| File: | llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp |
| Warning: | line 892, 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 "llvm/Transforms/Utils/ModuleUtils.h" | |||
| 114 | #include <algorithm> | |||
| 115 | #include <cassert> | |||
| 116 | #include <cstdint> | |||
| 117 | #include <memory> | |||
| 118 | #include <numeric> | |||
| 119 | #include <string> | |||
| 120 | #include <unordered_map> | |||
| 121 | #include <utility> | |||
| 122 | #include <vector> | |||
| 123 | ||||
| 124 | using namespace llvm; | |||
| 125 | using ProfileCount = Function::ProfileCount; | |||
| 126 | using VPCandidateInfo = ValueProfileCollector::CandidateInfo; | |||
| 127 | ||||
| 128 | #define DEBUG_TYPE"pgo-instrumentation" "pgo-instrumentation" | |||
| 129 | ||||
| 130 | STATISTIC(NumOfPGOInstrument, "Number of edges instrumented.")static llvm::Statistic NumOfPGOInstrument = {"pgo-instrumentation" , "NumOfPGOInstrument", "Number of edges instrumented."}; | |||
| 131 | STATISTIC(NumOfPGOSelectInsts, "Number of select instruction instrumented.")static llvm::Statistic NumOfPGOSelectInsts = {"pgo-instrumentation" , "NumOfPGOSelectInsts", "Number of select instruction instrumented." }; | |||
| 132 | STATISTIC(NumOfPGOMemIntrinsics, "Number of mem intrinsics instrumented.")static llvm::Statistic NumOfPGOMemIntrinsics = {"pgo-instrumentation" , "NumOfPGOMemIntrinsics", "Number of mem intrinsics instrumented." }; | |||
| 133 | STATISTIC(NumOfPGOEdge, "Number of edges.")static llvm::Statistic NumOfPGOEdge = {"pgo-instrumentation", "NumOfPGOEdge", "Number of edges."}; | |||
| 134 | STATISTIC(NumOfPGOBB, "Number of basic-blocks.")static llvm::Statistic NumOfPGOBB = {"pgo-instrumentation", "NumOfPGOBB" , "Number of basic-blocks."}; | |||
| 135 | STATISTIC(NumOfPGOSplit, "Number of critical edge splits.")static llvm::Statistic NumOfPGOSplit = {"pgo-instrumentation" , "NumOfPGOSplit", "Number of critical edge splits."}; | |||
| 136 | STATISTIC(NumOfPGOFunc, "Number of functions having valid profile counts.")static llvm::Statistic NumOfPGOFunc = {"pgo-instrumentation", "NumOfPGOFunc", "Number of functions having valid profile counts." }; | |||
| 137 | STATISTIC(NumOfPGOMismatch, "Number of functions having mismatch profile.")static llvm::Statistic NumOfPGOMismatch = {"pgo-instrumentation" , "NumOfPGOMismatch", "Number of functions having mismatch profile." }; | |||
| 138 | STATISTIC(NumOfPGOMissing, "Number of functions without profile.")static llvm::Statistic NumOfPGOMissing = {"pgo-instrumentation" , "NumOfPGOMissing", "Number of functions without profile."}; | |||
| 139 | STATISTIC(NumOfPGOICall, "Number of indirect call value instrumentations.")static llvm::Statistic NumOfPGOICall = {"pgo-instrumentation" , "NumOfPGOICall", "Number of indirect call value instrumentations." }; | |||
| 140 | STATISTIC(NumOfCSPGOInstrument, "Number of edges instrumented in CSPGO.")static llvm::Statistic NumOfCSPGOInstrument = {"pgo-instrumentation" , "NumOfCSPGOInstrument", "Number of edges instrumented in CSPGO." }; | |||
| 141 | STATISTIC(NumOfCSPGOSelectInsts,static llvm::Statistic NumOfCSPGOSelectInsts = {"pgo-instrumentation" , "NumOfCSPGOSelectInsts", "Number of select instruction instrumented in CSPGO." } | |||
| 142 | "Number of select instruction instrumented in CSPGO.")static llvm::Statistic NumOfCSPGOSelectInsts = {"pgo-instrumentation" , "NumOfCSPGOSelectInsts", "Number of select instruction instrumented in CSPGO." }; | |||
| 143 | STATISTIC(NumOfCSPGOMemIntrinsics,static llvm::Statistic NumOfCSPGOMemIntrinsics = {"pgo-instrumentation" , "NumOfCSPGOMemIntrinsics", "Number of mem intrinsics instrumented in CSPGO." } | |||
| 144 | "Number of mem intrinsics instrumented in CSPGO.")static llvm::Statistic NumOfCSPGOMemIntrinsics = {"pgo-instrumentation" , "NumOfCSPGOMemIntrinsics", "Number of mem intrinsics instrumented in CSPGO." }; | |||
| 145 | STATISTIC(NumOfCSPGOEdge, "Number of edges in CSPGO.")static llvm::Statistic NumOfCSPGOEdge = {"pgo-instrumentation" , "NumOfCSPGOEdge", "Number of edges in CSPGO."}; | |||
| 146 | STATISTIC(NumOfCSPGOBB, "Number of basic-blocks in CSPGO.")static llvm::Statistic NumOfCSPGOBB = {"pgo-instrumentation", "NumOfCSPGOBB", "Number of basic-blocks in CSPGO."}; | |||
| 147 | STATISTIC(NumOfCSPGOSplit, "Number of critical edge splits in CSPGO.")static llvm::Statistic NumOfCSPGOSplit = {"pgo-instrumentation" , "NumOfCSPGOSplit", "Number of critical edge splits in CSPGO." }; | |||
| 148 | STATISTIC(NumOfCSPGOFunc,static llvm::Statistic NumOfCSPGOFunc = {"pgo-instrumentation" , "NumOfCSPGOFunc", "Number of functions having valid profile counts in CSPGO." } | |||
| 149 | "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." }; | |||
| 150 | STATISTIC(NumOfCSPGOMismatch,static llvm::Statistic NumOfCSPGOMismatch = {"pgo-instrumentation" , "NumOfCSPGOMismatch", "Number of functions having mismatch profile in CSPGO." } | |||
| 151 | "Number of functions having mismatch profile in CSPGO.")static llvm::Statistic NumOfCSPGOMismatch = {"pgo-instrumentation" , "NumOfCSPGOMismatch", "Number of functions having mismatch profile in CSPGO." }; | |||
| 152 | STATISTIC(NumOfCSPGOMissing, "Number of functions without profile in CSPGO.")static llvm::Statistic NumOfCSPGOMissing = {"pgo-instrumentation" , "NumOfCSPGOMissing", "Number of functions without profile in CSPGO." }; | |||
| 153 | ||||
| 154 | // Command line option to specify the file to read profile from. This is | |||
| 155 | // mainly used for testing. | |||
| 156 | static cl::opt<std::string> | |||
| 157 | PGOTestProfileFile("pgo-test-profile-file", cl::init(""), cl::Hidden, | |||
| 158 | cl::value_desc("filename"), | |||
| 159 | cl::desc("Specify the path of profile data file. This is" | |||
| 160 | "mainly for test purpose.")); | |||
| 161 | static cl::opt<std::string> PGOTestProfileRemappingFile( | |||
| 162 | "pgo-test-profile-remapping-file", cl::init(""), cl::Hidden, | |||
| 163 | cl::value_desc("filename"), | |||
| 164 | cl::desc("Specify the path of profile remapping file. This is mainly for " | |||
| 165 | "test purpose.")); | |||
| 166 | ||||
| 167 | // Command line option to disable value profiling. The default is false: | |||
| 168 | // i.e. value profiling is enabled by default. This is for debug purpose. | |||
| 169 | static cl::opt<bool> DisableValueProfiling("disable-vp", cl::init(false), | |||
| 170 | cl::Hidden, | |||
| 171 | cl::desc("Disable Value Profiling")); | |||
| 172 | ||||
| 173 | // Command line option to set the maximum number of VP annotations to write to | |||
| 174 | // the metadata for a single indirect call callsite. | |||
| 175 | static cl::opt<unsigned> MaxNumAnnotations( | |||
| 176 | "icp-max-annotations", cl::init(3), cl::Hidden, cl::ZeroOrMore, | |||
| 177 | cl::desc("Max number of annotations for a single indirect " | |||
| 178 | "call callsite")); | |||
| 179 | ||||
| 180 | // Command line option to set the maximum number of value annotations | |||
| 181 | // to write to the metadata for a single memop intrinsic. | |||
| 182 | static cl::opt<unsigned> MaxNumMemOPAnnotations( | |||
| 183 | "memop-max-annotations", cl::init(4), cl::Hidden, cl::ZeroOrMore, | |||
| 184 | cl::desc("Max number of preicise value annotations for a single memop" | |||
| 185 | "intrinsic")); | |||
| 186 | ||||
| 187 | // Command line option to control appending FunctionHash to the name of a COMDAT | |||
| 188 | // function. This is to avoid the hash mismatch caused by the preinliner. | |||
| 189 | static cl::opt<bool> DoComdatRenaming( | |||
| 190 | "do-comdat-renaming", cl::init(false), cl::Hidden, | |||
| 191 | cl::desc("Append function hash to the name of COMDAT function to avoid " | |||
| 192 | "function hash mismatch due to the preinliner")); | |||
| 193 | ||||
| 194 | // Command line option to enable/disable the warning about missing profile | |||
| 195 | // information. | |||
| 196 | static cl::opt<bool> | |||
| 197 | PGOWarnMissing("pgo-warn-missing-function", cl::init(false), cl::Hidden, | |||
| 198 | cl::desc("Use this option to turn on/off " | |||
| 199 | "warnings about missing profile data for " | |||
| 200 | "functions.")); | |||
| 201 | ||||
| 202 | namespace llvm { | |||
| 203 | // Command line option to enable/disable the warning about a hash mismatch in | |||
| 204 | // the profile data. | |||
| 205 | cl::opt<bool> | |||
| 206 | NoPGOWarnMismatch("no-pgo-warn-mismatch", cl::init(false), cl::Hidden, | |||
| 207 | cl::desc("Use this option to turn off/on " | |||
| 208 | "warnings about profile cfg mismatch.")); | |||
| 209 | } // namespace llvm | |||
| 210 | ||||
| 211 | // Command line option to enable/disable the warning about a hash mismatch in | |||
| 212 | // the profile data for Comdat functions, which often turns out to be false | |||
| 213 | // positive due to the pre-instrumentation inline. | |||
| 214 | static cl::opt<bool> | |||
| 215 | NoPGOWarnMismatchComdat("no-pgo-warn-mismatch-comdat", cl::init(true), | |||
| 216 | cl::Hidden, | |||
| 217 | cl::desc("The option is used to turn on/off " | |||
| 218 | "warnings about hash mismatch for comdat " | |||
| 219 | "functions.")); | |||
| 220 | ||||
| 221 | // Command line option to enable/disable select instruction instrumentation. | |||
| 222 | static cl::opt<bool> | |||
| 223 | PGOInstrSelect("pgo-instr-select", cl::init(true), cl::Hidden, | |||
| 224 | cl::desc("Use this option to turn on/off SELECT " | |||
| 225 | "instruction instrumentation. ")); | |||
| 226 | ||||
| 227 | // Command line option to turn on CFG dot or text dump of raw profile counts | |||
| 228 | static cl::opt<PGOViewCountsType> PGOViewRawCounts( | |||
| 229 | "pgo-view-raw-counts", cl::Hidden, | |||
| 230 | cl::desc("A boolean option to show CFG dag or text " | |||
| 231 | "with raw profile counts from " | |||
| 232 | "profile data. See also option " | |||
| 233 | "-pgo-view-counts. To limit graph " | |||
| 234 | "display to only one function, use " | |||
| 235 | "filtering option -view-bfi-func-name."), | |||
| 236 | cl::values(clEnumValN(PGOVCT_None, "none", "do not show.")llvm::cl::OptionEnumValue { "none", int(PGOVCT_None), "do not show." }, | |||
| 237 | clEnumValN(PGOVCT_Graph, "graph", "show a graph.")llvm::cl::OptionEnumValue { "graph", int(PGOVCT_Graph), "show a graph." }, | |||
| 238 | clEnumValN(PGOVCT_Text, "text", "show in text.")llvm::cl::OptionEnumValue { "text", int(PGOVCT_Text), "show in text." })); | |||
| 239 | ||||
| 240 | // Command line option to enable/disable memop intrinsic call.size profiling. | |||
| 241 | static cl::opt<bool> | |||
| 242 | PGOInstrMemOP("pgo-instr-memop", cl::init(true), cl::Hidden, | |||
| 243 | cl::desc("Use this option to turn on/off " | |||
| 244 | "memory intrinsic size profiling.")); | |||
| 245 | ||||
| 246 | // Emit branch probability as optimization remarks. | |||
| 247 | static cl::opt<bool> | |||
| 248 | EmitBranchProbability("pgo-emit-branch-prob", cl::init(false), cl::Hidden, | |||
| 249 | cl::desc("When this option is on, the annotated " | |||
| 250 | "branch probability will be emitted as " | |||
| 251 | "optimization remarks: -{Rpass|" | |||
| 252 | "pass-remarks}=pgo-instrumentation")); | |||
| 253 | ||||
| 254 | static cl::opt<bool> PGOInstrumentEntry( | |||
| 255 | "pgo-instrument-entry", cl::init(false), cl::Hidden, | |||
| 256 | cl::desc("Force to instrument function entry basicblock.")); | |||
| 257 | ||||
| 258 | static cl::opt<bool> | |||
| 259 | PGOFixEntryCount("pgo-fix-entry-count", cl::init(true), cl::Hidden, | |||
| 260 | cl::desc("Fix function entry count in profile use.")); | |||
| 261 | ||||
| 262 | static cl::opt<bool> PGOVerifyHotBFI( | |||
| 263 | "pgo-verify-hot-bfi", cl::init(false), cl::Hidden, | |||
| 264 | cl::desc("Print out the non-match BFI count if a hot raw profile count " | |||
| 265 | "becomes non-hot, or a cold raw profile count becomes hot. " | |||
| 266 | "The print is enabled under -Rpass-analysis=pgo, or " | |||
| 267 | "internal option -pass-remakrs-analysis=pgo.")); | |||
| 268 | ||||
| 269 | static cl::opt<bool> PGOVerifyBFI( | |||
| 270 | "pgo-verify-bfi", cl::init(false), cl::Hidden, | |||
| 271 | cl::desc("Print out mismatched BFI counts after setting profile metadata " | |||
| 272 | "The print is enabled under -Rpass-analysis=pgo, or " | |||
| 273 | "internal option -pass-remakrs-analysis=pgo.")); | |||
| 274 | ||||
| 275 | static cl::opt<unsigned> PGOVerifyBFIRatio( | |||
| 276 | "pgo-verify-bfi-ratio", cl::init(2), cl::Hidden, | |||
| 277 | cl::desc("Set the threshold for pgo-verify-bfi: only print out " | |||
| 278 | "mismatched BFI if the difference percentage is greater than " | |||
| 279 | "this value (in percentage).")); | |||
| 280 | ||||
| 281 | static cl::opt<unsigned> PGOVerifyBFICutoff( | |||
| 282 | "pgo-verify-bfi-cutoff", cl::init(5), cl::Hidden, | |||
| 283 | cl::desc("Set the threshold for pgo-verify-bfi: skip the counts whose " | |||
| 284 | "profile count value is below.")); | |||
| 285 | ||||
| 286 | namespace llvm { | |||
| 287 | // Command line option to turn on CFG dot dump after profile annotation. | |||
| 288 | // Defined in Analysis/BlockFrequencyInfo.cpp: -pgo-view-counts | |||
| 289 | extern cl::opt<PGOViewCountsType> PGOViewCounts; | |||
| 290 | ||||
| 291 | // Command line option to specify the name of the function for CFG dump | |||
| 292 | // Defined in Analysis/BlockFrequencyInfo.cpp: -view-bfi-func-name= | |||
| 293 | extern cl::opt<std::string> ViewBlockFreqFuncName; | |||
| 294 | ||||
| 295 | extern cl::opt<bool> DebugInfoCorrelate; | |||
| 296 | } // namespace llvm | |||
| 297 | ||||
| 298 | static cl::opt<bool> | |||
| 299 | PGOOldCFGHashing("pgo-instr-old-cfg-hashing", cl::init(false), cl::Hidden, | |||
| 300 | cl::desc("Use the old CFG function hashing")); | |||
| 301 | ||||
| 302 | // Return a string describing the branch condition that can be | |||
| 303 | // used in static branch probability heuristics: | |||
| 304 | static std::string getBranchCondString(Instruction *TI) { | |||
| 305 | BranchInst *BI = dyn_cast<BranchInst>(TI); | |||
| 306 | if (!BI || !BI->isConditional()) | |||
| 307 | return std::string(); | |||
| 308 | ||||
