[yaml2obj/obj2yaml] - Add support for .stack_sizes sections.
[llvm-complete.git] / lib / Transforms / Instrumentation / PGOInstrumentation.cpp
blobe776d59cccb5ba7f4446caf0e886401ab7e3c81a
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.
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.
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.
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.
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).
48 //===----------------------------------------------------------------------===//
50 #include "CFGMST.h"
51 #include "llvm/ADT/APInt.h"
52 #include "llvm/ADT/ArrayRef.h"
53 #include "llvm/ADT/STLExtras.h"
54 #include "llvm/ADT/SmallVector.h"
55 #include "llvm/ADT/Statistic.h"
56 #include "llvm/ADT/StringRef.h"
57 #include "llvm/ADT/Triple.h"
58 #include "llvm/ADT/Twine.h"
59 #include "llvm/ADT/iterator.h"
60 #include "llvm/ADT/iterator_range.h"
61 #include "llvm/Analysis/BlockFrequencyInfo.h"
62 #include "llvm/Analysis/BranchProbabilityInfo.h"
63 #include "llvm/Analysis/CFG.h"
64 #include "llvm/Analysis/IndirectCallVisitor.h"
65 #include "llvm/Analysis/LoopInfo.h"
66 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
67 #include "llvm/Analysis/ProfileSummaryInfo.h"
68 #include "llvm/IR/Attributes.h"
69 #include "llvm/IR/BasicBlock.h"
70 #include "llvm/IR/CFG.h"
71 #include "llvm/IR/CallSite.h"
72 #include "llvm/IR/Comdat.h"
73 #include "llvm/IR/Constant.h"
74 #include "llvm/IR/Constants.h"
75 #include "llvm/IR/DiagnosticInfo.h"
76 #include "llvm/IR/Dominators.h"
77 #include "llvm/IR/Function.h"
78 #include "llvm/IR/GlobalAlias.h"
79 #include "llvm/IR/GlobalValue.h"
80 #include "llvm/IR/GlobalVariable.h"
81 #include "llvm/IR/IRBuilder.h"
82 #include "llvm/IR/InstVisitor.h"
83 #include "llvm/IR/InstrTypes.h"
84 #include "llvm/IR/Instruction.h"
85 #include "llvm/IR/Instructions.h"
86 #include "llvm/IR/IntrinsicInst.h"
87 #include "llvm/IR/Intrinsics.h"
88 #include "llvm/IR/LLVMContext.h"
89 #include "llvm/IR/MDBuilder.h"
90 #include "llvm/IR/Module.h"
91 #include "llvm/IR/PassManager.h"
92 #include "llvm/IR/ProfileSummary.h"
93 #include "llvm/IR/Type.h"
94 #include "llvm/IR/Value.h"
95 #include "llvm/Pass.h"
96 #include "llvm/ProfileData/InstrProf.h"
97 #include "llvm/ProfileData/InstrProfReader.h"
98 #include "llvm/Support/BranchProbability.h"
99 #include "llvm/Support/Casting.h"
100 #include "llvm/Support/CommandLine.h"
101 #include "llvm/Support/DOTGraphTraits.h"
102 #include "llvm/Support/Debug.h"
103 #include "llvm/Support/Error.h"
104 #include "llvm/Support/ErrorHandling.h"
105 #include "llvm/Support/GraphWriter.h"
106 #include "llvm/Support/JamCRC.h"
107 #include "llvm/Support/raw_ostream.h"
108 #include "llvm/Transforms/Instrumentation.h"
109 #include "llvm/Transforms/Instrumentation/PGOInstrumentation.h"
110 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
111 #include "llvm/Transforms/Utils/MisExpect.h"
112 #include <algorithm>
113 #include <cassert>
114 #include <cstdint>
115 #include <memory>
116 #include <numeric>
117 #include <string>
118 #include <unordered_map>
119 #include <utility>
120 #include <vector>
122 using namespace llvm;
123 using ProfileCount = Function::ProfileCount;
125 #define DEBUG_TYPE "pgo-instrumentation"
127 STATISTIC(NumOfPGOInstrument, "Number of edges instrumented.");
128 STATISTIC(NumOfPGOSelectInsts, "Number of select instruction instrumented.");
