the warning once the clang option is finalized. | |||
| 1825 | if (!PGOReader->isIRLevelProfile()) { | |||
| 1826 | Ctx.diagnose(DiagnosticInfoPGOProfile( | |||
| 1827 | ProfileFileName.data(), "Not an IR level instrumentation profile")); | |||
| 1828 | return false; | |||
| 1829 | } | |||
| 1830 | ||||
| 1831 | // Add the profile summary (read from the header of the indexed summary) here | |||
| 1832 | // so that we can use it below when reading counters (which checks if the | |||
| 1833 | // function should be marked with a cold or inlinehint attribute). | |||
| 1834 | M.setProfileSummary(PGOReader->getSummary(IsCS).getMD(M.getContext()), | |||
| 1835 | IsCS ? ProfileSummary::PSK_CSInstr | |||
| 1836 | : ProfileSummary::PSK_Instr); | |||
| 1837 | PSI->refresh(); | |||
| 1838 | ||||
| 1839 | std::unordered_multimap<Comdat *, GlobalValue *> ComdatMembers; | |||
| 1840 | collectComdatMembers(M, ComdatMembers); | |||
| 1841 | std::vector<Function *> HotFunctions; | |||
| 1842 | std::vector<Function *> ColdFunctions; | |||
| 1843 | ||||
| 1844 | // If the profile marked as always instrument the entry BB, do the | |||
| 1845 | // same. Note this can be overwritten by the internal option in CFGMST.h | |||
| 1846 | bool InstrumentFuncEntry = PGOReader->instrEntryBBEnabled(); | |||
| 1847 | if (PGOInstrumentEntry.getNumOccurrences() > 0) | |||
| 1848 | InstrumentFuncEntry = PGOInstrumentEntry; | |||
| 1849 | for (auto &F : M) { | |||
| 1850 | if (F.isDeclaration()) | |||
| 1851 | continue; | |||
| 1852 | auto &TLI = LookupTLI(F); | |||
| 1853 | auto *BPI = LookupBPI(F); | |||
| 1854 | auto *BFI = LookupBFI(F); | |||
| 1855 | // Split indirectbr critical edges here before computing the MST rather than | |||
| 1856 | // later in getInstrBB() to avoid invalidating it. | |||
| 1857 | SplitIndirectBrCriticalEdges(F, BPI, BFI); | |||
| 1858 | PGOUseFunc Func(F, &M, TLI, ComdatMembers, BPI, BFI, PSI, IsCS, | |||
| 1859 | InstrumentFuncEntry); | |||
| 1860 | // When AllMinusOnes is true, it means the profile for the function | |||
| 1861 | // is unrepresentative and this function is actually hot. Set the | |||
| 1862 | // entry count of the function to be multiple times of hot threshold | |||
| 1863 | // and drop all its internal counters. | |||
| 1864 | bool AllMinusOnes = false; | |||
| 1865 | bool AllZeros = false; | |||
| 1866 | if (!Func.readCounters(PGOReader.get(), AllZeros, AllMinusOnes)) | |||
| 1867 | continue; | |||
| 1868 | if (AllZeros) { | |||
| 1869 | F.setEntryCount(ProfileCount(0, Function::PCT_Real)); | |||
| 1870 | if (Func.getProgramMaxCount() != 0) | |||
| 1871 | ColdFunctions.push_back(&F); | |||
| 1872 | continue; | |||
| 1873 | } | |||
| 1874 | const unsigned MultiplyFactor = 3; | |||
| 1875 | if (AllMinusOnes) { | |||
| 1876 | uint64_t HotThreshold = PSI->getHotCountThreshold(); | |||
| 1877 | if (HotThreshold) | |||
| 1878 | F.setEntryCount( | |||
| 1879 | ProfileCount(HotThreshold * MultiplyFactor, Function::PCT_Real)); | |||
| 1880 | HotFunctions.push_back(&F); | |||