| 309 | Value *Cond = BI->getCondition(); | |||
| 310 | ICmpInst *CI = dyn_cast<ICmpInst>(Cond); | |||
| 311 | if (!CI) | |||
| 312 | return std::string(); | |||
| 313 | ||||
| 314 | std::string result; | |||
| 315 | raw_string_ostream OS(result); | |||
| 316 | OS << CmpInst::getPredicateName(CI->getPredicate()) << "_"; | |||
| 317 | CI->getOperand(0)->getType()->print(OS, true); | |||
| 318 | ||||
| 319 | Value *RHS = CI->getOperand(1); | |||
| 320 | ConstantInt *CV = dyn_cast<ConstantInt>(RHS); | |||
| 321 | if (CV) { | |||
| 322 | if (CV->isZero()) | |||
| 323 | OS << "_Zero"; | |||
| 324 | else if (CV->isOne()) | |||
| 325 | OS << "_One"; | |||
| 326 | else if (CV->isMinusOne()) | |||
| 327 | OS << "_MinusOne"; | |||
| 328 | else | |||
| 329 | OS << "_Const"; | |||
| 330 | } | |||
| 331 | OS.flush(); | |||
| 332 | return result; | |||
| 333 | } | |||
| 334 | ||||
| 335 | static const char *ValueProfKindDescr[] = { | |||
| 336 | #define VALUE_PROF_KIND(Enumerator, Value, Descr) Descr, | |||
| 337 | #include "llvm/ProfileData/InstrProfData.inc" | |||
| 338 | }; | |||
| 339 | ||||
| 340 | namespace { | |||
| 341 | ||||
| 342 | /// The select instruction visitor plays three roles specified | |||
| 343 | /// by the mode. In \c VM_counting mode, it simply counts the number of | |||
| 344 | /// select instructions. In \c VM_instrument mode, it inserts code to count | |||
| 345 | /// the number times TrueValue of select is taken. In \c VM_annotate mode, | |||
| 346 | /// it reads the profile data and annotate the select instruction with metadata. | |||
| 347 | enum VisitMode { VM_counting, VM_instrument, VM_annotate }; | |||
| 348 | class PGOUseFunc; | |||
| 349 | ||||
| 350 | /// Instruction Visitor class to visit select instructions. | |||
| 351 | struct SelectInstVisitor : public InstVisitor<SelectInstVisitor> { | |||
| 352 | Function &F; | |||
| 353 | unsigned NSIs = 0; // Number of select instructions instrumented. | |||
| 354 | VisitMode Mode = VM_counting; // Visiting mode. | |||
| 355 | unsigned *CurCtrIdx = nullptr; // Pointer to current counter index. | |||
| 356 | unsigned TotalNumCtrs = 0; // Total number of counters | |||
| 357 | GlobalVariable *FuncNameVar = nullptr; | |||
| 358 | uint64_t FuncHash = 0; | |||
| 359 | PGOUseFunc *UseFunc = nullptr; | |||
| 360 | ||||
| 361 | SelectInstVisitor(Function &Func) : F(Func) {} | |||
| 362 | ||||
| 363 | void countSelects(Function &Func) { | |||
| 364 | NSIs = 0; | |||
| 365 | Mode = VM_counting; | |||
| 366 | visit(Func); | |||
| 367 | } | |||
| 368 | ||||
| 369 | // Visit the IR stream and instrument all select instructions. \p | |||
| 370 | // Ind is a pointer to the counter index variable; \p TotalNC | |||
| 371 | // is the total number of counters; \p FNV is the pointer to the | |||
| 372 | // PGO function name var; \p FHash is the function hash. | |||
| 373 | void instrumentSelects(Function &Func, unsigned *Ind, unsigned TotalNC, | |||
| 374 | GlobalVariable *FNV, uint64_t FHash) { | |||
| 375 | Mode = VM_instrument; | |||
| 376 | CurCtrIdx = Ind; | |||
| 377 | TotalNumCtrs = TotalNC; | |||
| 378 | FuncHash = FHash; | |||
| 379 | FuncNameVar = FNV; | |||
| 380 | visit(Func); | |||
| 381 | } | |||
| 382 | ||||
| 383 | // Visit the IR stream and annotate all select instructions. | |||
| 384 | void annotateSelects(Function &Func, PGOUseFunc *UF, unsigned *Ind) { | |||
| 385 | Mode = VM_annotate; | |||
| 386 | UseFunc = UF; | |||
| 387 | CurCtrIdx = Ind; | |||
| 388 | visit(Func); | |||
| 389 | } | |||
| 390 | ||||
| 391 | void instrumentOneSelectInst(SelectInst &SI); | |||
| 392 | void annotateOneSelectInst(SelectInst &SI); | |||
| 393 | ||||
| 394 | // Visit \p SI instruction and perform tasks according to visit mode. | |||
| 395 | void visitSelectInst(SelectInst &SI); | |||
| 396 | ||||
| 397 | // Return the number of select instructions. This needs be called after | |||
| 398 | // countSelects(). | |||
| 399 | unsigned getNumOfSelectInsts() const { return NSIs; } | |||
| 400 | }; | |||
| 401 | ||||
| 402 | ||||
| 403 | class PGOInstrumentationGenLegacyPass : public ModulePass { | |||
| 404 | public: | |||
| 405 | static char ID; | |||
| 406 | ||||
| 407 | PGOInstrumentationGenLegacyPass(bool IsCS = false) | |||
| 408 | : ModulePass(ID), IsCS(IsCS) { | |||
| 409 | initializePGOInstrumentationGenLegacyPassPass( | |||
| 410 | *PassRegistry::getPassRegistry()); | |||
| 411 | } | |||
| 412 | ||||
| 413 | StringRef getPassName() const override { return "PGOInstrumentationGenPass"; } | |||
| 414 | ||||
| 415 | private: | |||
| 416 | // Is this is context-sensitive instrumentation. | |||
| 417 | bool IsCS; | |||
| 418 | bool runOnModule(Module &M) override; | |||
| 419 | ||||
| 420 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
| 421 | AU.addRequired<BlockFrequencyInfoWrapperPass>(); | |||
| 422 | AU.addRequired<TargetLibraryInfoWrapperPass>(); | |||
| 423 | } | |||
| 424 | }; | |||
| 425 | ||||
| 426 | class PGOInstrumentationUseLegacyPass : public ModulePass { | |||
| 427 | public: | |||
| 428 | static char ID; | |||
| 429 | ||||
| 430 | // Provide the profile filename as the parameter. | |||
| 431 | PGOInstrumentationUseLegacyPass(std::string Filename = "", bool IsCS = false) | |||
| 432 | : ModulePass(ID), ProfileFileName(std::move(Filename)), IsCS(IsCS) { | |||
| 433 | if (!PGOTestProfileFile.empty()) | |||
| 434 | ProfileFileName = PGOTestProfileFile; | |||
| 435 | initializePGOInstrumentationUseLegacyPassPass( | |||
| 436 | *PassRegistry::getPassRegistry()); | |||
| 437 | } | |||
| 438 | ||||
| 439 | StringRef getPassName() const override { return "PGOInstrumentationUsePass"; } | |||
| 440 | ||||
| 441 | private: | |||
| 442 | std::string ProfileFileName; | |||
| 443 | // Is this is context-sensitive instrumentation use. | |||
| 444 | bool IsCS; | |||
| 445 | ||||
| 446 | bool runOnModule(Module &M) override; | |||
| 447 | ||||
| 448 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
| 449 | AU.addRequired<ProfileSummaryInfoWrapperPass>(); | |||
| 450 | AU.addRequired<BlockFrequencyInfoWrapperPass>(); | |||
| 451 | AU.addRequired<TargetLibraryInfoWrapperPass>(); | |||
| 452 | } | |||
| 453 | }; | |||
| 454 | ||||
| 455 | class PGOInstrumentationGenCreateVarLegacyPass : public ModulePass { | |||
| 456 | public: | |||
| 457 | static char ID; | |||
| 458 | StringRef getPassName() const override { | |||
| 459 | return "PGOInstrumentationGenCreateVarPass"; | |||
| 460 | } | |||
| 461 | PGOInstrumentationGenCreateVarLegacyPass(std::string CSInstrName = "") | |||
| 462 | : ModulePass(ID), InstrProfileOutput(CSInstrName) { | |||
| 463 | initializePGOInstrumentationGenCreateVarLegacyPassPass( | |||
| 464 | *PassRegistry::getPassRegistry()); | |||
| 465 | } | |||
| 466 | ||||
| 467 | private: | |||
| 468 | bool runOnModule(Module &M) override { | |||
| 469 | createProfileFileNameVar(M, InstrProfileOutput); | |||
| 470 | // The variable in a comdat may be discarded by LTO. Ensure the | |||
| 471 | // declaration will be retained. | |||
| 472 | appendToCompilerUsed(M, createIRLevelProfileFlagVar(M, /*IsCS=*/true, | |||
| 473 | PGOInstrumentEntry, | |||
| 474 | DebugInfoCorrelate)); | |||
| 475 | return false; | |||
| 476 | } | |||
| 477 | std::string InstrProfileOutput; | |||
| 478 | }; | |||
| 479 | ||||
| 480 | } // end anonymous namespace | |||
| 481 | ||||
| 482 | char PGOInstrumentationGenLegacyPass::ID = 0; | |||
| 483 | ||||
| 484 | INITIALIZE_PASS_BEGIN(PGOInstrumentationGenLegacyPass, "pgo-instr-gen",static void *initializePGOInstrumentationGenLegacyPassPassOnce (PassRegistry &Registry) { | |||
| 485 | "PGO instrumentation.", false, false)static void *initializePGOInstrumentationGenLegacyPassPassOnce (PassRegistry &Registry) { | |||
| 486 | INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)initializeBlockFrequencyInfoWrapperPassPass(Registry); | |||
| 487 | INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)initializeBranchProbabilityInfoWrapperPassPass(Registry); | |||
| 488 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); | |||
| 489 | 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)); } | |||
| 490 | "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)); } | |||
| 491 | ||||
| 492 | ModulePass *llvm::createPGOInstrumentationGenLegacyPass(bool IsCS) { | |||
| 493 | return new PGOInstrumentationGenLegacyPass(IsCS); | |||
| 494 | } | |||
| 495 | ||||
| 496 | char PGOInstrumentationUseLegacyPass::ID = 0; | |||
| 497 | ||||
| 498 | INITIALIZE_PASS_BEGIN(PGOInstrumentationUseLegacyPass, "pgo-instr-use",static void *initializePGOInstrumentationUseLegacyPassPassOnce (PassRegistry &Registry) { | |||
| 499 | "Read PGO instrumentation profile.", false, false)static void *initializePGOInstrumentationUseLegacyPassPassOnce (PassRegistry &Registry) { | |||
| 500 | INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)initializeBlockFrequencyInfoWrapperPassPass(Registry); | |||
| 501 | INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)initializeBranchProbabilityInfoWrapperPassPass(Registry); | |||
| 502 | INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)initializeProfileSummaryInfoWrapperPassPass(Registry); | |||
| 503 | 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)); } | |||
| 504 | "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)); } | |||
| 505 | ||||
| 506 | ModulePass *llvm::createPGOInstrumentationUseLegacyPass(StringRef Filename, | |||
| 507 | bool IsCS) { | |||
| 508 | return new PGOInstrumentationUseLegacyPass(Filename.str(), IsCS); | |||
| 509 | } | |||
| 510 | ||||
| 511 | char PGOInstrumentationGenCreateVarLegacyPass::ID = 0; | |||
| 512 | ||||
| 513 | 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)); } | |||
| 514 | "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)); } | |||
| 515 | "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)); } | |||
| 516 | 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)); } | |||
| 517 | ||||
| 518 | ModulePass * | |||
| 519 | llvm::createPGOInstrumentationGenCreateVarLegacyPass(StringRef CSInstrName) { | |||
| 520 | return new PGOInstrumentationGenCreateVarLegacyPass(std::string(CSInstrName)); | |||
| 521 | } | |||
| 522 | ||||
| 523 | namespace { | |||
| 524 | ||||
| 525 | /// An MST based instrumentation for PGO | |||
| 526 | /// | |||
| 527 | /// Implements a Minimum Spanning Tree (MST) based instrumentation for PGO | |||
| 528 | /// in the function level. | |||
| 529 | struct PGOEdge { | |||
| 530 | // This class implements the CFG edges. Note the CFG can be a multi-graph. | |||
| 531 | // So there might be multiple edges with same SrcBB and DestBB. | |||
| 532 | const BasicBlock *SrcBB; | |||
| 533 | const BasicBlock *DestBB; | |||
| 534 | uint64_t Weight; | |||
| 535 | bool InMST = false; | |||
| 536 | bool Removed = false; | |||
| 537 | bool IsCritical = false; | |||
| 538 | ||||
| 539 | PGOEdge(const BasicBlock *Src, const BasicBlock *Dest, uint64_t W = 1) | |||
| 540 | : SrcBB(Src), DestBB(Dest), Weight(W) {} | |||
| 541 | ||||
| 542 | // Return the information string of an edge. | |||
| 543 | std::string infoString() const { | |||
| 544 | return (Twine(Removed ? "-" : " ") + (InMST ? " " : "*") + | |||
| 545 | (IsCritical ? "c" : " ") + " W=" + Twine(Weight)).str(); | |||
| 546 | } | |||
| 547 | }; | |||
| 548 | ||||
| 549 | // This class stores the auxiliary information for each BB. | |||
| 550 | struct BBInfo { | |||
| 551 | BBInfo *Group; | |||
| 552 | uint32_t Index; | |||
| 553 | uint32_t Rank = 0; | |||
| 554 | ||||
| 555 | BBInfo(unsigned IX) : Group(this), Index(IX) {} | |||
| 556 | ||||
| 557 | // Return the information string of this object. | |||
| 558 | std::string infoString() const { | |||
| 559 | return (Twine("Index=") + Twine(Index)).str(); | |||
| 560 | } | |||
| 561 | ||||
| 562 | // Empty function -- only applicable to UseBBInfo. | |||
| 563 | void addOutEdge(PGOEdge *E LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__))) {} | |||
| 564 | ||||
| 565 | // Empty function -- only applicable to UseBBInfo. | |||
| 566 | void addInEdge(PGOEdge *E LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__))) {} | |||
| 567 | }; | |||
| 568 | ||||
| 569 | // This class implements the CFG edges. Note the CFG can be a multi-graph. | |||
| 570 | template <class Edge, class BBInfo> class FuncPGOInstrumentation { | |||
| 571 | private: | |||
| 572 | Function &F; | |||
| 573 | ||||
| 574 | // Is this is context-sensitive instrumentation. | |||
| 575 | bool IsCS; | |||
| 576 | ||||
| 577 | // A map that stores the Comdat group in function F. | |||
| 578 | std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers; | |||
| 579 | ||||
| 580 | ValueProfileCollector VPC; | |||
| 581 | ||||
| 582 | void computeCFGHash(); | |||
| 583 | void renameComdatFunction(); | |||
| 584 | ||||
| 585 | public: | |||
| 586 | std::vector<std::vector<VPCandidateInfo>> ValueSites; | |||
| 587 | SelectInstVisitor SIVisitor; | |||
| 588 | std::string FuncName; | |||
| 589 | GlobalVariable *FuncNameVar; | |||
| 590 | ||||
| 591 | // CFG hash value for this function. | |||
| 592 | uint64_t FunctionHash = 0; | |||
| 593 | ||||
| 594 | // The Minimum Spanning Tree of function CFG. | |||
| 595 | CFGMST<Edge, BBInfo> MST; | |||
| 596 | ||||
| 597 | // Collect all the BBs that will be instrumented, and store them in | |||
| 598 | // InstrumentBBs. | |||
| 599 | void getInstrumentBBs(std::vector<BasicBlock *> &InstrumentBBs); | |||
| 600 | ||||
| 601 | // Give an edge, find the BB that will be instrumented. | |||
| 602 | // Return nullptr if there is no BB to be instrumented. | |||
| 603 | BasicBlock *getInstrBB(Edge *E); | |||
| 604 | ||||
| 605 | // Return the auxiliary BB information. | |||
| 606 | BBInfo &getBBInfo(const BasicBlock *BB) const { return MST.getBBInfo(BB); } | |||
| 607 | ||||
| 608 | // Return the auxiliary BB information if available. | |||
| 609 | BBInfo *findBBInfo(const BasicBlock *BB) const { return MST.findBBInfo(BB); } | |||
| 610 | ||||
| 611 | // Dump edges and BB information. | |||
| 612 | void dumpInfo(std::string Str = "") const { | |||
| 613 | MST.dumpEdges(dbgs(), Twine("Dump Function ") + FuncName + " Hash: " + | |||