129 STATISTIC(NumOfPGOMemIntrinsics, "Number of mem intrinsics instrumented.");
130 STATISTIC(NumOfPGOEdge, "Number of edges.");
131 STATISTIC(NumOfPGOBB, "Number of basic-blocks.");
132 STATISTIC(NumOfPGOSplit, "Number of critical edge splits.");
133 STATISTIC(NumOfPGOFunc, "Number of functions having valid profile counts.");
134 STATISTIC(NumOfPGOMismatch, "Number of functions having mismatch profile.");
135 STATISTIC(NumOfPGOMissing, "Number of functions without profile.");
136 STATISTIC(NumOfPGOICall, "Number of indirect call value instrumentations.");
137 STATISTIC(NumOfCSPGOInstrument, "Number of edges instrumented in CSPGO.");
138 STATISTIC(NumOfCSPGOSelectInsts,
139 "Number of select instruction instrumented in CSPGO.");
140 STATISTIC(NumOfCSPGOMemIntrinsics,
141 "Number of mem intrinsics instrumented in CSPGO.");
142 STATISTIC(NumOfCSPGOEdge, "Number of edges in CSPGO.");
143 STATISTIC(NumOfCSPGOBB, "Number of basic-blocks in CSPGO.");
144 STATISTIC(NumOfCSPGOSplit, "Number of critical edge splits in CSPGO.");
145 STATISTIC(NumOfCSPGOFunc,
146 "Number of functions having valid profile counts in CSPGO.");
147 STATISTIC(NumOfCSPGOMismatch,
148 "Number of functions having mismatch profile in CSPGO.");
149 STATISTIC(NumOfCSPGOMissing, "Number of functions without profile in CSPGO.");
151 // Command line option to specify the file to read profile from. This is
152 // mainly used for testing.
153 static cl::opt<std::string>
154 PGOTestProfileFile("pgo-test-profile-file", cl::init(""), cl::Hidden,
155 cl::value_desc("filename"),
156 cl::desc("Specify the path of profile data file. This is"
157 "mainly for test purpose."));
158 static cl::opt<std::string> PGOTestProfileRemappingFile(
159 "pgo-test-profile-remapping-file", cl::init(""), cl::Hidden,
160 cl::value_desc("filename"),
161 cl::desc("Specify the path of profile remapping file. This is mainly for "
162 "test purpose."));
164 // Command line option to disable value profiling. The default is false:
165 // i.e. value profiling is enabled by default. This is for debug purpose.
166 static cl::opt<bool> DisableValueProfiling("disable-vp", cl::init(false),
167 cl::Hidden,
168 cl::desc("Disable Value Profiling"));
170 // Command line option to set the maximum number of VP annotations to write to
171 // the metadata for a single indirect call callsite.
172 static cl::opt<unsigned> MaxNumAnnotations(
173 "icp-max-annotations", cl::init(3), cl::Hidden, cl::ZeroOrMore,
174 cl::desc("Max number of annotations for a single indirect "
175 "call callsite"));
177 // Command line option to set the maximum number of value annotations
178 // to write to the metadata for a single memop intrinsic.
179 static cl::opt<unsigned> MaxNumMemOPAnnotations(
180 "memop-max-annotations", cl::init(4), cl::Hidden, cl::ZeroOrMore,
181 cl::desc("Max number of preicise value annotations for a single memop"
182 "intrinsic"));
184 // Command line option to control appending FunctionHash to the name of a COMDAT
185 // function. This is to avoid the hash mismatch caused by the preinliner.
186 static cl::opt<bool> DoComdatRenaming(
187 "do-comdat-renaming", cl::init(false), cl::Hidden,
188 cl::desc("Append function hash to the name of COMDAT function to avoid "
189 "function hash mismatch due to the preinliner"));
191 // Command line option to enable/disable the warning about missing profile
192 // information.