| 1881 | continue; | |||
| 1882 | } | |||
| 1883 | Func.populateCounters(); | |||
| 1884 | Func.setBranchWeights(); | |||
| 1885 | Func.annotateValueSites(); | |||
| 1886 | Func.annotateIrrLoopHeaderWeights(); | |||
| 1887 | PGOUseFunc::FuncFreqAttr FreqAttr = Func.getFuncFreqAttr(); | |||
| 1888 | if (FreqAttr == PGOUseFunc::FFA_Cold) | |||
| 1889 | ColdFunctions.push_back(&F); | |||
| 1890 | else if (FreqAttr == PGOUseFunc::FFA_Hot) | |||
| 1891 | HotFunctions.push_back(&F); | |||
| 1892 | if (PGOViewCounts != PGOVCT_None && | |||
| 1893 | (ViewBlockFreqFuncName.empty() || | |||
| 1894 | F.getName().equals(ViewBlockFreqFuncName))) { | |||
| 1895 | LoopInfo LI{DominatorTree(F)}; | |||
| 1896 | std::unique_ptr<BranchProbabilityInfo> NewBPI = | |||
| 1897 | std::make_unique<BranchProbabilityInfo>(F, LI); | |||
| 1898 | std::unique_ptr<BlockFrequencyInfo> NewBFI = | |||
| 1899 | std::make_unique<BlockFrequencyInfo>(F, *NewBPI, LI); | |||
| 1900 | if (PGOViewCounts == PGOVCT_Graph) | |||
| 1901 | NewBFI->view(); | |||
| 1902 | else if (PGOViewCounts == PGOVCT_Text) { | |||
| 1903 | dbgs() << "pgo-view-counts: " << Func.getFunc().getName() << "\n"; | |||
| 1904 | NewBFI->print(dbgs()); | |||
| 1905 | } | |||
| 1906 | } | |||
| 1907 | if (PGOViewRawCounts != PGOVCT_None && | |||
| 1908 | (ViewBlockFreqFuncName.empty() || | |||
| 1909 | F.getName().equals(ViewBlockFreqFuncName))) { | |||
| 1910 | if (PGOViewRawCounts == PGOVCT_Graph) | |||
| 1911 | if (ViewBlockFreqFuncName.empty()) | |||
| 1912 | WriteGraph(&Func, Twine("PGORawCounts_") + Func.getFunc().getName()); | |||
| 1913 | else | |||
| 1914 | ViewGraph(&Func, Twine("PGORawCounts_") + Func.getFunc().getName()); | |||
| 1915 | else if (PGOViewRawCounts == PGOVCT_Text) { | |||
| 1916 | dbgs() << "pgo-view-raw-counts: " << Func.getFunc().getName() << "\n"; | |||
| 1917 | Func.dumpInfo(); | |||
| 1918 | } | |||
| 1919 | } | |||
| 1920 | ||||
| 1921 | if (PGOVerifyBFI || PGOVerifyHotBFI || PGOFixEntryCount) { | |||
| 1922 | LoopInfo LI{DominatorTree(F)}; | |||
| 1923 | BranchProbabilityInfo NBPI(F, LI); | |||
| 1924 | ||||
| 1925 | // Fix func entry count. | |||
| 1926 | if (PGOFixEntryCount) | |||
| 1927 | fixFuncEntryCount(Func, LI, NBPI); | |||
| 1928 | ||||
| 1929 | // Verify BlockFrequency information. | |||
| 1930 | uint64_t HotCountThreshold = 0, ColdCountThreshold = 0; | |||
| 1931 | if (PGOVerifyHotBFI) { | |||
| 1932 | HotCountThreshold = PSI->getOrCompHotCountThreshold(); | |||
| 1933 | ColdCountThreshold = PSI->getOrCompColdCountThreshold(); | |||
| 1934 | } | |||
| 1935 | verifyFuncBFI(Func, LI, NBPI, HotCountThreshold, ColdCountThreshold); | |||
| 1936 | } | |||
| 1937 | } | |||
| 1938 | ||||
| 1939 | // Set function hotness attribute from the profile. | |||
| 1940 | // We have to apply these attributes at the end because their presence | |||
| 1941 | // can affect the BranchProbabilityInfo of any callers, resulting in an | |||
| 1942 | // inconsistent MST between prof-gen and prof-use. | |||