| 614 | Twine(FunctionHash) + "\t" + Str); | |||
| 615 | } | |||
| 616 | ||||
| 617 | FuncPGOInstrumentation( | |||
| 618 | Function &Func, TargetLibraryInfo &TLI, | |||
| 619 | std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers, | |||
| 620 | bool CreateGlobalVar = false, BranchProbabilityInfo *BPI = nullptr, | |||
| 621 | BlockFrequencyInfo *BFI = nullptr, bool IsCS = false, | |||
| 622 | bool InstrumentFuncEntry = true) | |||
| 623 | : F(Func), IsCS(IsCS), ComdatMembers(ComdatMembers), VPC(Func, TLI), | |||
| 624 | ValueSites(IPVK_Last + 1), SIVisitor(Func), | |||
| 625 | MST(F, InstrumentFuncEntry, BPI, BFI) { | |||
| 626 | // This should be done before CFG hash computation. | |||
| 627 | SIVisitor.countSelects(Func); | |||
| 628 | ValueSites[IPVK_MemOPSize] = VPC.get(IPVK_MemOPSize); | |||
| 629 | if (!IsCS) { | |||
| 630 | NumOfPGOSelectInsts += SIVisitor.getNumOfSelectInsts(); | |||
| 631 | NumOfPGOMemIntrinsics += ValueSites[IPVK_MemOPSize].size(); | |||
| 632 | NumOfPGOBB += MST.BBInfos.size(); | |||
| 633 | ValueSites[IPVK_IndirectCallTarget] = VPC.get(IPVK_IndirectCallTarget); | |||
| 634 | } else { | |||
| 635 | NumOfCSPGOSelectInsts += SIVisitor.getNumOfSelectInsts(); | |||
| 636 | NumOfCSPGOMemIntrinsics += ValueSites[IPVK_MemOPSize].size(); | |||
| 637 | NumOfCSPGOBB += MST.BBInfos.size(); | |||
| 638 | } | |||
| 639 | ||||
| 640 | FuncName = getPGOFuncName(F); | |||
| 641 | computeCFGHash(); | |||
| 642 | if (!ComdatMembers.empty()) | |||
| 643 | renameComdatFunction(); | |||
| 644 | LLVM_DEBUG(dumpInfo("after CFGMST"))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dumpInfo("after CFGMST"); } } while (false); | |||
| 645 | ||||
| 646 | for (auto &E : MST.AllEdges) { | |||
| 647 | if (E->Removed) | |||
| 648 | continue; | |||
| 649 | IsCS ? NumOfCSPGOEdge++ : NumOfPGOEdge++; | |||
| 650 | if (!E->InMST) | |||
| 651 | IsCS ? NumOfCSPGOInstrument++ : NumOfPGOInstrument++; | |||
| 652 | } | |||
| 653 | ||||
| 654 | if (CreateGlobalVar) | |||
| 655 | FuncNameVar = createPGOFuncNameVar(F, FuncName); | |||
| 656 | } | |||
| 657 | }; | |||
| 658 | ||||
| 659 | } // end anonymous namespace | |||
| 660 | ||||
| 661 | // Compute Hash value for the CFG: the lower 32 bits are CRC32 of the index | |||
| 662 | // value of each BB in the CFG. The higher 32 bits are the CRC32 of the numbers | |||
| 663 | // of selects, indirect calls, mem ops and edges. | |||
| 664 | template <class Edge, class BBInfo> | |||
| 665 | void FuncPGOInstrumentation<Edge, BBInfo>::computeCFGHash() { | |||
| 666 | std::vector<uint8_t> Indexes; | |||
| 667 | JamCRC JC; | |||
| 668 | for (auto &BB : F) { | |||
| 669 | const Instruction *TI = BB.getTerminator(); | |||
| 670 | for (unsigned I = 0, E = TI->getNumSuccessors(); I != E; ++I) { | |||
| 671 | BasicBlock *Succ = TI->getSuccessor(I); | |||
| 672 | auto BI = findBBInfo(Succ); | |||
| 673 | if (BI == nullptr) | |||
| 674 | continue; | |||
| 675 | uint32_t Index = BI->Index; | |||
| 676 | for (int J = 0; J < 4; J++) | |||
| 677 | Indexes.push_back((uint8_t)(Index >> (J * 8))); | |||
| 678 | } | |||
| 679 | } | |||
| 680 | JC.update(Indexes); | |||
| 681 | ||||
| 682 | JamCRC JCH; | |||
| 683 | if (PGOOldCFGHashing) { | |||
| 684 | // Hash format for context sensitive profile. Reserve 4 bits for other | |||
| 685 | // information. | |||
| 686 | FunctionHash = (uint64_t)SIVisitor.getNumOfSelectInsts() << 56 | | |||
| 687 | (uint64_t)ValueSites[IPVK_IndirectCallTarget].size() << 48 | | |||
| 688 | //(uint64_t)ValueSites[IPVK_MemOPSize].size() << 40 | | |||
| 689 | (uint64_t)MST.AllEdges.size() << 32 | JC.getCRC(); | |||
| 690 | } else { | |||
| 691 | // The higher 32 bits. | |||
| 692 | auto updateJCH = [&JCH](uint64_t Num) { | |||
| 693 | uint8_t Data[8]; | |||
| 694 | support::endian::write64le(Data, Num); | |||
| 695 | JCH.update(Data); | |||
| 696 | }; | |||
| 697 | updateJCH((uint64_t)SIVisitor.getNumOfSelectInsts()); | |||
| 698 | updateJCH((uint64_t)ValueSites[IPVK_IndirectCallTarget].size()); | |||
| 699 | updateJCH((uint64_t)ValueSites[IPVK_MemOPSize].size()); | |||
| 700 | updateJCH((uint64_t)MST.AllEdges.size()); | |||
| 701 | ||||
| 702 | // Hash format for context sensitive profile. Reserve 4 bits for other | |||
| 703 | // information. | |||
| 704 | FunctionHash = (((uint64_t)JCH.getCRC()) << 28) + JC.getCRC(); | |||
| 705 | } | |||
| 706 | ||||
| 707 | // Reserve bit 60-63 for other information purpose. | |||
| 708 | FunctionHash &= 0x0FFFFFFFFFFFFFFF; | |||
| 709 | if (IsCS) | |||
| 710 | NamedInstrProfRecord::setCSFlagInHash(FunctionHash); | |||
| 711 | LLVM_DEBUG(dbgs() << "Function Hash Computation for " << F.getName() << ":\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Function Hash Computation for " << F.getName() << ":\n" << " CRC = " << JC.getCRC() << ", Selects = " << SIVisitor.getNumOfSelectInsts () << ", Edges = " << MST.AllEdges.size() << ", ICSites = " << ValueSites[IPVK_IndirectCallTarget]. size(); } } while (false) | |||
| 712 | << " CRC = " << JC.getCRC()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Function Hash Computation for " << F.getName() << ":\n" << " CRC = " << JC.getCRC() << ", Selects = " << SIVisitor.getNumOfSelectInsts () << ", Edges = " << MST.AllEdges.size() << ", ICSites = " << ValueSites[IPVK_IndirectCallTarget]. size(); } } while (false) | |||
| 713 | << ", Selects = " << SIVisitor.getNumOfSelectInsts()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Function Hash Computation for " << F.getName() << ":\n" << " CRC = " << JC.getCRC() << ", Selects = " << SIVisitor.getNumOfSelectInsts () << ", Edges = " << MST.AllEdges.size() << ", ICSites = " << ValueSites[IPVK_IndirectCallTarget]. size(); } } while (false) | |||
| 714 | << ", Edges = " << MST.AllEdges.size() << ", ICSites = "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Function Hash Computation for " << F.getName() << ":\n" << " CRC = " << JC.getCRC() << ", Selects = " << SIVisitor.getNumOfSelectInsts () << ", Edges = " << MST.AllEdges.size() << ", ICSites = " << ValueSites[IPVK_IndirectCallTarget]. size(); } } while (false) | |||
| 715 | << ValueSites[IPVK_IndirectCallTarget].size())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Function Hash Computation for " << F.getName() << ":\n" << " CRC = " << JC.getCRC() << ", Selects = " << SIVisitor.getNumOfSelectInsts () << ", Edges = " << MST.AllEdges.size() << ", ICSites = " << ValueSites[IPVK_IndirectCallTarget]. size(); } } while (false); | |||
| 716 | if (!PGOOldCFGHashing) { | |||
| 717 | LLVM_DEBUG(dbgs() << ", Memops = " << ValueSites[IPVK_MemOPSize].size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << ", Memops = " << ValueSites[IPVK_MemOPSize].size() << ", High32 CRC = " << JCH.getCRC(); } } while (false) | |||
| 718 | << ", High32 CRC = " << JCH.getCRC())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << ", Memops = " << ValueSites[IPVK_MemOPSize].size() << ", High32 CRC = " << JCH.getCRC(); } } while (false); | |||
| 719 | } | |||
| 720 | LLVM_DEBUG(dbgs() << ", Hash = " << FunctionHash << "\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << ", Hash = " << FunctionHash << "\n";; } } while (false); | |||
| 721 | } | |||
| 722 | ||||
| 723 | // Check if we can safely rename this Comdat function. | |||
| 724 | static bool canRenameComdat( | |||
| 725 | Function &F, | |||
| 726 | std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers) { | |||
| 727 | if (!DoComdatRenaming || !canRenameComdatFunc(F, true)) | |||
| 728 | return false; | |||
| 729 | ||||
| 730 | // FIXME: Current only handle those Comdat groups that only containing one | |||
| 731 | // function. | |||
| 732 | // (1) For a Comdat group containing multiple functions, we need to have a | |||
| 733 | // unique postfix based on the hashes for each function. There is a | |||
| 734 | // non-trivial code refactoring to do this efficiently. | |||
| 735 | // (2) Variables can not be renamed, so we can not rename Comdat function in a | |||
| 736 | // group including global vars. | |||
| 737 | Comdat *C = F.getComdat(); | |||
| 738 | for (auto &&CM : make_range(ComdatMembers.equal_range(C))) { | |||
| 739 | assert(!isa<GlobalAlias>(CM.second))(static_cast <bool> (!isa<GlobalAlias>(CM.second) ) ? void (0) : __assert_fail ("!isa<GlobalAlias>(CM.second)" , "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 739, __extension__ __PRETTY_FUNCTION__)); | |||
| 740 | Function *FM = dyn_cast<Function>(CM.second); | |||
| 741 | if (FM != &F) | |||
| 742 | return false; | |||
| 743 | } | |||
| 744 | return true; | |||
| 745 | } | |||
| 746 | ||||
| 747 | // Append the CFGHash to the Comdat function name. | |||
| 748 | template <class Edge, class BBInfo> | |||
| 749 | void FuncPGOInstrumentation<Edge, BBInfo>::renameComdatFunction() { | |||
| 750 | if (!canRenameComdat(F, ComdatMembers)) | |||
| 751 | return; | |||
| 752 | std::string OrigName = F.getName().str(); | |||
| 753 | std::string NewFuncName = | |||
| 754 | Twine(F.getName() + "." + Twine(FunctionHash)).str(); | |||
| 755 | F.setName(Twine(NewFuncName)); | |||
| 756 | GlobalAlias::create(GlobalValue::WeakAnyLinkage, OrigName, &F); | |||
| 757 | FuncName = Twine(FuncName + "." + Twine(FunctionHash)).str(); | |||
| 758 | Comdat *NewComdat; | |||
| 759 | Module *M = F.getParent(); | |||
| 760 | // For AvailableExternallyLinkage functions, change the linkage to | |||
| 761 | // LinkOnceODR and put them into comdat. This is because after renaming, there | |||
| 762 | // is no backup external copy available for the function. | |||
| 763 | if (!F.hasComdat()) { | |||
| 764 | assert(F.getLinkage() == GlobalValue::AvailableExternallyLinkage)(static_cast <bool> (F.getLinkage() == GlobalValue::AvailableExternallyLinkage ) ? void (0) : __assert_fail ("F.getLinkage() == GlobalValue::AvailableExternallyLinkage" , "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 764, __extension__ __PRETTY_FUNCTION__)); | |||
| 765 | NewComdat = M->getOrInsertComdat(StringRef(NewFuncName)); | |||
| 766 | F.setLinkage(GlobalValue::LinkOnceODRLinkage); | |||
| 767 | F.setComdat(NewComdat); | |||
| 768 | return; | |||
| 769 | } | |||
| 770 | ||||
| 771 | // This function belongs to a single function Comdat group. | |||
| 772 | Comdat *OrigComdat = F.getComdat(); | |||
| 773 | std::string NewComdatName = | |||
| 774 | Twine(OrigComdat->getName() + "." + Twine(FunctionHash)).str(); | |||
| 775 | NewComdat = M->getOrInsertComdat(StringRef(NewComdatName)); | |||
| 776 | NewComdat->setSelectionKind(OrigComdat->getSelectionKind()); | |||
| 777 | ||||
| 778 | for (auto &&CM : make_range(ComdatMembers.equal_range(OrigComdat))) { | |||
| 779 | // Must be a function. | |||
| 780 | cast<Function>(CM.second)->setComdat(NewComdat); | |||
| 781 | } | |||
| 782 | } | |||
| 783 | ||||
| 784 | // Collect all the BBs that will be instruments and return them in | |||
| 785 | // InstrumentBBs and setup InEdges/OutEdge for UseBBInfo. | |||
| 786 | template <class Edge, class BBInfo> | |||
| 787 | void FuncPGOInstrumentation<Edge, BBInfo>::getInstrumentBBs( | |||
| 788 | std::vector<BasicBlock *> &InstrumentBBs) { | |||
| 789 | // Use a worklist as we will update the vector during the iteration. | |||
| 790 | std::vector<Edge *> EdgeList; | |||
| 791 | EdgeList.reserve(MST.AllEdges.size()); | |||
| 792 | for (auto &E : MST.AllEdges) | |||
| 793 | EdgeList.push_back(E.get()); | |||
| 794 | ||||
| 795 | for (auto &E : EdgeList) { | |||
| 796 | BasicBlock *InstrBB = getInstrBB(E); | |||
| 797 | if (InstrBB) | |||
| 798 | InstrumentBBs.push_back(InstrBB); | |||
| 799 | } | |||
| 800 | ||||
| 801 | // Set up InEdges/OutEdges for all BBs. | |||
| 802 | for (auto &E : MST.AllEdges) { | |||
| 803 | if (E->Removed) | |||
| 804 | continue; | |||
| 805 | const BasicBlock *SrcBB = E->SrcBB; | |||
| 806 | const BasicBlock *DestBB = E->DestBB; | |||
| 807 | BBInfo &SrcInfo = getBBInfo(SrcBB); | |||
| 808 | BBInfo &DestInfo = getBBInfo(DestBB); | |||
| 809 | SrcInfo.addOutEdge(E.get()); | |||
| 810 | DestInfo.addInEdge(E.get()); | |||
| 811 | } | |||
| 812 | } | |||
| 813 | ||||
| 814 | // Given a CFG E to be instrumented, find which BB to place the instrumented | |||
| 815 | // code. The function will split the critical edge if necessary. | |||
| 816 | template <class Edge, class BBInfo> | |||
| 817 | BasicBlock *FuncPGOInstrumentation<Edge, BBInfo>::getInstrBB(Edge *E) { | |||
| 818 | if (E->InMST || E->Removed) | |||
| 819 | return nullptr; | |||
| 820 | ||||
| 821 | BasicBlock *SrcBB = const_cast<BasicBlock *>(E->SrcBB); | |||
| 822 | BasicBlock *DestBB = const_cast<BasicBlock *>(E->DestBB); | |||
| 823 | // For a fake edge, instrument the real BB. | |||
| 824 | if (SrcBB == nullptr) | |||
| 825 | return DestBB; | |||
| 826 | if (DestBB == nullptr) | |||
| 827 | return SrcBB; | |||
| 828 | ||||
| 829 | auto canInstrument = [](BasicBlock *BB) -> BasicBlock * { | |||
| 830 | // There are basic blocks (such as catchswitch) cannot be instrumented. | |||
| 831 | // If the returned first insertion point is the end of BB, skip this BB. | |||
| 832 | if (BB->getFirstInsertionPt() == BB->end()) | |||
| 833 | return nullptr; | |||
| 834 | return BB; | |||
| 835 | }; | |||
| 836 | ||||
| 837 | // Instrument the SrcBB if it has a single successor, | |||
| 838 | // otherwise, the DestBB if this is not a critical edge. | |||
| 839 | Instruction *TI = SrcBB->getTerminator(); | |||
| 840 | if (TI->getNumSuccessors() <= 1) | |||
| 841 | return canInstrument(SrcBB); | |||
| 842 | if (!E->IsCritical) | |||
| 843 | return canInstrument(DestBB); | |||
| 844 | ||||
| 845 | // Some IndirectBr critical edges cannot be split by the previous | |||
| 846 | // SplitIndirectBrCriticalEdges call. Bail out. | |||
| 847 | unsigned SuccNum = GetSuccessorNumber(SrcBB, DestBB); | |||
| 848 | BasicBlock *InstrBB = | |||
| 849 | isa<IndirectBrInst>(TI) ? nullptr : SplitCriticalEdge(TI, SuccNum); | |||
| 850 | if (!InstrBB) { | |||
| 851 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Fail to split critical edge: not instrument this edge.\n" ; } } while (false) | |||
| 852 | dbgs() << "Fail to split critical edge: not instrument this edge.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Fail to split critical edge: not instrument this edge.\n" ; } } while (false); | |||