193 static cl::opt<bool>
194 PGOWarnMissing("pgo-warn-missing-function", cl::init(false), cl::Hidden,
195 cl::desc("Use this option to turn on/off "
196 "warnings about missing profile data for "
197 "functions."));
199 // Command line option to enable/disable the warning about a hash mismatch in
200 // the profile data.
201 static cl::opt<bool>
202 NoPGOWarnMismatch("no-pgo-warn-mismatch", cl::init(false), cl::Hidden,
203 cl::desc("Use this option to turn off/on "
204 "warnings about profile cfg mismatch."));
206 // Command line option to enable/disable the warning about a hash mismatch in
207 // the profile data for Comdat functions, which often turns out to be false
208 // positive due to the pre-instrumentation inline.
209 static cl::opt<bool>
210 NoPGOWarnMismatchComdat("no-pgo-warn-mismatch-comdat", cl::init(true),
211 cl::Hidden,
212 cl::desc("The option is used to turn on/off "
213 "warnings about hash mismatch for comdat "
214 "functions."));
216 // Command line option to enable/disable select instruction instrumentation.
217 static cl::opt<bool>
218 PGOInstrSelect("pgo-instr-select", cl::init(true), cl::Hidden,
219 cl::desc("Use this option to turn on/off SELECT "
220 "instruction instrumentation. "));
222 // Command line option to turn on CFG dot or text dump of raw profile counts
223 static cl::opt<PGOViewCountsType> PGOViewRawCounts(
224 "pgo-view-raw-counts", cl::Hidden,
225 cl::desc("A boolean option to show CFG dag or text "
226 "with raw profile counts from "
227 "profile data. See also option "
228 "-pgo-view-counts. To limit graph "
229 "display to only one function, use "
230 "filtering option -view-bfi-func-name."),
231 cl::values(clEnumValN(PGOVCT_None, "none", "do not show."),
232 clEnumValN(PGOVCT_Graph, "graph", "show a graph."),
233 clEnumValN(PGOVCT_Text, "text", "show in text.")));
235 // Command line option to enable/disable memop intrinsic call.size profiling.
236 static cl::opt<bool>
237 PGOInstrMemOP("pgo-instr-memop", cl::init(true), cl::Hidden,
238 cl::desc("Use this option to turn on/off "
239 "memory intrinsic size profiling."));
241 // Emit branch probability as optimization remarks.
242 static cl::opt<bool>
243 EmitBranchProbability("pgo-emit-branch-prob", cl::init(false), cl::Hidden,
244 cl::desc("When this option is on, the annotated "
245 "branch probability will be emitted as "
246 "optimization remarks: -{Rpass|"
247 "pass-remarks}=pgo-instrumentation"));
249 // Command line option to turn on CFG dot dump after profile annotation.
250 // Defined in Analysis/BlockFrequencyInfo.cpp: -pgo-view-counts
251 extern cl::opt<PGOViewCountsType> PGOViewCounts;
253 // Command line option to specify the name of the function for CFG dump
254 // Defined in Analysis/BlockFrequencyInfo.cpp: -view-bfi-func-name=
255 extern cl::opt<std::string> ViewBlockFreqFuncName;
257 // Return a string describing the branch condition that can be
258 // used in static branch probability heuristics:
259 static std::string getBranchCondString(Instruction *TI) {
260 BranchInst *BI = dyn_cast<BranchInst>(TI);
261 if (!BI || !BI->isConditional())
262 return std::string();
264 Value *Cond = BI->getCondition();
265 ICmpInst *CI = dyn_cast<ICmpInst>(Cond);
266 if (!CI)
267 return std::string();
269 std::string result;
270 raw_string_ostream OS(result);
271 OS << CmpInst::getPredicateName(CI->getPredicate()) << "_";
272 CI->getOperand(0)->getType()->print(OS, true);
274 Value *RHS = CI->getOperand(1);
275 ConstantInt *CV = dyn_cast<ConstantInt>(RHS);
276 if (CV) {
277 if (CV->isZero())
278 OS << "_Zero";
279 else if (CV->isOne())
280 OS << "_One";
281 else if (CV->isMinusOne())
282 OS << "_MinusOne";
283 else
284 OS << "_Const";
286 OS.flush();
287 return result;
290 namespace {
292 /// The select instruction visitor plays three roles specified
293 /// by the mode. In \c VM_counting mode, it simply counts the number of
294 /// select instructions. In \c VM_instrument mode, it inserts code to count
295 /// the number times TrueValue of select is taken. In \c VM_annotate mode,
296 /// it reads the profile data and annotate the select instruction with metadata.