| 1943 | for (auto &F : HotFunctions) { | |||
| 1944 | F->addFnAttr(Attribute::InlineHint); | |||
| 1945 | LLVM_DEBUG(dbgs() << "Set inline attribute to function: " << F->getName()do { } while (false) | |||
| 1946 | << "\n")do { } while (false); | |||
| 1947 | } | |||
| 1948 | for (auto &F : ColdFunctions) { | |||
| 1949 | // Only set when there is no Attribute::Hot set by the user. For Hot | |||
| 1950 | // attribute, user's annotation has the precedence over the profile. | |||
| 1951 | if (F->hasFnAttribute(Attribute::Hot)) { | |||
| 1952 | auto &Ctx = M.getContext(); | |||
| 1953 | std::string Msg = std::string("Function ") + F->getName().str() + | |||
| 1954 | std::string(" is annotated as a hot function but" | |||
| 1955 | " the profile is cold"); | |||
| 1956 | Ctx.diagnose( | |||
| 1957 | DiagnosticInfoPGOProfile(M.getName().data(), Msg, DS_Warning)); | |||
| 1958 | continue; | |||
| 1959 | } | |||
| 1960 | F->addFnAttr(Attribute::Cold); | |||
| 1961 | LLVM_DEBUG(dbgs() << "Set cold attribute to function: " << F->getName()do { } while (false) | |||
| 1962 | << "\n")do { } while (false); | |||
| 1963 | } | |||
| 1964 | return true; | |||
| 1965 | } | |||
| 1966 | ||||
| 1967 | PGOInstrumentationUse::PGOInstrumentationUse(std::string Filename, | |||
| 1968 | std::string RemappingFilename, | |||
| 1969 | bool IsCS) | |||
| 1970 | : ProfileFileName(std::move(Filename)), | |||
| 1971 | ProfileRemappingFileName(std::move(RemappingFilename)), IsCS(IsCS) { | |||
| 1972 | if (!PGOTestProfileFile.empty()) | |||
| 1973 | ProfileFileName = PGOTestProfileFile; | |||
| 1974 | if (!PGOTestProfileRemappingFile.empty()) | |||
| 1975 | ProfileRemappingFileName = PGOTestProfileRemappingFile; | |||
| 1976 | } | |||
| 1977 | ||||
| 1978 | PreservedAnalyses PGOInstrumentationUse::run(Module &M, | |||
| 1979 | ModuleAnalysisManager &AM) { | |||
| 1980 | ||||
| 1981 | auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); | |||
| 1982 | auto LookupTLI = [&FAM](Function &F) -> TargetLibraryInfo & { | |||
| 1983 | return FAM.getResult<TargetLibraryAnalysis>(F); | |||
| 1984 | }; | |||
| 1985 | auto LookupBPI = [&FAM](Function &F) { | |||
| 1986 | return &FAM.getResult<BranchProbabilityAnalysis>(F); | |||
| 1987 | }; | |||
| 1988 | auto LookupBFI = [&FAM](Function &F) { | |||
| 1989 | return &FAM.getResult<BlockFrequencyAnalysis>(F); | |||
| 1990 | }; | |||
| 1991 | ||||
| 1992 | auto *PSI = &AM.getResult<ProfileSummaryAnalysis>(M); | |||
| 1993 | ||||
| 1994 | if (!annotateAllFunctions(M, ProfileFileName, ProfileRemappingFileName, | |||
| 1995 | LookupTLI, LookupBPI, LookupBFI, PSI, IsCS)) | |||
| 1996 | return PreservedAnalyses::all(); | |||
| 1997 | ||||
| 1998 | return PreservedAnalyses::none(); | |||
| 1999 | } | |||
| 2000 | ||||
| 2001 | bool PGOInstrumentationUseLegacyPass::runOnModule(Module &M) { | |||
| 2002 | if (skipModule(M)) | |||
| 2003 | return false; | |||
| 2004 | ||||
| 2005 | auto LookupTLI = [this](Function &F) -> TargetLibraryInfo & { | |||