| 853 | return nullptr; | |||
| 854 | } | |||
| 855 | // For a critical edge, we have to split. Instrument the newly | |||
| 856 | // created BB. | |||
| 857 | IsCS ? NumOfCSPGOSplit++ : NumOfPGOSplit++; | |||
| 858 | LLVM_DEBUG(dbgs() << "Split critical edge: " << getBBInfo(SrcBB).Indexdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Split critical edge: " << getBBInfo(SrcBB).Index << " --> " << getBBInfo(DestBB).Index << "\n"; } } while (false) | |||
| 859 | << " --> " << getBBInfo(DestBB).Index << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Split critical edge: " << getBBInfo(SrcBB).Index << " --> " << getBBInfo(DestBB).Index << "\n"; } } while (false); | |||
| 860 | // Need to add two new edges. First one: Add new edge of SrcBB->InstrBB. | |||
| 861 | MST.addEdge(SrcBB, InstrBB, 0); | |||
| 862 | // Second one: Add new edge of InstrBB->DestBB. | |||
| 863 | Edge &NewEdge1 = MST.addEdge(InstrBB, DestBB, 0); | |||
| 864 | NewEdge1.InMST = true; | |||
| 865 | E->Removed = true; | |||
| 866 | ||||
| 867 | return canInstrument(InstrBB); | |||
| 868 | } | |||
| 869 | ||||
| 870 | // When generating value profiling calls on Windows routines that make use of | |||
| 871 | // handler funclets for exception processing an operand bundle needs to attached | |||
| 872 | // to the called function. This routine will set \p OpBundles to contain the | |||
| 873 | // funclet information, if any is needed, that should be placed on the generated | |||
| 874 | // value profiling call for the value profile candidate call. | |||
| 875 | static void | |||
| 876 | populateEHOperandBundle(VPCandidateInfo &Cand, | |||
| 877 | DenseMap<BasicBlock *, ColorVector> &BlockColors, | |||
| 878 | SmallVectorImpl<OperandBundleDef> &OpBundles) { | |||
| 879 | auto *OrigCall = dyn_cast<CallBase>(Cand.AnnotatedInst); | |||
| 880 | if (OrigCall
| |||
| 881 | // The instrumentation call should belong to the same funclet as a | |||
| 882 | // non-intrinsic call, so just copy the operand bundle, if any exists. | |||
| 883 | Optional<OperandBundleUse> ParentFunclet = | |||
| 884 | OrigCall->getOperandBundle(LLVMContext::OB_funclet); | |||
| 885 | if (ParentFunclet) | |||
| 886 | OpBundles.emplace_back(OperandBundleDef(*ParentFunclet)); | |||
| 887 | } else { | |||
| 888 | // Intrinsics or other instructions do not get funclet information from the | |||
| 889 | // front-end. Need to use the BlockColors that was computed by the routine | |||
| 890 | // colorEHFunclets to determine whether a funclet is needed. | |||
| 891 | if (!BlockColors.empty()) { | |||
| 892 | const ColorVector &CV = BlockColors.find(OrigCall->getParent())->second; | |||
| ||||
| 893 | assert(CV.size() == 1 && "non-unique color for block!")(static_cast <bool> (CV.size() == 1 && "non-unique color for block!" ) ? void (0) : __assert_fail ("CV.size() == 1 && \"non-unique color for block!\"" , "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 893, __extension__ __PRETTY_FUNCTION__)); | |||
| 894 | Instruction *EHPad = CV.front()->getFirstNonPHI(); | |||
| 895 | if (EHPad->isEHPad()) | |||
| 896 | OpBundles.emplace_back("funclet", EHPad); | |||
| 897 | } | |||
| 898 | } | |||
| 899 | } | |||
| 900 | ||||
| 901 | // Visit all edge and instrument the edges not in MST, and do value profiling. | |||
| 902 | // Critical edges will be split. | |||
| 903 | static void instrumentOneFunc( | |||
| 904 | Function &F, Module *M, TargetLibraryInfo &TLI, BranchProbabilityInfo *BPI, | |||
| 905 | BlockFrequencyInfo *BFI, | |||
| 906 | std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers, | |||
| 907 | bool IsCS) { | |||
| 908 | // Split indirectbr critical edges here before computing the MST rather than | |||
| 909 | // later in getInstrBB() to avoid invalidating it. | |||
| 910 | SplitIndirectBrCriticalEdges(F, BPI, BFI); | |||
| 911 | ||||
| 912 | FuncPGOInstrumentation<PGOEdge, BBInfo> FuncInfo( | |||
| 913 | F, TLI, ComdatMembers, true, BPI, BFI, IsCS, PGOInstrumentEntry); | |||
| 914 | std::vector<BasicBlock *> InstrumentBBs; | |||
| 915 | FuncInfo.getInstrumentBBs(InstrumentBBs); | |||
| 916 | unsigned NumCounters = | |||
| 917 | InstrumentBBs.size() + FuncInfo.SIVisitor.getNumOfSelectInsts(); | |||
| 918 | ||||
| 919 | uint32_t I = 0; | |||
| 920 | Type *I8PtrTy = Type::getInt8PtrTy(M->getContext()); | |||
| 921 | for (auto *InstrBB : InstrumentBBs) { | |||
| 922 | IRBuilder<> Builder(InstrBB, InstrBB->getFirstInsertionPt()); | |||
| 923 | assert(Builder.GetInsertPoint() != InstrBB->end() &&(static_cast <bool> (Builder.GetInsertPoint() != InstrBB ->end() && "Cannot get the Instrumentation point") ? void (0) : __assert_fail ("Builder.GetInsertPoint() != InstrBB->end() && \"Cannot get the Instrumentation point\"" , "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 924, __extension__ __PRETTY_FUNCTION__)) | |||
| 924 | "Cannot get the Instrumentation point")(static_cast <bool> (Builder.GetInsertPoint() != InstrBB ->end() && "Cannot get the Instrumentation point") ? void (0) : __assert_fail ("Builder.GetInsertPoint() != InstrBB->end() && \"Cannot get the Instrumentation point\"" , "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 924, __extension__ __PRETTY_FUNCTION__)); | |||
| 925 | Builder.CreateCall( | |||
| 926 | Intrinsic::getDeclaration(M, Intrinsic::instrprof_increment), | |||
| 927 | {ConstantExpr::getBitCast(FuncInfo.FuncNameVar, I8PtrTy), | |||
| 928 | Builder.getInt64(FuncInfo.FunctionHash), Builder.getInt32(NumCounters), | |||
| 929 | Builder.getInt32(I++)}); | |||
| 930 | } | |||
| 931 | ||||
| 932 | // Now instrument select instructions: | |||
| 933 | FuncInfo.SIVisitor.instrumentSelects(F, &I, NumCounters, FuncInfo.FuncNameVar, | |||
| 934 | FuncInfo.FunctionHash); | |||
| 935 | assert(I == NumCounters)(static_cast <bool> (I == NumCounters) ? void (0) : __assert_fail ("I == NumCounters", "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 935, __extension__ __PRETTY_FUNCTION__)); | |||
| 936 | ||||
| 937 | if (DisableValueProfiling) | |||
| 938 | return; | |||
| 939 | ||||
| 940 | NumOfPGOICall += FuncInfo.ValueSites[IPVK_IndirectCallTarget].size(); | |||
| 941 | ||||
| 942 | // Intrinsic function calls do not have funclet operand bundles needed for | |||
| 943 | // Windows exception handling attached to them. However, if value profiling is | |||
| 944 | // inserted for one of these calls, then a funclet value will need to be set | |||
| 945 | // on the instrumentation call based on the funclet coloring. | |||
| 946 | DenseMap<BasicBlock *, ColorVector> BlockColors; | |||
| 947 | if (F.hasPersonalityFn() && | |||
| 948 | isFuncletEHPersonality(classifyEHPersonality(F.getPersonalityFn()))) | |||
| 949 | BlockColors = colorEHFunclets(F); | |||
| 950 | ||||
| 951 | // For each VP Kind, walk the VP candidates and instrument each one. | |||
| 952 | for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) { | |||
| 953 | unsigned SiteIndex = 0; | |||
| 954 | if (Kind
| |||
| 955 | continue; | |||
| 956 | ||||
| 957 | for (VPCandidateInfo Cand : FuncInfo.ValueSites[Kind]) { | |||
| 958 | LLVM_DEBUG(dbgs() << "Instrument one VP " << ValueProfKindDescr[Kind]do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Instrument one VP " << ValueProfKindDescr[Kind] << " site: CallSite Index = " << SiteIndex << "\n"; } } while (false) | |||
| 959 | << " site: CallSite Index = " << SiteIndex << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Instrument one VP " << ValueProfKindDescr[Kind] << " site: CallSite Index = " << SiteIndex << "\n"; } } while (false); | |||
| 960 | ||||
| 961 | IRBuilder<> Builder(Cand.InsertPt); | |||
| 962 | assert(Builder.GetInsertPoint() != Cand.InsertPt->getParent()->end() &&(static_cast <bool> (Builder.GetInsertPoint() != Cand.InsertPt ->getParent()->end() && "Cannot get the Instrumentation point" ) ? void (0) : __assert_fail ("Builder.GetInsertPoint() != Cand.InsertPt->getParent()->end() && \"Cannot get the Instrumentation point\"" , "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 963, __extension__ __PRETTY_FUNCTION__)) | |||
| 963 | "Cannot get the Instrumentation point")(static_cast <bool> (Builder.GetInsertPoint() != Cand.InsertPt ->getParent()->end() && "Cannot get the Instrumentation point" ) ? void (0) : __assert_fail ("Builder.GetInsertPoint() != Cand.InsertPt->getParent()->end() && \"Cannot get the Instrumentation point\"" , "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 963, __extension__ __PRETTY_FUNCTION__)); | |||
| 964 | ||||
| 965 | Value *ToProfile = nullptr; | |||
| 966 | if (Cand.V->getType()->isIntegerTy()) | |||
| 967 | ToProfile = Builder.CreateZExtOrTrunc(Cand.V, Builder.getInt64Ty()); | |||
| 968 | else if (Cand.V->getType()->isPointerTy()) | |||
| 969 | ToProfile = Builder.CreatePtrToInt(Cand.V, Builder.getInt64Ty()); | |||
| 970 | assert(ToProfile && "value profiling Value is of unexpected type")(static_cast <bool> (ToProfile && "value profiling Value is of unexpected type" ) ? void (0) : __assert_fail ("ToProfile && \"value profiling Value is of unexpected type\"" , "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 970, __extension__ __PRETTY_FUNCTION__)); | |||
| 971 | ||||
| 972 | SmallVector<OperandBundleDef, 1> OpBundles; | |||
| 973 | populateEHOperandBundle(Cand, BlockColors, OpBundles); | |||
| 974 | Builder.CreateCall( | |||
| 975 | Intrinsic::getDeclaration(M, Intrinsic::instrprof_value_profile), | |||
| 976 | {ConstantExpr::getBitCast(FuncInfo.FuncNameVar, I8PtrTy), | |||
| 977 | Builder.getInt64(FuncInfo.FunctionHash), ToProfile, | |||
| 978 | Builder.getInt32(Kind), Builder.getInt32(SiteIndex++)}, | |||
| 979 | OpBundles); | |||
| 980 | } | |||
| 981 | } // IPVK_First <= Kind <= IPVK_Last | |||
| 982 | } | |||
| 983 | ||||
| 984 | namespace { | |||
| 985 | ||||
| 986 | // This class represents a CFG edge in profile use compilation. | |||
| 987 | struct PGOUseEdge : public PGOEdge { | |||
| 988 | bool CountValid = false; | |||
| 989 | uint64_t CountValue = 0; | |||
| 990 | ||||
| 991 | PGOUseEdge(const BasicBlock *Src, const BasicBlock *Dest, uint64_t W = 1) | |||
| 992 | : PGOEdge(Src, Dest, W) {} | |||
| 993 | ||||
| 994 | // Set edge count value | |||
| 995 | void setEdgeCount(uint64_t Value) { | |||
| 996 | CountValue = Value; | |||
| 997 | CountValid = true; | |||
| 998 | } | |||
| 999 | ||||
| 1000 | // Return the information string for this object. | |||
| 1001 | std::string infoString() const { | |||
| 1002 | if (!CountValid) | |||
| 1003 | return PGOEdge::infoString(); | |||
| 1004 | return (Twine(PGOEdge::infoString()) + " Count=" + Twine(CountValue)) | |||
| 1005 | .str(); | |||
| 1006 | } | |||
| 1007 | }; | |||
| 1008 | ||||
| 1009 | using DirectEdges = SmallVector<PGOUseEdge *, 2>; | |||
| 1010 | ||||
| 1011 | // This class stores the auxiliary information for each BB. | |||
| 1012 | struct UseBBInfo : public BBInfo { | |||
| 1013 | uint64_t CountValue = 0; | |||
| 1014 | bool CountValid; | |||
| 1015 | int32_t UnknownCountInEdge = 0; | |||
| 1016 | int32_t UnknownCountOutEdge = 0; | |||
| 1017 | DirectEdges InEdges; | |||
| 1018 | DirectEdges OutEdges; | |||
| 1019 | ||||
| 1020 | UseBBInfo(unsigned IX) : BBInfo(IX), CountValid(false) {} | |||
| 1021 | ||||
| 1022 | UseBBInfo(unsigned IX, uint64_t C) | |||
| 1023 | : BBInfo(IX), CountValue(C), CountValid(true) {} | |||
| 1024 | ||||
| 1025 | // Set the profile count value for this BB. | |||
| 1026 | void setBBInfoCount(uint64_t Value) { | |||
| 1027 | CountValue = Value; | |||
| 1028 | CountValid = true; | |||
| 1029 | } | |||
| 1030 | ||||
| 1031 | // Return the information string of this object. | |||
| 1032 | std::string infoString() const { | |||
| 1033 | if (!CountValid) | |||
| 1034 | return BBInfo::infoString(); | |||
| 1035 | return (Twine(BBInfo::infoString()) + " Count=" + Twine(CountValue)).str(); | |||
| 1036 | } | |||
| 1037 | ||||
| 1038 | // Add an OutEdge and update the edge count. | |||
| 1039 | void addOutEdge(PGOUseEdge *E) { | |||
| 1040 | OutEdges.push_back(E); | |||
| 1041 | UnknownCountOutEdge++; | |||
| 1042 | } | |||
| 1043 | ||||
| 1044 | // Add an InEdge and update the edge count. | |||
| 1045 | void addInEdge(PGOUseEdge *E) { | |||
| 1046 | InEdges.push_back(E); | |||
| 1047 | UnknownCountInEdge++; | |||
| 1048 | } | |||
| 1049 | }; | |||
| 1050 | ||||
| 1051 | } // end anonymous namespace | |||
| 1052 | ||||
| 1053 | // Sum up the count values for all the edges. | |||
| 1054 | static uint64_t sumEdgeCount(const ArrayRef<PGOUseEdge *> Edges) { | |||
| 1055 | uint64_t Total = 0; | |||
| 1056 | for (auto &E : Edges) { | |||
| 1057 | if (E->Removed) | |||
| 1058 | continue; | |||
| 1059 | Total += E->CountValue; | |||
| 1060 | } | |||
| 1061 | return Total; | |||
| 1062 | } | |||
| 1063 | ||||
| 1064 | namespace { | |||
| 1065 | ||||
| 1066 | class PGOUseFunc { | |||
| 1067 | public: | |||
| 1068 | PGOUseFunc(Function &Func, Module *Modu, TargetLibraryInfo &TLI, | |||
| 1069 | std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers, | |||
| 1070 | BranchProbabilityInfo *BPI, BlockFrequencyInfo *BFIin, | |||
| 1071 | ProfileSummaryInfo *PSI, bool IsCS, bool InstrumentFuncEntry) | |||
| 1072 | : F(Func), M(Modu), BFI(BFIin), PSI(PSI), | |||
| 1073 | FuncInfo(Func, TLI, ComdatMembers, false, BPI, BFIin, IsCS, | |||
| 1074 | InstrumentFuncEntry), | |||
| 1075 | FreqAttr(FFA_Normal), IsCS(IsCS) {} | |||
| 1076 | ||||
| 1077 | // Read counts for the instrumented BB from profile. | |||
| 1078 | bool readCounters(IndexedInstrProfReader *PGOReader, bool &AllZeros, | |||
| 1079 | bool &AllMinusOnes); | |||
| 1080 | ||||
| 1081 | // Populate the counts for all BBs. | |||
| 1082 | void populateCounters(); | |||
| 1083 | ||||
| 1084 | // Set the branch weights based on the count values. | |||
| 1085 | void setBranchWeights(); | |||
| 1086 | ||||
| 1087 | // Annotate the value profile call sites for all value kind. | |||
| 1088 | void annotateValueSites(); | |||
| 1089 | ||||
| 1090 | // Annotate the value profile call sites for one value kind. | |||
| 1091 | void annotateValueSites(uint32_t Kind); | |||