297 enum VisitMode { VM_counting, VM_instrument, VM_annotate };
298 class PGOUseFunc;
300 /// Instruction Visitor class to visit select instructions.
301 struct SelectInstVisitor : public InstVisitor<SelectInstVisitor> {
302 Function &F;
303 unsigned NSIs = 0; // Number of select instructions instrumented.
304 VisitMode Mode = VM_counting; // Visiting mode.
305 unsigned *CurCtrIdx = nullptr; // Pointer to current counter index.
306 unsigned TotalNumCtrs = 0; // Total number of counters
307 GlobalVariable *FuncNameVar = nullptr;
308 uint64_t FuncHash = 0;
309 PGOUseFunc *UseFunc = nullptr;
311 SelectInstVisitor(Function &Func) : F(Func) {}
313 void countSelects(Function &Func) {
314 NSIs = 0;
315 Mode = VM_counting;
316 visit(Func);
319 // Visit the IR stream and instrument all select instructions. \p
320 // Ind is a pointer to the counter index variable; \p TotalNC
321 // is the total number of counters; \p FNV is the pointer to the
322 // PGO function name var; \p FHash is the function hash.
323 void instrumentSelects(Function &Func, unsigned *Ind, unsigned TotalNC,
324 GlobalVariable *FNV, uint64_t FHash) {
325 Mode = VM_instrument;
326 CurCtrIdx = Ind;
327 TotalNumCtrs = TotalNC;
328 FuncHash = FHash;
329 FuncNameVar = FNV;
330 visit(Func);
333 // Visit the IR stream and annotate all select instructions.
334 void annotateSelects(Function &Func, PGOUseFunc *UF, unsigned *Ind) {
335 Mode = VM_annotate;
336 UseFunc = UF;
337 CurCtrIdx = Ind;
338 visit(Func);
341 void instrumentOneSelectInst(SelectInst &SI);
342 void annotateOneSelectInst(SelectInst &SI);
344 // Visit \p SI instruction and perform tasks according to visit mode.
345 void visitSelectInst(SelectInst &SI);
347 // Return the number of select instructions. This needs be called after
348 // countSelects().
349 unsigned getNumOfSelectInsts() const { return NSIs; }
352 /// Instruction Visitor class to visit memory intrinsic calls.
353 struct MemIntrinsicVisitor : public InstVisitor<MemIntrinsicVisitor> {
354 Function &F;
355 unsigned NMemIs = 0; // Number of memIntrinsics instrumented.
356 VisitMode Mode = VM_counting; // Visiting mode.
357 unsigned CurCtrId = 0; // Current counter index.
358 unsigned TotalNumCtrs = 0; // Total number of counters
359 GlobalVariable *FuncNameVar = nullptr;
360 uint64_t FuncHash = 0;
361 PGOUseFunc *UseFunc = nullptr;
362 std::vector<Instruction *> Candidates;
364 MemIntrinsicVisitor(Function &Func) : F(Func) {}
366 void countMemIntrinsics(Function &Func) {
367 NMemIs = 0;
368 Mode = VM_counting;
369 visit(Func);
372 void instrumentMemIntrinsics(Function &Func, unsigned TotalNC,
373 GlobalVariable *FNV, uint64_t FHash) {
374 Mode = VM_instrument;
375 TotalNumCtrs = TotalNC;
376 FuncHash = FHash;
377 FuncNameVar = FNV;
378 visit(Func);
381 std::vector<Instruction *> findMemIntrinsics(Function &Func) {
382 Candidates.clear();
383 Mode = VM_annotate;
384 visit(Func);
385 return Candidates;
388 // Visit the IR stream and annotate all mem intrinsic call instructions.