| 2006 | return this->getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); | |||
| 2007 | }; | |||
| 2008 | auto LookupBPI = [this](Function &F) { | |||
| 2009 | return &this->getAnalysis<BranchProbabilityInfoWrapperPass>(F).getBPI(); | |||
| 2010 | }; | |||
| 2011 | auto LookupBFI = [this](Function &F) { | |||
| 2012 | return &this->getAnalysis<BlockFrequencyInfoWrapperPass>(F).getBFI(); | |||
| 2013 | }; | |||
| 2014 | ||||
| 2015 | auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); | |||
| 2016 | return annotateAllFunctions(M, ProfileFileName, "", LookupTLI, LookupBPI, | |||
| 2017 | LookupBFI, PSI, IsCS); | |||
| 2018 | } | |||
| 2019 | ||||
| 2020 | static std::string getSimpleNodeName(const BasicBlock *Node) { | |||
| 2021 | if (!Node->getName().empty()) | |||
| 2022 | return std::string(Node->getName()); | |||
| 2023 | ||||
| 2024 | std::string SimpleNodeName; | |||
| 2025 | raw_string_ostream OS(SimpleNodeName); | |||
| 2026 | Node->printAsOperand(OS, false); | |||
| 2027 | return OS.str(); | |||
| 2028 | } | |||
| 2029 | ||||
| 2030 | void llvm::setProfMetadata(Module *M, Instruction *TI, | |||
| 2031 | ArrayRef<uint64_t> EdgeCounts, | |||
| 2032 | uint64_t MaxCount) { | |||
| 2033 | MDBuilder MDB(M->getContext()); | |||
| 2034 | assert(MaxCount > 0 && "Bad max count")((void)0); | |||
| 2035 | uint64_t Scale = calculateCountScale(MaxCount); | |||
| 2036 | SmallVector<unsigned, 4> Weights; | |||
| 2037 | for (const auto &ECI : EdgeCounts) | |||
| 2038 | Weights.push_back(scaleBranchCount(ECI, Scale)); | |||
| 2039 | ||||
| 2040 | LLVM_DEBUG(dbgs() << "Weight is: "; for (const auto &Wdo { } while (false) | |||
| 2041 | : Weights) {do { } while (false) | |||
| 2042 | dbgs() << W << " ";do { } while (false) | |||
| 2043 | } dbgs() << "\n";)do { } while (false); | |||
| 2044 | ||||
| 2045 | TI->setMetadata(LLVMContext::MD_prof, MDB.createBranchWeights(Weights)); | |||
| 2046 | if (EmitBranchProbability) { | |||
| 2047 | std::string BrCondStr = getBranchCondString(TI); | |||
| 2048 | if (BrCondStr.empty()) | |||
| 2049 | return; | |||
| 2050 | ||||
| 2051 | uint64_t WSum = | |||
| 2052 | std::accumulate(Weights.begin(), Weights.end(), (uint64_t)0, | |||
| 2053 | [](uint64_t w1, uint64_t w2) { return w1 + w2; }); | |||
| 2054 | uint64_t TotalCount = | |||
| 2055 | std::accumulate(EdgeCounts.begin(), EdgeCounts.end(), (uint64_t)0, | |||
| 2056 | [](uint64_t c1, uint64_t c2) { return c1 + c2; }); | |||
| 2057 | Scale = calculateCountScale(WSum); | |||
| 2058 | BranchProbability BP(scaleBranchCount(Weights[0], Scale), | |||
| 2059 | scaleBranchCount(WSum, Scale)); | |||
| 2060 | std::string BranchProbStr; | |||
| 2061 | raw_string_ostream OS(BranchProbStr); | |||
| 2062 | OS << BP; | |||
| 2063 | OS << " (total count : " << TotalCount << ")"; | |||
| 2064 | OS.flush(); | |||
| 2065 | Function *F = TI->getParent()->getParent(); | |||
| 2066 | OptimizationRemarkEmitter ORE(F); | |||
| 2067 | ORE.emit([&]() { | |||
| 2068 | return OptimizationRemark(DEBUG_TYPE"pgo-instrumentation", "pgo-instrumentation", TI) | |||