| 1092 | ||||
| 1093 | // Annotate the irreducible loop header weights. | |||
| 1094 | void annotateIrrLoopHeaderWeights(); | |||
| 1095 | ||||
| 1096 | // The hotness of the function from the profile count. | |||
| 1097 | enum FuncFreqAttr { FFA_Normal, FFA_Cold, FFA_Hot }; | |||
| 1098 | ||||
| 1099 | // Return the function hotness from the profile. | |||
| 1100 | FuncFreqAttr getFuncFreqAttr() const { return FreqAttr; } | |||
| 1101 | ||||
| 1102 | // Return the function hash. | |||
| 1103 | uint64_t getFuncHash() const { return FuncInfo.FunctionHash; } | |||
| 1104 | ||||
| 1105 | // Return the profile record for this function; | |||
| 1106 | InstrProfRecord &getProfileRecord() { return ProfileRecord; } | |||
| 1107 | ||||
| 1108 | // Return the auxiliary BB information. | |||
| 1109 | UseBBInfo &getBBInfo(const BasicBlock *BB) const { | |||
| 1110 | return FuncInfo.getBBInfo(BB); | |||
| 1111 | } | |||
| 1112 | ||||
| 1113 | // Return the auxiliary BB information if available. | |||
| 1114 | UseBBInfo *findBBInfo(const BasicBlock *BB) const { | |||
| 1115 | return FuncInfo.findBBInfo(BB); | |||
| 1116 | } | |||
| 1117 | ||||
| 1118 | Function &getFunc() const { return F; } | |||
| 1119 | ||||
| 1120 | void dumpInfo(std::string Str = "") const { | |||
| 1121 | FuncInfo.dumpInfo(Str); | |||
| 1122 | } | |||
| 1123 | ||||
| 1124 | uint64_t getProgramMaxCount() const { return ProgramMaxCount; } | |||
| 1125 | private: | |||
| 1126 | Function &F; | |||
| 1127 | Module *M; | |||
| 1128 | BlockFrequencyInfo *BFI; | |||
| 1129 | ProfileSummaryInfo *PSI; | |||
| 1130 | ||||
| 1131 | // This member stores the shared information with class PGOGenFunc. | |||
| 1132 | FuncPGOInstrumentation<PGOUseEdge, UseBBInfo> FuncInfo; | |||
| 1133 | ||||
| 1134 | // The maximum count value in the profile. This is only used in PGO use | |||
| 1135 | // compilation. | |||
| 1136 | uint64_t ProgramMaxCount; | |||
| 1137 | ||||
| 1138 | // Position of counter that remains to be read. | |||
| 1139 | uint32_t CountPosition = 0; | |||
| 1140 | ||||
| 1141 | // Total size of the profile count for this function. | |||
| 1142 | uint32_t ProfileCountSize = 0; | |||
| 1143 | ||||
| 1144 | // ProfileRecord for this function. | |||
| 1145 | InstrProfRecord ProfileRecord; | |||
| 1146 | ||||
| 1147 | // Function hotness info derived from profile. | |||
| 1148 | FuncFreqAttr FreqAttr; | |||
| 1149 | ||||
| 1150 | // Is to use the context sensitive profile. | |||
| 1151 | bool IsCS; | |||
| 1152 | ||||
| 1153 | // Find the Instrumented BB and set the value. Return false on error. | |||
| 1154 | bool setInstrumentedCounts(const std::vector<uint64_t> &CountFromProfile); | |||
| 1155 | ||||
| 1156 | // Set the edge counter value for the unknown edge -- there should be only | |||
| 1157 | // one unknown edge. | |||
| 1158 | void setEdgeCount(DirectEdges &Edges, uint64_t Value); | |||
| 1159 | ||||
| 1160 | // Return FuncName string; | |||
| 1161 | std::string getFuncName() const { return FuncInfo.FuncName; } | |||
| 1162 | ||||
| 1163 | // Set the hot/cold inline hints based on the count values. | |||
| 1164 | // FIXME: This function should be removed once the functionality in | |||
| 1165 | // the inliner is implemented. | |||
| 1166 | void markFunctionAttributes(uint64_t EntryCount, uint64_t MaxCount) { | |||
| 1167 | if (PSI->isHotCount(EntryCount)) | |||
| 1168 | FreqAttr = FFA_Hot; | |||
| 1169 | else if (PSI->isColdCount(MaxCount)) | |||
| 1170 | FreqAttr = FFA_Cold; | |||
| 1171 | } | |||
| 1172 | }; | |||
| 1173 | ||||
| 1174 | } // end anonymous namespace | |||
| 1175 | ||||
| 1176 | // Visit all the edges and assign the count value for the instrumented | |||
| 1177 | // edges and the BB. Return false on error. | |||
| 1178 | bool PGOUseFunc::setInstrumentedCounts( | |||
| 1179 | const std::vector<uint64_t> &CountFromProfile) { | |||
| 1180 | ||||
| 1181 | std::vector<BasicBlock *> InstrumentBBs; | |||
| 1182 | FuncInfo.getInstrumentBBs(InstrumentBBs); | |||
| 1183 | unsigned NumCounters = | |||
| 1184 | InstrumentBBs.size() + FuncInfo.SIVisitor.getNumOfSelectInsts(); | |||
| 1185 | // The number of counters here should match the number of counters | |||
| 1186 | // in profile. Return if they mismatch. | |||
| 1187 | if (NumCounters != CountFromProfile.size()) { | |||
| 1188 | return false; | |||
| 1189 | } | |||
| 1190 | auto *FuncEntry = &*F.begin(); | |||
| 1191 | ||||
| 1192 | // Set the profile count to the Instrumented BBs. | |||
| 1193 | uint32_t I = 0; | |||
| 1194 | for (BasicBlock *InstrBB : InstrumentBBs) { | |||
| 1195 | uint64_t CountValue = CountFromProfile[I++]; | |||
| 1196 | UseBBInfo &Info = getBBInfo(InstrBB); | |||
| 1197 | // If we reach here, we know that we have some nonzero count | |||
| 1198 | // values in this function. The entry count should not be 0. | |||
| 1199 | // Fix it if necessary. | |||
| 1200 | if (InstrBB == FuncEntry && CountValue == 0) | |||
| 1201 | CountValue = 1; | |||
| 1202 | Info.setBBInfoCount(CountValue); | |||
| 1203 | } | |||
| 1204 | ProfileCountSize = CountFromProfile.size(); | |||
| 1205 | CountPosition = I; | |||
| 1206 | ||||
| 1207 | // Set the edge count and update the count of unknown edges for BBs. | |||
| 1208 | auto setEdgeCount = [this](PGOUseEdge *E, uint64_t Value) -> void { | |||
| 1209 | E->setEdgeCount(Value); | |||
| 1210 | this->getBBInfo(E->SrcBB).UnknownCountOutEdge--; | |||
| 1211 | this->getBBInfo(E->DestBB).UnknownCountInEdge--; | |||
| 1212 | }; | |||
| 1213 | ||||
| 1214 | // Set the profile count the Instrumented edges. There are BBs that not in | |||
| 1215 | // MST but not instrumented. Need to set the edge count value so that we can | |||
| 1216 | // populate the profile counts later. | |||
| 1217 | for (auto &E : FuncInfo.MST.AllEdges) { | |||
| 1218 | if (E->Removed || E->InMST) | |||
| 1219 | continue; | |||
| 1220 | const BasicBlock *SrcBB = E->SrcBB; | |||
| 1221 | UseBBInfo &SrcInfo = getBBInfo(SrcBB); | |||
| 1222 | ||||
| 1223 | // If only one out-edge, the edge profile count should be the same as BB | |||
| 1224 | // profile count. | |||
| 1225 | if (SrcInfo.CountValid && SrcInfo.OutEdges.size() == 1) | |||
| 1226 | setEdgeCount(E.get(), SrcInfo.CountValue); | |||
| 1227 | else { | |||
| 1228 | const BasicBlock *DestBB = E->DestBB; | |||
| 1229 | UseBBInfo &DestInfo = getBBInfo(DestBB); | |||
| 1230 | // If only one in-edge, the edge profile count should be the same as BB | |||
| 1231 | // profile count. | |||
| 1232 | if (DestInfo.CountValid && DestInfo.InEdges.size() == 1) | |||
| 1233 | setEdgeCount(E.get(), DestInfo.CountValue); | |||
| 1234 | } | |||
| 1235 | if (E->CountValid) | |||
| 1236 | continue; | |||
| 1237 | // E's count should have been set from profile. If not, this meenas E skips | |||
| 1238 | // the instrumentation. We set the count to 0. | |||
| 1239 | setEdgeCount(E.get(), 0); | |||
| 1240 | } | |||
| 1241 | return true; | |||
| 1242 | } | |||
| 1243 | ||||
| 1244 | // Set the count value for the unknown edge. There should be one and only one | |||
| 1245 | // unknown edge in Edges vector. | |||
| 1246 | void PGOUseFunc::setEdgeCount(DirectEdges &Edges, uint64_t Value) { | |||
| 1247 | for (auto &E : Edges) { | |||
| 1248 | if (E->CountValid) | |||
| 1249 | continue; | |||
| 1250 | E->setEdgeCount(Value); | |||
| 1251 | ||||
| 1252 | getBBInfo(E->SrcBB).UnknownCountOutEdge--; | |||
| 1253 | getBBInfo(E->DestBB).UnknownCountInEdge--; | |||
| 1254 | return; | |||
| 1255 | } | |||
| 1256 | llvm_unreachable("Cannot find the unknown count edge")::llvm::llvm_unreachable_internal("Cannot find the unknown count edge" , "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 1256); | |||
| 1257 | } | |||
| 1258 | ||||
| 1259 | // Emit function metadata indicating PGO profile mismatch. | |||
| 1260 | static void annotateFunctionWithHashMismatch(Function &F, | |||
| 1261 | LLVMContext &ctx) { | |||
| 1262 | const char MetadataName[] = "instr_prof_hash_mismatch"; | |||
| 1263 | SmallVector<Metadata *, 2> Names; | |||
| 1264 | // If this metadata already exists, ignore. | |||
| 1265 | auto *Existing = F.getMetadata(LLVMContext::MD_annotation); | |||
| 1266 | if (Existing) { | |||
| 1267 | MDTuple *Tuple = cast<MDTuple>(Existing); | |||
| 1268 | for (auto &N : Tuple->operands()) { | |||
| 1269 | if (cast<MDString>(N.get())->getString() == MetadataName) | |||
| 1270 | return; | |||
| 1271 | Names.push_back(N.get()); | |||
| 1272 | } | |||
| 1273 | } | |||
| 1274 | ||||
| 1275 | MDBuilder MDB(ctx); | |||
| 1276 | Names.push_back(MDB.createString(MetadataName)); | |||
| 1277 | MDNode *MD = MDTuple::get(ctx, Names); | |||
| 1278 | F.setMetadata(LLVMContext::MD_annotation, MD); | |||
| 1279 | } | |||
| 1280 | ||||
| 1281 | // Read the profile from ProfileFileName and assign the value to the | |||
| 1282 | // instrumented BB and the edges. This function also updates ProgramMaxCount. | |||
| 1283 | // Return true if the profile are successfully read, and false on errors. | |||
| 1284 | bool PGOUseFunc::readCounters(IndexedInstrProfReader *PGOReader, bool &AllZeros, | |||
| 1285 | bool &AllMinusOnes) { | |||
| 1286 | auto &Ctx = M->getContext(); | |||
| 1287 | Expected<InstrProfRecord> Result = | |||
| 1288 | PGOReader->getInstrProfRecord(FuncInfo.FuncName, FuncInfo.FunctionHash); | |||
| 1289 | if (Error E = Result.takeError()) { | |||
| 1290 | handleAllErrors(std::move(E), [&](const InstrProfError &IPE) { | |||
| 1291 | auto Err = IPE.get(); | |||
| 1292 | bool SkipWarning = false; | |||
| 1293 | LLVM_DEBUG(dbgs() << "Error in reading profile for Func "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Error in reading profile for Func " << FuncInfo.FuncName << ": "; } } while (false) | |||
| 1294 | << FuncInfo.FuncName << ": ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Error in reading profile for Func " << FuncInfo.FuncName << ": "; } } while (false); | |||
| 1295 | if (Err == instrprof_error::unknown_function) { | |||
| 1296 | IsCS ? NumOfCSPGOMissing++ : NumOfPGOMissing++; | |||
| 1297 | SkipWarning = !PGOWarnMissing; | |||
| 1298 | LLVM_DEBUG(dbgs() << "unknown function")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "unknown function" ; } } while (false); | |||
| 1299 | } else if (Err == instrprof_error::hash_mismatch || | |||
| 1300 | Err == instrprof_error::malformed) { | |||
| 1301 | IsCS ? NumOfCSPGOMismatch++ : NumOfPGOMismatch++; | |||
| 1302 | SkipWarning = | |||
| 1303 | NoPGOWarnMismatch || | |||
| 1304 | (NoPGOWarnMismatchComdat && | |||
| 1305 | (F.hasComdat() || | |||
| 1306 | F.getLinkage() == GlobalValue::AvailableExternallyLinkage)); | |||
| 1307 | LLVM_DEBUG(dbgs() << "hash mismatch (skip=" << SkipWarning << ")")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "hash mismatch (skip=" << SkipWarning << ")"; } } while (false); | |||
| 1308 | // Emit function metadata indicating PGO profile mismatch. | |||
| 1309 | annotateFunctionWithHashMismatch(F, M->getContext()); | |||
| 1310 | } | |||
| 1311 | ||||
| 1312 | LLVM_DEBUG(dbgs() << " IsCS=" << IsCS << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << " IsCS=" << IsCS << "\n"; } } while (false); | |||
| 1313 | if (SkipWarning) | |||
| 1314 | return; | |||
| 1315 | ||||
| 1316 | std::string Msg = IPE.message() + std::string(" ") + F.getName().str() + | |||
| 1317 | std::string(" Hash = ") + | |||
| 1318 | std::to_string(FuncInfo.FunctionHash); | |||
| 1319 | ||||
| 1320 | Ctx.diagnose( | |||
| 1321 | DiagnosticInfoPGOProfile(M->getName().data(), Msg, DS_Warning)); | |||
| 1322 | }); | |||
| 1323 | return false; | |||
| 1324 | } | |||
| 1325 | ProfileRecord = std::move(Result.get()); | |||
| 1326 | std::vector<uint64_t> &CountFromProfile = ProfileRecord.Counts; | |||
| 1327 | ||||
| 1328 | IsCS ? NumOfCSPGOFunc++ : NumOfPGOFunc++; | |||
| 1329 | LLVM_DEBUG(dbgs() << CountFromProfile.size() << " counts\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << CountFromProfile.size () << " counts\n"; } } while (false); | |||
| 1330 | AllMinusOnes = (CountFromProfile.size() > 0); | |||
| 1331 | uint64_t ValueSum = 0; | |||
| 1332 | for (unsigned I = 0, S = CountFromProfile.size(); I < S; I++) { | |||
| 1333 | LLVM_DEBUG(dbgs() << " " << I << ": " << CountFromProfile[I] << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << " " << I << ": " << CountFromProfile[I] << "\n"; } } while ( false); | |||
| 1334 | ValueSum += CountFromProfile[I]; | |||
| 1335 | if (CountFromProfile[I] != (uint64_t)-1) | |||
| 1336 | AllMinusOnes = false; | |||
| 1337 | } | |||
| 1338 | AllZeros = (ValueSum == 0); | |||
| 1339 | ||||
| 1340 | LLVM_DEBUG(dbgs() << "SUM = " << ValueSum << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "SUM = " << ValueSum << "\n"; } } while (false); | |||
| 1341 | ||||
| 1342 | getBBInfo(nullptr).UnknownCountOutEdge = 2; | |||
| 1343 | getBBInfo(nullptr).UnknownCountInEdge = 2; | |||
| 1344 | ||||
| 1345 | if (!setInstrumentedCounts(CountFromProfile)) { | |||
| 1346 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Inconsistent number of counts, skipping this function" ; } } while (false) | |||
| 1347 | dbgs() << "Inconsistent number of counts, skipping this function")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Inconsistent number of counts, skipping this function" ; } } while (false); | |||
| 1348 | Ctx.diagnose(DiagnosticInfoPGOProfile( | |||
| 1349 | M->getName().data(), | |||
| 1350 | Twine("Inconsistent number of counts in ") + F.getName().str() | |||
| 1351 | + Twine(": the profile may be stale or there is a function name collision."), | |||
| 1352 | DS_Warning)); | |||
| 1353 | return false; | |||
| 1354 | } | |||
| 1355 | ProgramMaxCount = PGOReader->getMaximumFunctionCount(IsCS); | |||
| 1356 | return true; | |||
| 1357 | } | |||
| 1358 | ||||
| 1359 | // Populate the counters from instrumented BBs to all BBs. | |||
| 1360 | // In the end of this operation, all BBs should have a valid count value. | |||
| 1361 | void PGOUseFunc::populateCounters() { | |||
| 1362 | bool Changes = true; | |||
| 1363 | unsigned NumPasses = 0; | |||
| 1364 | while (Changes) { | |||
| 1365 | NumPasses++; | |||
| 1366 | Changes = false; | |||