389 void instrumentOneMemIntrinsic(MemIntrinsic &MI);
391 // Visit \p MI instruction and perform tasks according to visit mode.
392 void visitMemIntrinsic(MemIntrinsic &SI);
394 unsigned getNumOfMemIntrinsics() const { return NMemIs; }
397 class PGOInstrumentationGenLegacyPass : public ModulePass {
398 public:
399 static char ID;
401 PGOInstrumentationGenLegacyPass(bool IsCS = false)
402 : ModulePass(ID), IsCS(IsCS) {
403 initializePGOInstrumentationGenLegacyPassPass(
404 *PassRegistry::getPassRegistry());
407 StringRef getPassName() const override { return "PGOInstrumentationGenPass"; }
409 private:
410 // Is this is context-sensitive instrumentation.
411 bool IsCS;
412 bool runOnModule(Module &M) override;
414 void getAnalysisUsage(AnalysisUsage &AU) const override {
415 AU.addRequired<BlockFrequencyInfoWrapperPass>();
419 class PGOInstrumentationUseLegacyPass : public ModulePass {
420 public:
421 static char ID;
423 // Provide the profile filename as the parameter.
424 PGOInstrumentationUseLegacyPass(std::string Filename = "", bool IsCS = false)
425 : ModulePass(ID), ProfileFileName(std::move(Filename)), IsCS(IsCS) {
426 if (!PGOTestProfileFile.empty())
427 ProfileFileName = PGOTestProfileFile;
428 initializePGOInstrumentationUseLegacyPassPass(
429 *PassRegistry::getPassRegistry());
432 StringRef getPassName() const override { return "PGOInstrumentationUsePass"; }
434 private:
435 std::string ProfileFileName;
436 // Is this is context-sensitive instrumentation use.
437 bool IsCS;
439 bool runOnModule(Module &M) override;
441 void getAnalysisUsage(AnalysisUsage &AU) const override {
442 AU.addRequired<ProfileSummaryInfoWrapperPass>();
443 AU.addRequired<BlockFrequencyInfoWrapperPass>();
447 class PGOInstrumentationGenCreateVarLegacyPass : public ModulePass {
448 public:
449 static char ID;
450 StringRef getPassName() const override {
451 return "PGOInstrumentationGenCreateVarPass";
453 PGOInstrumentationGenCreateVarLegacyPass(std::string CSInstrName = "")
454 : ModulePass(ID), InstrProfileOutput(CSInstrName) {
455 initializePGOInstrumentationGenCreateVarLegacyPassPass(
456 *PassRegistry::getPassRegistry());
459 private:
460 bool runOnModule(Module &M) override {
461 createProfileFileNameVar(M, InstrProfileOutput);
462 createIRLevelProfileFlagVar(M, true);
463 return false;
465 std::string InstrProfileOutput;
468 } // end anonymous namespace
470 char PGOInstrumentationGenLegacyPass::ID = 0;
472 INITIALIZE_PASS_BEGIN(PGOInstrumentationGenLegacyPass, "pgo-instr-gen",
473 "PGO instrumentation.", false, false)
474 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
475 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)
476 INITIALIZE_PASS_END(PGOInstrumentationGenLegacyPass, "pgo-instr-gen",
477 "PGO instrumentation.", false, false)
479 ModulePass *llvm::createPGOInstrumentationGenLegacyPass(bool IsCS) {
480 return new PGOInstrumentationGenLegacyPass(IsCS);
483 char PGOInstrumentationUseLegacyPass::ID = 0;
485 INITIALIZE_PASS_BEGIN(PGOInstrumentationUseLegacyPass, "pgo-instr-use",
486 "Read PGO instrumentation profile.", false, false)
487 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
488 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)