| 2069 | << BrCondStr << " is true with probability : " << BranchProbStr; | |||
| 2070 | }); | |||
| 2071 | } | |||
| 2072 | } | |||
| 2073 | ||||
| 2074 | namespace llvm { | |||
| 2075 | ||||
| 2076 | void setIrrLoopHeaderMetadata(Module *M, Instruction *TI, uint64_t Count) { | |||
| 2077 | MDBuilder MDB(M->getContext()); | |||
| 2078 | TI->setMetadata(llvm::LLVMContext::MD_irr_loop, | |||
| 2079 | MDB.createIrrLoopHeaderWeight(Count)); | |||
| 2080 | } | |||
| 2081 | ||||
| 2082 | template <> struct GraphTraits<PGOUseFunc *> { | |||
| 2083 | using NodeRef = const BasicBlock *; | |||
| 2084 | using ChildIteratorType = const_succ_iterator; | |||
| 2085 | using nodes_iterator = pointer_iterator<Function::const_iterator>; | |||
| 2086 | ||||
| 2087 | static NodeRef getEntryNode(const PGOUseFunc *G) { | |||
| 2088 | return &G->getFunc().front(); | |||
| 2089 | } | |||
| 2090 | ||||
| 2091 | static ChildIteratorType child_begin(const NodeRef N) { | |||
| 2092 | return succ_begin(N); | |||
| 2093 | } | |||
| 2094 | ||||
| 2095 | static ChildIteratorType child_end(const NodeRef N) { return succ_end(N); } | |||
| 2096 | ||||
| 2097 | static nodes_iterator nodes_begin(const PGOUseFunc *G) { | |||
| 2098 | return nodes_iterator(G->getFunc().begin()); | |||
| 2099 | } | |||
| 2100 | ||||
| 2101 | static nodes_iterator nodes_end(const PGOUseFunc *G) { | |||
| 2102 | return nodes_iterator(G->getFunc().end()); | |||
| 2103 | } | |||
| 2104 | }; | |||
| 2105 | ||||
| 2106 | template <> struct DOTGraphTraits<PGOUseFunc *> : DefaultDOTGraphTraits { | |||
| 2107 | explicit DOTGraphTraits(bool isSimple = false) | |||
| 2108 | : DefaultDOTGraphTraits(isSimple) {} | |||
| 2109 | ||||
| 2110 | static std::string getGraphName(const PGOUseFunc *G) { | |||
| 2111 | return std::string(G->getFunc().getName()); | |||
| 2112 | } | |||
| 2113 | ||||
| 2114 | std::string getNodeLabel(const BasicBlock *Node, const PGOUseFunc *Graph) { | |||
| 2115 | std::string Result; | |||
| 2116 | raw_string_ostream OS(Result); | |||
| 2117 | ||||
| 2118 | OS << getSimpleNodeName(Node) << ":\\l"; | |||
| 2119 | UseBBInfo *BI = Graph->findBBInfo(Node); | |||
| 2120 | OS << "Count : "; | |||
| 2121 | if (BI && BI->CountValid) | |||
| 2122 | OS << BI->CountValue << "\\l"; | |||
| 2123 | else | |||
| 2124 | OS << "Unknown\\l"; | |||
| 2125 | ||||
| 2126 | if (!PGOInstrSelect) | |||
| 2127 | return Result; | |||
| 2128 | ||||
| 2129 | for (const Instruction &I : *Node) { | |||
| 2130 | if (!isa<SelectInst>(&I)) | |||
| 2131 | continue; | |||
| 2132 | // Display scaled counts for SELECT instruction: | |||
| 2133 | OS << "SELECT : { T = "; | |||
| 2134 | uint64_t TC, FC; | |||
| 2135 | bool HasProf = I.extractProfMetadata(TC, FC); | |||
| 2136 | if (!HasProf) | |||
| 2137 | OS << "Unknown, F = Unknown }\\l"; | |||
| 2138 | else | |||
| 2139 | OS << TC << ", F = " << FC << " }\\l"; | |||
| 2140 | } | |||
| 2141 | return Result; | |||
| 2142 | } | |||
| 2143 | }; | |||
| 2144 | ||||
| 2145 | } // end namespace llvm |