| 1367 | ||||
| 1368 | // For efficient traversal, it's better to start from the end as most | |||
| 1369 | // of the instrumented edges are at the end. | |||
| 1370 | for (auto &BB : reverse(F)) { | |||
| 1371 | UseBBInfo *Count = findBBInfo(&BB); | |||
| 1372 | if (Count == nullptr) | |||
| 1373 | continue; | |||
| 1374 | if (!Count->CountValid) { | |||
| 1375 | if (Count->UnknownCountOutEdge == 0) { | |||
| 1376 | Count->CountValue = sumEdgeCount(Count->OutEdges); | |||
| 1377 | Count->CountValid = true; | |||
| 1378 | Changes = true; | |||
| 1379 | } else if (Count->UnknownCountInEdge == 0) { | |||
| 1380 | Count->CountValue = sumEdgeCount(Count->InEdges); | |||
| 1381 | Count->CountValid = true; | |||
| 1382 | Changes = true; | |||
| 1383 | } | |||
| 1384 | } | |||
| 1385 | if (Count->CountValid) { | |||
| 1386 | if (Count->UnknownCountOutEdge == 1) { | |||
| 1387 | uint64_t Total = 0; | |||
| 1388 | uint64_t OutSum = sumEdgeCount(Count->OutEdges); | |||
| 1389 | // If the one of the successor block can early terminate (no-return), | |||
| 1390 | // we can end up with situation where out edge sum count is larger as | |||
| 1391 | // the source BB's count is collected by a post-dominated block. | |||
| 1392 | if (Count->CountValue > OutSum) | |||
| 1393 | Total = Count->CountValue - OutSum; | |||
| 1394 | setEdgeCount(Count->OutEdges, Total); | |||
| 1395 | Changes = true; | |||
| 1396 | } | |||
| 1397 | if (Count->UnknownCountInEdge == 1) { | |||
| 1398 | uint64_t Total = 0; | |||
| 1399 | uint64_t InSum = sumEdgeCount(Count->InEdges); | |||
| 1400 | if (Count->CountValue > InSum) | |||
| 1401 | Total = Count->CountValue - InSum; | |||
| 1402 | setEdgeCount(Count->InEdges, Total); | |||
| 1403 | Changes = true; | |||
| 1404 | } | |||
| 1405 | } | |||
| 1406 | } | |||
| 1407 | } | |||
| 1408 | ||||
| 1409 | LLVM_DEBUG(dbgs() << "Populate counts in " << NumPasses << " passes.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Populate counts in " << NumPasses << " passes.\n"; } } while (false); | |||
| 1410 | #ifndef NDEBUG | |||
| 1411 | // Assert every BB has a valid counter. | |||
| 1412 | for (auto &BB : F) { | |||
| 1413 | auto BI = findBBInfo(&BB); | |||
| 1414 | if (BI == nullptr) | |||
| 1415 | continue; | |||
| 1416 | assert(BI->CountValid && "BB count is not valid")(static_cast <bool> (BI->CountValid && "BB count is not valid" ) ? void (0) : __assert_fail ("BI->CountValid && \"BB count is not valid\"" , "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 1416, __extension__ __PRETTY_FUNCTION__)); | |||
| 1417 | } | |||
| 1418 | #endif | |||
| 1419 | uint64_t FuncEntryCount = getBBInfo(&*F.begin()).CountValue; | |||
| 1420 | uint64_t FuncMaxCount = FuncEntryCount; | |||
| 1421 | for (auto &BB : F) { | |||
| 1422 | auto BI = findBBInfo(&BB); | |||
| 1423 | if (BI == nullptr) | |||
| 1424 | continue; | |||
| 1425 | FuncMaxCount = std::max(FuncMaxCount, BI->CountValue); | |||
| 1426 | } | |||
| 1427 | ||||
| 1428 | // Fix the obviously inconsistent entry count. | |||
| 1429 | if (FuncMaxCount > 0 && FuncEntryCount == 0) | |||
| 1430 | FuncEntryCount = 1; | |||
| 1431 | F.setEntryCount(ProfileCount(FuncEntryCount, Function::PCT_Real)); | |||
| 1432 | markFunctionAttributes(FuncEntryCount, FuncMaxCount); | |||
| 1433 | ||||
| 1434 | // Now annotate select instructions | |||
| 1435 | FuncInfo.SIVisitor.annotateSelects(F, this, &CountPosition); | |||
| 1436 | assert(CountPosition == ProfileCountSize)(static_cast <bool> (CountPosition == ProfileCountSize) ? void (0) : __assert_fail ("CountPosition == ProfileCountSize" , "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 1436, __extension__ __PRETTY_FUNCTION__)); | |||
| 1437 | ||||
| 1438 | LLVM_DEBUG(FuncInfo.dumpInfo("after reading profile."))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { FuncInfo.dumpInfo("after reading profile." ); } } while (false); | |||
| 1439 | } | |||
| 1440 | ||||
| 1441 | // Assign the scaled count values to the BB with multiple out edges. | |||
| 1442 | void PGOUseFunc::setBranchWeights() { | |||
| 1443 | // Generate MD_prof metadata for every branch instruction. | |||
| 1444 | LLVM_DEBUG(dbgs() << "\nSetting branch weights for func " << F.getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "\nSetting branch weights for func " << F.getName() << " IsCS=" << IsCS << "\n"; } } while (false) | |||
| 1445 | << " IsCS=" << IsCS << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "\nSetting branch weights for func " << F.getName() << " IsCS=" << IsCS << "\n"; } } while (false); | |||
| 1446 | for (auto &BB : F) { | |||
| 1447 | Instruction *TI = BB.getTerminator(); | |||
| 1448 | if (TI->getNumSuccessors() < 2) | |||
| 1449 | continue; | |||
| 1450 | if (!(isa<BranchInst>(TI) || isa<SwitchInst>(TI) || | |||
| 1451 | isa<IndirectBrInst>(TI) || isa<InvokeInst>(TI))) | |||
| 1452 | continue; | |||
| 1453 | ||||
| 1454 | if (getBBInfo(&BB).CountValue == 0) | |||
| 1455 | continue; | |||
| 1456 | ||||
| 1457 | // We have a non-zero Branch BB. | |||
| 1458 | const UseBBInfo &BBCountInfo = getBBInfo(&BB); | |||
| 1459 | unsigned Size = BBCountInfo.OutEdges.size(); | |||
| 1460 | SmallVector<uint64_t, 2> EdgeCounts(Size, 0); | |||
| 1461 | uint64_t MaxCount = 0; | |||
| 1462 | for (unsigned s = 0; s < Size; s++) { | |||
| 1463 | const PGOUseEdge *E = BBCountInfo.OutEdges[s]; | |||
| 1464 | const BasicBlock *SrcBB = E->SrcBB; | |||
| 1465 | const BasicBlock *DestBB = E->DestBB; | |||
| 1466 | if (DestBB == nullptr) | |||
| 1467 | continue; | |||
| 1468 | unsigned SuccNum = GetSuccessorNumber(SrcBB, DestBB); | |||
| 1469 | uint64_t EdgeCount = E->CountValue; | |||
| 1470 | if (EdgeCount > MaxCount) | |||
| 1471 | MaxCount = EdgeCount; | |||
| 1472 | EdgeCounts[SuccNum] = EdgeCount; | |||
| 1473 | } | |||
| 1474 | setProfMetadata(M, TI, EdgeCounts, MaxCount); | |||
| 1475 | } | |||
| 1476 | } | |||
| 1477 | ||||
| 1478 | static bool isIndirectBrTarget(BasicBlock *BB) { | |||
| 1479 | for (BasicBlock *Pred : predecessors(BB)) { | |||
| 1480 | if (isa<IndirectBrInst>(Pred->getTerminator())) | |||
| 1481 | return true; | |||
| 1482 | } | |||
| 1483 | return false; | |||
| 1484 | } | |||
| 1485 | ||||
| 1486 | void PGOUseFunc::annotateIrrLoopHeaderWeights() { | |||
| 1487 | LLVM_DEBUG(dbgs() << "\nAnnotating irreducible loop header weights.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "\nAnnotating irreducible loop header weights.\n" ; } } while (false); | |||
| 1488 | // Find irr loop headers | |||
| 1489 | for (auto &BB : F) { | |||
| 1490 | // As a heuristic also annotate indrectbr targets as they have a high chance | |||
| 1491 | // to become an irreducible loop header after the indirectbr tail | |||
| 1492 | // duplication. | |||
| 1493 | if (BFI->isIrrLoopHeader(&BB) || isIndirectBrTarget(&BB)) { | |||
| 1494 | Instruction *TI = BB.getTerminator(); | |||
| 1495 | const UseBBInfo &BBCountInfo = getBBInfo(&BB); | |||
| 1496 | setIrrLoopHeaderMetadata(M, TI, BBCountInfo.CountValue); | |||
| 1497 | } | |||
| 1498 | } | |||
| 1499 | } | |||
| 1500 | ||||
| 1501 | void SelectInstVisitor::instrumentOneSelectInst(SelectInst &SI) { | |||
| 1502 | Module *M = F.getParent(); | |||
| 1503 | IRBuilder<> Builder(&SI); | |||
| 1504 | Type *Int64Ty = Builder.getInt64Ty(); | |||
| 1505 | Type *I8PtrTy = Builder.getInt8PtrTy(); | |||
| 1506 | auto *Step = Builder.CreateZExt(SI.getCondition(), Int64Ty); | |||
| 1507 | Builder.CreateCall( | |||
| 1508 | Intrinsic::getDeclaration(M, Intrinsic::instrprof_increment_step), | |||
| 1509 | {ConstantExpr::getBitCast(FuncNameVar, I8PtrTy), | |||
| 1510 | Builder.getInt64(FuncHash), Builder.getInt32(TotalNumCtrs), | |||
| 1511 | Builder.getInt32(*CurCtrIdx), Step}); | |||
| 1512 | ++(*CurCtrIdx); | |||
| 1513 | } | |||
| 1514 | ||||
| 1515 | void SelectInstVisitor::annotateOneSelectInst(SelectInst &SI) { | |||
| 1516 | std::vector<uint64_t> &CountFromProfile = UseFunc->getProfileRecord().Counts; | |||
| 1517 | assert(*CurCtrIdx < CountFromProfile.size() &&(static_cast <bool> (*CurCtrIdx < CountFromProfile.size () && "Out of bound access of counters") ? void (0) : __assert_fail ("*CurCtrIdx < CountFromProfile.size() && \"Out of bound access of counters\"" , "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 1518, __extension__ __PRETTY_FUNCTION__)) | |||
| 1518 | "Out of bound access of counters")(static_cast <bool> (*CurCtrIdx < CountFromProfile.size () && "Out of bound access of counters") ? void (0) : __assert_fail ("*CurCtrIdx < CountFromProfile.size() && \"Out of bound access of counters\"" , "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 1518, __extension__ __PRETTY_FUNCTION__)); | |||
| 1519 | uint64_t SCounts[2]; | |||
| 1520 | SCounts[0] = CountFromProfile[*CurCtrIdx]; // True count | |||
| 1521 | ++(*CurCtrIdx); | |||
| 1522 | uint64_t TotalCount = 0; | |||
| 1523 | auto BI = UseFunc->findBBInfo(SI.getParent()); | |||
| 1524 | if (BI != nullptr) | |||
| 1525 | TotalCount = BI->CountValue; | |||
| 1526 | // False Count | |||
| 1527 | SCounts[1] = (TotalCount > SCounts[0] ? TotalCount - SCounts[0] : 0); | |||
| 1528 | uint64_t MaxCount = std::max(SCounts[0], SCounts[1]); | |||
| 1529 | if (MaxCount) | |||
| 1530 | setProfMetadata(F.getParent(), &SI, SCounts, MaxCount); | |||
| 1531 | } | |||
| 1532 | ||||
| 1533 | void SelectInstVisitor::visitSelectInst(SelectInst &SI) { | |||
| 1534 | if (!PGOInstrSelect) | |||
| 1535 | return; | |||
| 1536 | // FIXME: do not handle this yet. | |||
| 1537 | if (SI.getCondition()->getType()->isVectorTy()) | |||
| 1538 | return; | |||
| 1539 | ||||
| 1540 | switch (Mode) { | |||
| 1541 | case VM_counting: | |||
| 1542 | NSIs++; | |||
| 1543 | return; | |||
| 1544 | case VM_instrument: | |||
| 1545 | instrumentOneSelectInst(SI); | |||
| 1546 | return; | |||
| 1547 | case VM_annotate: | |||
| 1548 | annotateOneSelectInst(SI); | |||
| 1549 | return; | |||
| 1550 | } | |||
| 1551 | ||||
| 1552 | llvm_unreachable("Unknown visiting mode")::llvm::llvm_unreachable_internal("Unknown visiting mode", "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 1552); | |||
| 1553 | } | |||
| 1554 | ||||
| 1555 | // Traverse all valuesites and annotate the instructions for all value kind. | |||
| 1556 | void PGOUseFunc::annotateValueSites() { | |||
| 1557 | if (DisableValueProfiling) | |||
| 1558 | return; | |||
| 1559 | ||||
| 1560 | // Create the PGOFuncName meta data. | |||
| 1561 | createPGOFuncNameMetadata(F, FuncInfo.FuncName); | |||
| 1562 | ||||
| 1563 | for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) | |||
| 1564 | annotateValueSites(Kind); | |||
| 1565 | } | |||
| 1566 | ||||
| 1567 | // Annotate the instructions for a specific value kind. | |||
| 1568 | void PGOUseFunc::annotateValueSites(uint32_t Kind) { | |||
| 1569 | assert(Kind <= IPVK_Last)(static_cast <bool> (Kind <= IPVK_Last) ? void (0) : __assert_fail ("Kind <= IPVK_Last", "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 1569, __extension__ __PRETTY_FUNCTION__)); | |||
| 1570 | unsigned ValueSiteIndex = 0; | |||
| 1571 | auto &ValueSites = FuncInfo.ValueSites[Kind]; | |||
| 1572 | unsigned NumValueSites = ProfileRecord.getNumValueSites(Kind); | |||
| 1573 | if (NumValueSites != ValueSites.size()) { | |||
| 1574 | auto &Ctx = M->getContext(); | |||
| 1575 | Ctx.diagnose(DiagnosticInfoPGOProfile( | |||
| 1576 | M->getName().data(), | |||
| 1577 | Twine("Inconsistent number of value sites for ") + | |||
| 1578 | Twine(ValueProfKindDescr[Kind]) + | |||
| 1579 | Twine(" profiling in \"") + F.getName().str() + | |||
| 1580 | Twine("\", possibly due to the use of a stale profile."), | |||
| 1581 | DS_Warning)); | |||
| 1582 | return; | |||
| 1583 | } | |||
| 1584 | ||||
| 1585 | for (VPCandidateInfo &I : ValueSites) { | |||
| 1586 | LLVM_DEBUG(dbgs() << "Read one value site profile (kind = " << Kinddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Read one value site profile (kind = " << Kind << "): Index = " << ValueSiteIndex << " out of " << NumValueSites << "\n"; } } while (false) | |||
| 1587 | << "): Index = " << ValueSiteIndex << " out of "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Read one value site profile (kind = " << Kind << "): Index = " << ValueSiteIndex << " out of " << NumValueSites << "\n"; } } while (false) | |||
| 1588 | << NumValueSites << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Read one value site profile (kind = " << Kind << "): Index = " << ValueSiteIndex << " out of " << NumValueSites << "\n"; } } while (false); | |||
| 1589 | annotateValueSite(*M, *I.AnnotatedInst, ProfileRecord, | |||
| 1590 | static_cast<InstrProfValueKind>(Kind), ValueSiteIndex, | |||
| 1591 | Kind == IPVK_MemOPSize ? MaxNumMemOPAnnotations | |||
| 1592 | : MaxNumAnnotations); | |||
| 1593 | ValueSiteIndex++; | |||
| 1594 | } | |||
| 1595 | } | |||
| 1596 | ||||
| 1597 | // Collect the set of members for each Comdat in module M and store | |||
| 1598 | // in ComdatMembers. | |||
| 1599 | static void collectComdatMembers( | |||
| 1600 | Module &M, | |||
| 1601 | std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers) { | |||
| 1602 | if (!DoComdatRenaming) | |||
| 1603 | return; | |||
| 1604 | for (Function &F : M) | |||
| 1605 | if (Comdat *C = F.getComdat()) | |||
| 1606 | ComdatMembers.insert(std::make_pair(C, &F)); | |||
| 1607 | for (GlobalVariable &GV : M.globals()) | |||
| 1608 | if (Comdat *C = GV.getComdat()) | |||
| 1609 | ComdatMembers.insert(std::make_pair(C, &GV)); | |||
| 1610 | for (GlobalAlias &GA : M.aliases()) | |||
| 1611 | if (Comdat *C = GA.getComdat()) | |||
| 1612 | ComdatMembers.insert(std::make_pair(C, &GA)); | |||
| 1613 | } | |||
| 1614 | ||||
| 1615 | static bool InstrumentAllFunctions( | |||
| 1616 | Module &M, function_ref<TargetLibraryInfo &(Function &)> LookupTLI, | |||
| 1617 | function_ref<BranchProbabilityInfo *(Function &)> LookupBPI, | |||
| 1618 | function_ref<BlockFrequencyInfo *(Function &)> LookupBFI, bool IsCS) { | |||
| 1619 | // For the context-sensitve instrumentation, we should have a separated pass | |||
| 1620 | // (before LTO/ThinLTO linking) to create these variables. | |||
| 1621 | if (!IsCS) | |||