489 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
490 INITIALIZE_PASS_END(PGOInstrumentationUseLegacyPass, "pgo-instr-use",
491 "Read PGO instrumentation profile.", false, false)
493 ModulePass *llvm::createPGOInstrumentationUseLegacyPass(StringRef Filename,
494 bool IsCS) {
495 return new PGOInstrumentationUseLegacyPass(Filename.str(), IsCS);
498 char PGOInstrumentationGenCreateVarLegacyPass::ID = 0;
500 INITIALIZE_PASS(PGOInstrumentationGenCreateVarLegacyPass,
501 "pgo-instr-gen-create-var",
502 "Create PGO instrumentation version variable for CSPGO.", false,
503 false)
505 ModulePass *
506 llvm::createPGOInstrumentationGenCreateVarLegacyPass(StringRef CSInstrName) {
507 return new PGOInstrumentationGenCreateVarLegacyPass(CSInstrName);
510 namespace {
512 /// An MST based instrumentation for PGO
514 /// Implements a Minimum Spanning Tree (MST) based instrumentation for PGO
515 /// in the function level.
516 struct PGOEdge {
517 // This class implements the CFG edges. Note the CFG can be a multi-graph.
518 // So there might be multiple edges with same SrcBB and DestBB.
519 const BasicBlock *SrcBB;
520 const BasicBlock *DestBB;
521 uint64_t Weight;
522 bool InMST = false;
523 bool Removed = false;
524 bool IsCritical = false;
526 PGOEdge(const BasicBlock *Src, const BasicBlock *Dest, uint64_t W = 1)
527 : SrcBB(Src), DestBB(Dest), Weight(W) {}
529 // Return the information string of an edge.
530 const std::string infoString() const {
531 return (Twine(Removed ? "-" : " ") + (InMST ? " " : "*") +
532 (IsCritical ? "c" : " ") + " W=" + Twine(Weight)).str();
536 // This class stores the auxiliary information for each BB.
537 struct BBInfo {
538 BBInfo *Group;
539 uint32_t Index;
540 uint32_t Rank = 0;
542 BBInfo(unsigned IX) : Group(this), Index(IX) {}
544 // Return the information string of this object.
545 const std::string infoString() const {
546 return (Twine("Index=") + Twine(Index)).str();
549 // Empty function -- only applicable to UseBBInfo.
550 void addOutEdge(PGOEdge *E LLVM_ATTRIBUTE_UNUSED) {}
552 // Empty function -- only applicable to UseBBInfo.
553 void addInEdge(PGOEdge *E LLVM_ATTRIBUTE_UNUSED) {}
556 // This class implements the CFG edges. Note the CFG can be a multi-graph.
557 template <class Edge, class BBInfo> class FuncPGOInstrumentation {
558 private:
559 Function &F;
561 // Is this is context-sensitive instrumentation.
562 bool IsCS;
564 // A map that stores the Comdat group in function F.
565 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers;
567 void computeCFGHash();
568 void renameComdatFunction();
570 public:
571 std::vector<std::vector<Instruction *>> ValueSites;
572 SelectInstVisitor SIVisitor;
573 MemIntrinsicVisitor MIVisitor;
574 std::string FuncName;
575 GlobalVariable *FuncNameVar;
577 // CFG hash value for this function.
578 uint64_t FunctionHash = 0;
580 // The Minimum Spanning Tree of function CFG.
581 CFGMST<Edge, BBInfo> MST;
583 // Collect all the BBs that will be instrumented, and store them in
584 // InstrumentBBs.
585 void getInstrumentBBs(std::vector<BasicBlock *> &InstrumentBBs);
587 // Give an edge, find the BB that will be instrumented.