| 1622 | createIRLevelProfileFlagVar(M, /*IsCS=*/false, PGOInstrumentEntry, | |||
| 1623 | DebugInfoCorrelate); | |||
| 1624 | std::unordered_multimap<Comdat *, GlobalValue *> ComdatMembers; | |||
| 1625 | collectComdatMembers(M, ComdatMembers); | |||
| 1626 | ||||
| 1627 | for (auto &F : M) { | |||
| 1628 | if (F.isDeclaration()) | |||
| 1629 | continue; | |||
| 1630 | if (F.hasFnAttribute(llvm::Attribute::NoProfile)) | |||
| 1631 | continue; | |||
| 1632 | auto &TLI = LookupTLI(F); | |||
| 1633 | auto *BPI = LookupBPI(F); | |||
| 1634 | auto *BFI = LookupBFI(F); | |||
| 1635 | instrumentOneFunc(F, &M, TLI, BPI, BFI, ComdatMembers, IsCS); | |||
| 1636 | } | |||
| 1637 | return true; | |||
| 1638 | } | |||
| 1639 | ||||
| 1640 | PreservedAnalyses | |||
| 1641 | PGOInstrumentationGenCreateVar::run(Module &M, ModuleAnalysisManager &AM) { | |||
| 1642 | createProfileFileNameVar(M, CSInstrName); | |||
| 1643 | // The variable in a comdat may be discarded by LTO. Ensure the declaration | |||
| 1644 | // will be retained. | |||
| 1645 | appendToCompilerUsed(M, createIRLevelProfileFlagVar(M, /*IsCS=*/true, | |||
| 1646 | PGOInstrumentEntry, | |||
| 1647 | DebugInfoCorrelate)); | |||
| 1648 | return PreservedAnalyses::all(); | |||
| 1649 | } | |||
| 1650 | ||||
| 1651 | bool PGOInstrumentationGenLegacyPass::runOnModule(Module &M) { | |||
| 1652 | if (skipModule(M)) | |||
| 1653 | return false; | |||
| 1654 | ||||
| 1655 | auto LookupTLI = [this](Function &F) -> TargetLibraryInfo & { | |||
| 1656 | return this->getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); | |||
| 1657 | }; | |||
| 1658 | auto LookupBPI = [this](Function &F) { | |||
| 1659 | return &this->getAnalysis<BranchProbabilityInfoWrapperPass>(F).getBPI(); | |||
| 1660 | }; | |||
| 1661 | auto LookupBFI = [this](Function &F) { | |||
| 1662 | return &this->getAnalysis<BlockFrequencyInfoWrapperPass>(F).getBFI(); | |||
| 1663 | }; | |||
| 1664 | return InstrumentAllFunctions(M, LookupTLI, LookupBPI, LookupBFI, IsCS); | |||
| 1665 | } | |||
| 1666 | ||||
| 1667 | PreservedAnalyses PGOInstrumentationGen::run(Module &M, | |||
| 1668 | ModuleAnalysisManager &AM) { | |||
| 1669 | auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); | |||
| 1670 | auto LookupTLI = [&FAM](Function &F) -> TargetLibraryInfo & { | |||
| 1671 | return FAM.getResult<TargetLibraryAnalysis>(F); | |||
| 1672 | }; | |||
| 1673 | auto LookupBPI = [&FAM](Function &F) { | |||
| 1674 | return &FAM.getResult<BranchProbabilityAnalysis>(F); | |||
| 1675 | }; | |||
| 1676 | auto LookupBFI = [&FAM](Function &F) { | |||
| 1677 | return &FAM.getResult<BlockFrequencyAnalysis>(F); | |||
| 1678 | }; | |||
| 1679 | ||||
| 1680 | if (!InstrumentAllFunctions(M, LookupTLI, LookupBPI, LookupBFI, IsCS)) | |||
| ||||
| 1681 | return PreservedAnalyses::all(); | |||
| 1682 | ||||
| 1683 | return PreservedAnalyses::none(); | |||
| 1684 | } | |||
| 1685 | ||||
| 1686 | // Using the ratio b/w sums of profile count values and BFI count values to | |||
| 1687 | // adjust the func entry count. | |||
| 1688 | static void fixFuncEntryCount(PGOUseFunc &Func, LoopInfo &LI, | |||
| 1689 | BranchProbabilityInfo &NBPI) { | |||
| 1690 | Function &F = Func.getFunc(); | |||
| 1691 | BlockFrequencyInfo NBFI(F, NBPI, LI); | |||
| 1692 | #ifndef NDEBUG | |||
| 1693 | auto BFIEntryCount = F.getEntryCount(); | |||
| 1694 | assert(BFIEntryCount.hasValue() && (BFIEntryCount->getCount() > 0) &&(static_cast <bool> (BFIEntryCount.hasValue() && (BFIEntryCount->getCount() > 0) && "Invalid BFI Entrycount" ) ? void (0) : __assert_fail ("BFIEntryCount.hasValue() && (BFIEntryCount->getCount() > 0) && \"Invalid BFI Entrycount\"" , "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 1695, __extension__ __PRETTY_FUNCTION__)) | |||
| 1695 | "Invalid BFI Entrycount")(static_cast <bool> (BFIEntryCount.hasValue() && (BFIEntryCount->getCount() > 0) && "Invalid BFI Entrycount" ) ? void (0) : __assert_fail ("BFIEntryCount.hasValue() && (BFIEntryCount->getCount() > 0) && \"Invalid BFI Entrycount\"" , "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 1695, __extension__ __PRETTY_FUNCTION__)); | |||
| 1696 | #endif | |||
| 1697 | auto SumCount = APFloat::getZero(APFloat::IEEEdouble()); | |||
| 1698 | auto SumBFICount = APFloat::getZero(APFloat::IEEEdouble()); | |||
| 1699 | for (auto &BBI : F) { | |||
| 1700 | uint64_t CountValue = 0; | |||
| 1701 | uint64_t BFICountValue = 0; | |||
| 1702 | if (!Func.findBBInfo(&BBI)) | |||
| 1703 | continue; | |||
| 1704 | auto BFICount = NBFI.getBlockProfileCount(&BBI); | |||
| 1705 | CountValue = Func.getBBInfo(&BBI).CountValue; | |||
| 1706 | BFICountValue = BFICount.getValue(); | |||
| 1707 | SumCount.add(APFloat(CountValue * 1.0), APFloat::rmNearestTiesToEven); | |||
| 1708 | SumBFICount.add(APFloat(BFICountValue * 1.0), APFloat::rmNearestTiesToEven); | |||
| 1709 | } | |||
| 1710 | if (SumCount.isZero()) | |||
| 1711 | return; | |||
| 1712 | ||||
| 1713 | assert(SumBFICount.compare(APFloat(0.0)) == APFloat::cmpGreaterThan &&(static_cast <bool> (SumBFICount.compare(APFloat(0.0)) == APFloat::cmpGreaterThan && "Incorrect sum of BFI counts" ) ? void (0) : __assert_fail ("SumBFICount.compare(APFloat(0.0)) == APFloat::cmpGreaterThan && \"Incorrect sum of BFI counts\"" , "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 1714, __extension__ __PRETTY_FUNCTION__)) | |||
| 1714 | "Incorrect sum of BFI counts")(static_cast <bool> (SumBFICount.compare(APFloat(0.0)) == APFloat::cmpGreaterThan && "Incorrect sum of BFI counts" ) ? void (0) : __assert_fail ("SumBFICount.compare(APFloat(0.0)) == APFloat::cmpGreaterThan && \"Incorrect sum of BFI counts\"" , "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 1714, __extension__ __PRETTY_FUNCTION__)); | |||
| 1715 | if (SumBFICount.compare(SumCount) == APFloat::cmpEqual) | |||
| 1716 | return; | |||
| 1717 | double Scale = (SumCount / SumBFICount).convertToDouble(); | |||
| 1718 | if (Scale < 1.001 && Scale > 0.999) | |||
| 1719 | return; | |||
| 1720 | ||||
| 1721 | uint64_t FuncEntryCount = Func.getBBInfo(&*F.begin()).CountValue; | |||
| 1722 | uint64_t NewEntryCount = 0.5 + FuncEntryCount * Scale; | |||
| 1723 | if (NewEntryCount == 0) | |||
| 1724 | NewEntryCount = 1; | |||
| 1725 | if (NewEntryCount != FuncEntryCount) { | |||
| 1726 | F.setEntryCount(ProfileCount(NewEntryCount, Function::PCT_Real)); | |||
| 1727 | LLVM_DEBUG(dbgs() << "FixFuncEntryCount: in " << F.getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "FixFuncEntryCount: in " << F.getName() << ", entry_count " << FuncEntryCount << " --> " << NewEntryCount << "\n"; } } while (false) | |||
| 1728 | << ", entry_count " << FuncEntryCount << " --> "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "FixFuncEntryCount: in " << F.getName() << ", entry_count " << FuncEntryCount << " --> " << NewEntryCount << "\n"; } } while (false) | |||
| 1729 | << NewEntryCount << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "FixFuncEntryCount: in " << F.getName() << ", entry_count " << FuncEntryCount << " --> " << NewEntryCount << "\n"; } } while (false); | |||
| 1730 | } | |||
| 1731 | } | |||
| 1732 | ||||
| 1733 | // Compare the profile count values with BFI count values, and print out | |||
| 1734 | // the non-matching ones. | |||
| 1735 | static void verifyFuncBFI(PGOUseFunc &Func, LoopInfo &LI, | |||
| 1736 | BranchProbabilityInfo &NBPI, | |||
| 1737 | uint64_t HotCountThreshold, | |||
| 1738 | uint64_t ColdCountThreshold) { | |||
| 1739 | Function &F = Func.getFunc(); | |||
| 1740 | BlockFrequencyInfo NBFI(F, NBPI, LI); | |||
| 1741 | // bool PrintFunc = false; | |||
| 1742 | bool HotBBOnly = PGOVerifyHotBFI; | |||
| 1743 | std::string Msg; | |||
| 1744 | OptimizationRemarkEmitter ORE(&F); | |||
| 1745 | ||||
| 1746 | unsigned BBNum = 0, BBMisMatchNum = 0, NonZeroBBNum = 0; | |||
| 1747 | for (auto &BBI : F) { | |||
| 1748 | uint64_t CountValue = 0; | |||
| 1749 | uint64_t BFICountValue = 0; | |||
| 1750 | ||||
| 1751 | if (Func.getBBInfo(&BBI).CountValid) | |||
| 1752 | CountValue = Func.getBBInfo(&BBI).CountValue; | |||
| 1753 | ||||
| 1754 | BBNum++; | |||
| 1755 | if (CountValue) | |||
| 1756 | NonZeroBBNum++; | |||
| 1757 | auto BFICount = NBFI.getBlockProfileCount(&BBI); | |||
| 1758 | if (BFICount) | |||
| 1759 | BFICountValue = BFICount.getValue(); | |||
| 1760 | ||||
| 1761 | if (HotBBOnly) { | |||
| 1762 | bool rawIsHot = CountValue >= HotCountThreshold; | |||
| 1763 | bool BFIIsHot = BFICountValue >= HotCountThreshold; | |||
| 1764 | bool rawIsCold = CountValue <= ColdCountThreshold; | |||
| 1765 | bool ShowCount = false; | |||
| 1766 | if (rawIsHot && !BFIIsHot) { | |||
| 1767 | Msg = "raw-Hot to BFI-nonHot"; | |||
| 1768 | ShowCount = true; | |||
| 1769 | } else if (rawIsCold && BFIIsHot) { | |||
| 1770 | Msg = "raw-Cold to BFI-Hot"; | |||
| 1771 | ShowCount = true; | |||
| 1772 | } | |||
| 1773 | if (!ShowCount) | |||
| 1774 | continue; | |||
| 1775 | } else { | |||
| 1776 | if ((CountValue < PGOVerifyBFICutoff) && | |||
| 1777 | (BFICountValue < PGOVerifyBFICutoff)) | |||
| 1778 | continue; | |||
| 1779 | uint64_t Diff = (BFICountValue >= CountValue) | |||
| 1780 | ? BFICountValue - CountValue | |||
| 1781 | : CountValue - BFICountValue; | |||
| 1782 | if (Diff <= CountValue / 100 * PGOVerifyBFIRatio) | |||
| 1783 | continue; | |||
| 1784 | } | |||
| 1785 | BBMisMatchNum++; | |||
| 1786 | ||||
| 1787 | ORE.emit([&]() { | |||
| 1788 | OptimizationRemarkAnalysis Remark(DEBUG_TYPE"pgo-instrumentation", "bfi-verify", | |||
| 1789 | F.getSubprogram(), &BBI); | |||
| 1790 | Remark << "BB " << ore::NV("Block", BBI.getName()) | |||
| 1791 | << " Count=" << ore::NV("Count", CountValue) | |||
| 1792 | << " BFI_Count=" << ore::NV("Count", BFICountValue); | |||
| 1793 | if (!Msg.empty()) | |||
| 1794 | Remark << " (" << Msg << ")"; | |||
| 1795 | return Remark; | |||
| 1796 | }); | |||
| 1797 | } | |||
| 1798 | if (BBMisMatchNum) | |||
| 1799 | ORE.emit([&]() { | |||
| 1800 | return OptimizationRemarkAnalysis(DEBUG_TYPE"pgo-instrumentation", "bfi-verify", | |||
| 1801 | F.getSubprogram(), &F.getEntryBlock()) | |||
| 1802 | << "In Func " << ore::NV("Function", F.getName()) | |||
| 1803 | << ": Num_of_BB=" << ore::NV("Count", BBNum) | |||
| 1804 | << ", Num_of_non_zerovalue_BB=" << ore::NV("Count", NonZeroBBNum) | |||
| 1805 | << ", Num_of_mis_matching_BB=" << ore::NV("Count", BBMisMatchNum); | |||
| 1806 | }); | |||
| 1807 | } | |||
| 1808 | ||||
| 1809 | static bool annotateAllFunctions( | |||
| 1810 | Module &M, StringRef ProfileFileName, StringRef ProfileRemappingFileName, | |||
| 1811 | function_ref<TargetLibraryInfo &(Function &)> LookupTLI, | |||
| 1812 | function_ref<BranchProbabilityInfo *(Function &)> LookupBPI, | |||
| 1813 | function_ref<BlockFrequencyInfo *(Function &)> LookupBFI, | |||
| 1814 | ProfileSummaryInfo *PSI, bool IsCS) { | |||
| 1815 | LLVM_DEBUG(dbgs() << "Read in profile counters: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Read in profile counters: " ; } } while (false); | |||
| 1816 | auto &Ctx = M.getContext(); | |||
| 1817 | // Read the counter array from file. | |||
| 1818 | auto ReaderOrErr = | |||
| 1819 | IndexedInstrProfReader::create(ProfileFileName, ProfileRemappingFileName); | |||
| 1820 | if (Error E = ReaderOrErr.takeError()) { | |||
| 1821 | handleAllErrors(std::move(E), [&](const ErrorInfoBase &EI) { | |||
| 1822 | Ctx.diagnose( | |||
| 1823 | DiagnosticInfoPGOProfile(ProfileFileName.data(), EI.message())); | |||
| 1824 | }); | |||
| 1825 | return false; | |||
| 1826 | } | |||
| 1827 | ||||
| 1828 | std::unique_ptr<IndexedInstrProfReader> PGOReader = | |||
| 1829 | std::move(ReaderOrErr.get()); | |||
| 1830 | if (!PGOReader) { | |||
| 1831 | Ctx.diagnose(DiagnosticInfoPGOProfile(ProfileFileName.data(), | |||
| 1832 | StringRef("Cannot get PGOReader"))); | |||
| 1833 | return false; | |||
| 1834 | } | |||
| 1835 | if (!PGOReader->hasCSIRLevelProfile() && IsCS) | |||
| 1836 | return false; | |||
| 1837 | ||||
| 1838 | // TODO: might need to change the warning once the clang option is finalized. | |||
| 1839 | if (!PGOReader->isIRLevelProfile()) { | |||
| 1840 | Ctx.diagnose(DiagnosticInfoPGOProfile( | |||
| 1841 | ProfileFileName.data(), "Not an IR level instrumentation profile")); | |||
| 1842 | return false; | |||
| 1843 | } | |||
| 1844 | ||||
| 1845 | // Add the profile summary (read from the header of the indexed summary) here | |||
| 1846 | // so that we can use it below when reading counters (which checks if the | |||
| 1847 | // function should be marked with a cold or inlinehint attribute). | |||
| 1848 | M.setProfileSummary(PGOReader->getSummary(IsCS).getMD(M.getContext()), | |||
| 1849 | IsCS ? ProfileSummary::PSK_CSInstr | |||
| 1850 | : ProfileSummary::PSK_Instr); | |||
| 1851 | PSI->refresh(); | |||
| 1852 | ||||
| 1853 | std::unordered_multimap<Comdat *, GlobalValue *> ComdatMembers; | |||
| 1854 | collectComdatMembers(M, ComdatMembers); | |||
| 1855 | std::vector<Function *> HotFunctions; | |||
| 1856 | std::vector<Function *> ColdFunctions; | |||
| 1857 | ||||
| 1858 | // If the profile marked as always instrument the entry BB, do the | |||
| 1859 | // same. Note this can be overwritten by the internal option in CFGMST.h | |||
| 1860 | bool InstrumentFuncEntry = PGOReader->instrEntryBBEnabled(); | |||
| 1861 | if (PGOInstrumentEntry.getNumOccurrences() > 0) | |||
| 1862 | InstrumentFuncEntry = PGOInstrumentEntry; | |||
| 1863 | for (auto &F : M) { | |||
| 1864 | if (F.isDeclaration()) | |||
| 1865 | continue; | |||
| 1866 | auto &TLI = LookupTLI(F); | |||
| 1867 | auto *BPI = LookupBPI(F); | |||
| 1868 | auto *BFI = LookupBFI(F); | |||
| 1869 | // Split indirectbr critical edges here before computing the MST rather than | |||
| 1870 | // later in getInstrBB() to avoid invalidating it. | |||
| 1871 | SplitIndirectBrCriticalEdges(F, BPI, BFI); | |||
| 1872 | PGOUseFunc Func(F, &M, TLI, ComdatMembers, BPI, BFI, PSI, IsCS, | |||
| 1873 | InstrumentFuncEntry); | |||
| 1874 | // When AllMinusOnes is true, it means the profile for the function | |||
| 1875 | // is unrepresentative and this function is actually hot. Set the | |||
| 1876 | // entry count of the function to be multiple times of hot threshold | |||
| 1877 | // and drop all its internal counters. | |||
| 1878 | bool AllMinusOnes = false; | |||
| 1879 | bool AllZeros = false; | |||
| 1880 | if (!Func.readCounters(PGOReader.get(), AllZeros, AllMinusOnes)) | |||
| 1881 | continue; | |||
| 1882 | if (AllZeros) { | |||
| 1883 | F.setEntryCount(ProfileCount(0, Function::PCT_Real)); | |||
| 1884 | if (Func.getProgramMaxCount() != 0) | |||
| 1885 | ColdFunctions.push_back(&F); | |||
| 1886 | continue; | |||
| 1887 | } | |||
| 1888 | const unsigned MultiplyFactor = 3; | |||
| 1889 | if (AllMinusOnes) { | |||
| 1890 | uint64_t HotThreshold = PSI->getHotCountThreshold(); | |||
| 1891 | if (HotThreshold) | |||
| 1892 | F.setEntryCount( | |||
| 1893 | ProfileCount(HotThreshold * MultiplyFactor, Function::PCT_Real)); | |||
| 1894 | HotFunctions.push_back(&F); | |||
| 1895 | continue; | |||
| 1896 | } | |||
| 1897 | Func.populateCounters(); | |||
| 1898 | Func.setBranchWeights(); | |||
| 1899 | Func.annotateValueSites(); | |||
| 1900 | Func.annotateIrrLoopHeaderWeights(); | |||
| 1901 | PGOUseFunc::FuncFreqAttr FreqAttr = Func.getFuncFreqAttr(); | |||
| 1902 | if (FreqAttr == PGOUseFunc::FFA_Cold) | |||
| 1903 | ColdFunctions.push_back(&F); | |||
| 1904 | else if (FreqAttr == PGOUseFunc::FFA_Hot) | |||
| 1905 | HotFunctions.push_back(&F); | |||
| 1906 | if (PGOViewCounts != PGOVCT_None && | |||
| 1907 | (ViewBlockFreqFuncName.empty() || | |||
| 1908 | F.getName().equals(ViewBlockFreqFuncName))) { | |||
| 1909 | LoopInfo LI{DominatorTree(F)}; | |||
| 1910 | std::unique_ptr<BranchProbabilityInfo> NewBPI = | |||
| 1911 | std::make_unique<BranchProbabilityInfo>(F, LI); | |||
| 1912 | std::unique_ptr<BlockFrequencyInfo> NewBFI = | |||
| 1913 | std::make_unique<BlockFrequencyInfo>(F, *NewBPI, LI); | |||
| 1914 | if (PGOViewCounts == PGOVCT_Graph) | |||
| 1915 | NewBFI->view(); | |||
| 1916 | else if (PGOViewCounts == PGOVCT_Text) { | |||
| 1917 | dbgs() << "pgo-view-counts: " << Func.getFunc().getName() << "\n"; | |||
| 1918 | NewBFI->print(dbgs()); | |||
| 1919 | } | |||
| 1920 | } | |||
| 1921 | if (PGOViewRawCounts != PGOVCT_None && | |||
| 1922 | (ViewBlockFreqFuncName.empty() || | |||
| 1923 | F.getName().equals(ViewBlockFreqFuncName))) { | |||
| 1924 | if (PGOViewRawCounts == PGOVCT_Graph) | |||
| 1925 | if (ViewBlockFreqFuncName.empty()) | |||
| 1926 | WriteGraph(&Func, Twine("PGORawCounts_") + Func.getFunc().getName()); | |||
| 1927 | else | |||
| 1928 | ViewGraph(&Func, Twine("PGORawCounts_") + Func.getFunc().getName()); | |||
| 1929 | else if (PGOViewRawCounts == PGOVCT_Text) { | |||
| 1930 | dbgs() << "pgo-view-raw-counts: " << Func.getFunc().getName() << "\n"; | |||
| 1931 | Func.dumpInfo(); | |||
| 1932 | } | |||
| 1933 | } | |||
| 1934 | ||||
| 1935 | if (PGOVerifyBFI || PGOVerifyHotBFI || PGOFixEntryCount) { | |||
| 1936 | LoopInfo LI{DominatorTree(F)}; | |||
| 1937 | BranchProbabilityInfo NBPI(F, LI); | |||
| 1938 | ||||
| 1939 | // Fix func entry count. | |||
| 1940 | if (PGOFixEntryCount) | |||
| 1941 | fixFuncEntryCount(Func, LI, NBPI); | |||
| 1942 | ||||
| 1943 | // Verify BlockFrequency information. | |||
| 1944 | uint64_t HotCountThreshold = 0, ColdCountThreshold = 0; | |||
| 1945 | if (PGOVerifyHotBFI) { | |||
| 1946 | HotCountThreshold = PSI->getOrCompHotCountThreshold(); | |||
| 1947 | ColdCountThreshold = PSI->getOrCompColdCountThreshold(); | |||
| 1948 | } | |||
| 1949 | verifyFuncBFI(Func, LI, NBPI, HotCountThreshold, ColdCountThreshold); | |||
| 1950 | } | |||
| 1951 | } | |||
| 1952 | ||||
| 1953 | // Set function hotness attribute from the profile. | |||
| 1954 | // We have to apply these attributes at the end because their presence | |||
| 1955 | // can affect the BranchProbabilityInfo of any callers, resulting in an | |||
| 1956 | // inconsistent MST between prof-gen and prof-use. | |||
| 1957 | for (auto &F : HotFunctions) { | |||
| 1958 | F->addFnAttr(Attribute::InlineHint); | |||
| 1959 | LLVM_DEBUG(dbgs() << "Set inline attribute to function: " << F->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Set inline attribute to function: " << F->getName() << "\n"; } } while (false) | |||
| 1960 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Set inline attribute to function: " << F->getName() << "\n"; } } while (false); | |||
| 1961 | } | |||
| 1962 | for (auto &F : ColdFunctions) { | |||
| 1963 | // Only set when there is no Attribute::Hot set by the user. For Hot | |||
| 1964 | // attribute, user's annotation has the precedence over the profile. | |||
| 1965 | if (F->hasFnAttribute(Attribute::Hot)) { | |||
| 1966 | auto &Ctx = M.getContext(); | |||
| 1967 | std::string Msg = std::string("Function ") + F->getName().str() + | |||
| 1968 | std::string(" is annotated as a hot function but" | |||
| 1969 | " the profile is cold"); | |||
| 1970 | Ctx.diagnose( | |||
| 1971 | DiagnosticInfoPGOProfile(M.getName().data(), Msg, DS_Warning)); | |||
| 1972 | continue; | |||
| 1973 | } | |||
| 1974 | F->addFnAttr(Attribute::Cold); | |||
| 1975 | LLVM_DEBUG(dbgs() << "Set cold attribute to function: " << F->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Set cold attribute to function: " << F->getName() << "\n"; } } while (false) | |||
| 1976 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Set cold attribute to function: " << F->getName() << "\n"; } } while (false); | |||
| 1977 | } | |||
| 1978 | return true; | |||
| 1979 | } | |||
| 1980 | ||||
| 1981 | PGOInstrumentationUse::PGOInstrumentationUse(std::string Filename, | |||
| 1982 | std::string RemappingFilename, | |||
| 1983 | bool IsCS) | |||
| 1984 | : ProfileFileName(std::move(Filename)), | |||
| 1985 | ProfileRemappingFileName(std::move(RemappingFilename)), IsCS(IsCS) { | |||
| 1986 | if (!PGOTestProfileFile.empty()) | |||
| 1987 | ProfileFileName = PGOTestProfileFile; | |||
| 1988 | if (!PGOTestProfileRemappingFile.empty()) | |||
| 1989 | ProfileRemappingFileName = PGOTestProfileRemappingFile; | |||
| 1990 | } | |||
| 1991 | ||||
| 1992 | PreservedAnalyses PGOInstrumentationUse::run(Module &M, | |||
| 1993 | ModuleAnalysisManager &AM) { | |||
| 1994 | ||||
| 1995 | auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); | |||
| 1996 | auto LookupTLI = [&FAM](Function &F) -> TargetLibraryInfo & { | |||
| 1997 | return FAM.getResult<TargetLibraryAnalysis>(F); | |||
| 1998 | }; | |||
| 1999 | auto LookupBPI = [&FAM](Function &F) { | |||
| 2000 | return &FAM.getResult<BranchProbabilityAnalysis>(F); | |||
| 2001 | }; | |||
| 2002 | auto LookupBFI = [&FAM](Function &F) { | |||
| 2003 | return &FAM.getResult<BlockFrequencyAnalysis>(F); | |||
| 2004 | }; | |||
| 2005 | ||||
| 2006 | auto *PSI = &AM.getResult<ProfileSummaryAnalysis>(M); | |||
| 2007 | ||||
| 2008 | if (!annotateAllFunctions(M, ProfileFileName, ProfileRemappingFileName, | |||
| 2009 | LookupTLI, LookupBPI, LookupBFI, PSI, IsCS)) | |||
| 2010 | return PreservedAnalyses::all(); | |||
| 2011 | ||||
| 2012 | return PreservedAnalyses::none(); | |||
| 2013 | } | |||
| 2014 | ||||
| 2015 | bool PGOInstrumentationUseLegacyPass::runOnModule(Module &M) { | |||
| 2016 | if (skipModule(M)) | |||
| 2017 | return false; | |||
| 2018 | ||||
| 2019 | auto LookupTLI = [this](Function &F) -> TargetLibraryInfo & { | |||
| 2020 | return this->getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); | |||
| 2021 | }; | |||
| 2022 | auto LookupBPI = [this](Function &F) { | |||
| 2023 | return &this->getAnalysis<BranchProbabilityInfoWrapperPass>(F).getBPI(); | |||
| 2024 | }; | |||
| 2025 | auto LookupBFI = [this](Function &F) { | |||
| 2026 | return &this->getAnalysis<BlockFrequencyInfoWrapperPass>(F).getBFI(); | |||
| 2027 | }; | |||
| 2028 | ||||
| 2029 | auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); | |||
| 2030 | return annotateAllFunctions(M, ProfileFileName, "", LookupTLI, LookupBPI, | |||
| 2031 | LookupBFI, PSI, IsCS); | |||
| 2032 | } | |||
| 2033 | ||||
| 2034 | static std::string getSimpleNodeName(const BasicBlock *Node) { | |||
| 2035 | if (!Node->getName().empty()) | |||
| 2036 | return std::string(Node->getName()); | |||
| 2037 | ||||
| 2038 | std::string SimpleNodeName; | |||
| 2039 | raw_string_ostream OS(SimpleNodeName); | |||
| 2040 | Node->printAsOperand(OS, false); | |||
| 2041 | return OS.str(); | |||
| 2042 | } | |||
| 2043 | ||||
| 2044 | void llvm::setProfMetadata(Module *M, Instruction *TI, | |||
| 2045 | ArrayRef<uint64_t> EdgeCounts, | |||
| 2046 | uint64_t MaxCount) { | |||
| 2047 | MDBuilder MDB(M->getContext()); | |||
| 2048 | assert(MaxCount > 0 && "Bad max count")(static_cast <bool> (MaxCount > 0 && "Bad max count" ) ? void (0) : __assert_fail ("MaxCount > 0 && \"Bad max count\"" , "llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp" , 2048, __extension__ __PRETTY_FUNCTION__)); | |||
| 2049 | uint64_t Scale = calculateCountScale(MaxCount); | |||
| 2050 | SmallVector<unsigned, 4> Weights; | |||
| 2051 | for (const auto &ECI : EdgeCounts) | |||
| 2052 | Weights.push_back(scaleBranchCount(ECI, Scale)); | |||
| 2053 | ||||
| 2054 | LLVM_DEBUG(dbgs() << "Weight is: "; for (const auto &Wdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Weight is: "; for (const auto &W : Weights) { dbgs() << W << " " ; } dbgs() << "\n";; } } while (false) | |||
| 2055 | : Weights) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Weight is: "; for (const auto &W : Weights) { dbgs() << W << " " ; } dbgs() << "\n";; } } while (false) | |||
| 2056 | dbgs() << W << " ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Weight is: "; for (const auto &W : Weights) { dbgs() << W << " " ; } dbgs() << "\n";; } } while (false) | |||
| 2057 | } dbgs() << "\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("pgo-instrumentation")) { dbgs() << "Weight is: "; for (const auto &W : Weights) { dbgs() << W << " " ; } dbgs() << "\n";; } } while (false); | |||
| 2058 | ||||
| 2059 | TI->setMetadata(LLVMContext::MD_prof, MDB.createBranchWeights(Weights)); | |||
| 2060 | if (EmitBranchProbability) { | |||
| 2061 | std::string BrCondStr = getBranchCondString(TI); | |||
| 2062 | if (BrCondStr.empty()) | |||
| 2063 | return; | |||
| 2064 | ||||
| 2065 | uint64_t WSum = | |||
| 2066 | std::accumulate(Weights.begin(), Weights.end(), (uint64_t)0, | |||
| 2067 | [](uint64_t w1, uint64_t w2) { return w1 + w2; }); | |||
| 2068 | uint64_t TotalCount = | |||
| 2069 | std::accumulate(EdgeCounts.begin(), EdgeCounts.end(), (uint64_t)0, | |||
| 2070 | [](uint64_t c1, uint64_t c2) { return c1 + c2; }); | |||
| 2071 | Scale = calculateCountScale(WSum); | |||
| 2072 | BranchProbability BP(scaleBranchCount(Weights[0], Scale), | |||
| 2073 | scaleBranchCount(WSum, Scale)); | |||
| 2074 | std::string BranchProbStr; | |||
| 2075 | raw_string_ostream OS(BranchProbStr); | |||
| 2076 | OS << BP; | |||
| 2077 | OS << " (total count : " << TotalCount << ")"; | |||
| 2078 | OS.flush(); | |||
| 2079 | Function *F = TI->getParent()->getParent(); | |||
| 2080 | OptimizationRemarkEmitter ORE(F); | |||
| 2081 | ORE.emit([&]() { | |||
| 2082 | return OptimizationRemark(DEBUG_TYPE"pgo-instrumentation", "pgo-instrumentation", TI) | |||
| 2083 | << BrCondStr << " is true with probability : " << BranchProbStr; | |||
| 2084 | }); | |||
| 2085 | } | |||
| 2086 | } | |||
| 2087 | ||||
| 2088 | namespace llvm { | |||
| 2089 | ||||
| 2090 | void setIrrLoopHeaderMetadata(Module *M, Instruction *TI, uint64_t Count) { | |||
| 2091 | MDBuilder MDB(M->getContext()); | |||
| 2092 | TI->setMetadata(llvm::LLVMContext::MD_irr_loop, | |||
| 2093 | MDB.createIrrLoopHeaderWeight(Count)); | |||
| 2094 | } | |||
| 2095 | ||||
| 2096 | template <> struct GraphTraits<PGOUseFunc *> { | |||
| 2097 | using NodeRef = const BasicBlock *; | |||
| 2098 | using ChildIteratorType = const_succ_iterator; | |||
| 2099 | using nodes_iterator = pointer_iterator<Function::const_iterator>; | |||
| 2100 | ||||
| 2101 | static NodeRef getEntryNode(const PGOUseFunc *G) { | |||
| 2102 | return &G->getFunc().front(); | |||
| 2103 | } | |||
| 2104 | ||||
| 2105 | static ChildIteratorType child_begin(const NodeRef N) { | |||
| 2106 | return succ_begin(N); | |||
| 2107 | } | |||
| 2108 | ||||
| 2109 | static ChildIteratorType child_end(const NodeRef N) { return succ_end(N); } | |||
| 2110 | ||||
| 2111 | static nodes_iterator nodes_begin(const PGOUseFunc *G) { | |||
| 2112 | return nodes_iterator(G->getFunc().begin()); | |||
| 2113 | } | |||
| 2114 | ||||
| 2115 | static nodes_iterator nodes_end(const PGOUseFunc *G) { | |||
| 2116 | return nodes_iterator(G->getFunc().end()); | |||
| 2117 | } | |||
| 2118 | }; | |||
| 2119 | ||||
| 2120 | template <> struct DOTGraphTraits<PGOUseFunc *> : DefaultDOTGraphTraits { | |||
| 2121 | explicit DOTGraphTraits(bool isSimple = false) | |||
| 2122 | : DefaultDOTGraphTraits(isSimple) {} | |||
| 2123 | ||||
| 2124 | static std::string getGraphName(const PGOUseFunc *G) { | |||
| 2125 | return std::string(G->getFunc().getName()); | |||
| 2126 | } | |||
| 2127 | ||||
| 2128 | std::string getNodeLabel(const BasicBlock *Node, const PGOUseFunc *Graph) { | |||
| 2129 | std::string Result; | |||
| 2130 | raw_string_ostream OS(Result); | |||
| 2131 | ||||
| 2132 | OS << getSimpleNodeName(Node) << ":\\l"; | |||
| 2133 | UseBBInfo *BI = Graph->findBBInfo(Node); | |||
| 2134 | OS << "Count : "; | |||
| 2135 | if (BI && BI->CountValid) | |||
| 2136 | OS << BI->CountValue << "\\l"; | |||
| 2137 | else | |||
| 2138 | OS << "Unknown\\l"; | |||
| 2139 | ||||
| 2140 | if (!PGOInstrSelect) | |||
| 2141 | return Result; | |||
| 2142 | ||||
| 2143 | for (const Instruction &I : *Node) { | |||
| 2144 | if (!isa<SelectInst>(&I)) | |||
| 2145 | continue; | |||
| 2146 | // Display scaled counts for SELECT instruction: | |||
| 2147 | OS << "SELECT : { T = "; | |||
| 2148 | uint64_t TC, FC; | |||
| 2149 | bool HasProf = I.extractProfMetadata(TC, FC); | |||
| 2150 | if (!HasProf) | |||
| 2151 | OS << "Unknown, F = Unknown }\\l"; | |||
| 2152 | else | |||
| 2153 | OS << TC << ", F = " << FC << " }\\l"; | |||
| 2154 | } | |||
| 2155 | return Result; | |||
| 2156 | } | |||
| 2157 | }; | |||
| 2158 | ||||
| 2159 | } // end namespace llvm |