588 // Return nullptr if there is no BB to be instrumented.
589 BasicBlock *getInstrBB(Edge *E);
591 // Return the auxiliary BB information.
592 BBInfo &getBBInfo(const BasicBlock *BB) const { return MST.getBBInfo(BB); }
594 // Return the auxiliary BB information if available.
595 BBInfo *findBBInfo(const BasicBlock *BB) const { return MST.findBBInfo(BB); }
597 // Dump edges and BB information.
598 void dumpInfo(std::string Str = "") const {
599 MST.dumpEdges(dbgs(), Twine("Dump Function ") + FuncName + " Hash: " +
600 Twine(FunctionHash) + "\t" + Str);
603 FuncPGOInstrumentation(
604 Function &Func,
605 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers,
606 bool CreateGlobalVar = false, BranchProbabilityInfo *BPI = nullptr,
607 BlockFrequencyInfo *BFI = nullptr, bool IsCS = false)
608 : F(Func), IsCS(IsCS), ComdatMembers(ComdatMembers),
609 ValueSites(IPVK_Last + 1), SIVisitor(Func), MIVisitor(Func),
610 MST(F, BPI, BFI) {
611 // This should be done before CFG hash computation.
612 SIVisitor.countSelects(Func);
613 MIVisitor.countMemIntrinsics(Func);
614 if (!IsCS) {
615 NumOfPGOSelectInsts += SIVisitor.getNumOfSelectInsts();
616 NumOfPGOMemIntrinsics += MIVisitor.getNumOfMemIntrinsics();
617 NumOfPGOBB += MST.BBInfos.size();
618 ValueSites[IPVK_IndirectCallTarget] = findIndirectCalls(Func);
619 } else {
620 NumOfCSPGOSelectInsts += SIVisitor.getNumOfSelectInsts();
621 NumOfCSPGOMemIntrinsics += MIVisitor.getNumOfMemIntrinsics();
622 NumOfCSPGOBB += MST.BBInfos.size();
624 ValueSites[IPVK_MemOPSize] = MIVisitor.findMemIntrinsics(Func);
626 FuncName = getPGOFuncName(F);
627 computeCFGHash();
628 if (!ComdatMembers.empty())
629 renameComdatFunction();
630 LLVM_DEBUG(dumpInfo("after CFGMST"));
632 for (auto &E : MST.AllEdges) {
633 if (E->Removed)
634 continue;
635 IsCS ? NumOfCSPGOEdge++ : NumOfPGOEdge++;
636 if (!E->InMST)
637 IsCS ? NumOfCSPGOInstrument++ : NumOfPGOInstrument++;
640 if (CreateGlobalVar)
641 FuncNameVar = createPGOFuncNameVar(F, FuncName);
645 } // end anonymous namespace
647 // Compute Hash value for the CFG: the lower 32 bits are CRC32 of the index
648 // value of each BB in the CFG. The higher 32 bits record the number of edges.
649 template <class Edge, class BBInfo>
650 void FuncPGOInstrumentation<Edge, BBInfo>::computeCFGHash() {
651 std::vector<char> Indexes;
652 JamCRC JC;
653 for (auto &BB : F) {
654 const Instruction *TI = BB.getTerminator();
655 for (unsigned I = 0, E = TI->getNumSuccessors(); I != E; ++I) {
656 BasicBlock *Succ = TI->getSuccessor(I);
657 auto BI = findBBInfo(Succ);
658 if (BI == nullptr)
659 continue;
660 uint32_t Index = BI->Index;
661 for (int J = 0; J < 4; J++)
662 Indexes.push_back((char)(Index >> (J * 8)));
665 JC.update(Indexes);
667 // Hash format for context sensitive profile. Reserve 4 bits for other
668 // information.
669 FunctionHash = (uint64_t)SIVisitor.getNumOfSelectInsts() << 56 |
670 (uint64_t)ValueSites[IPVK_IndirectCallTarget].size() << 48 |
671 //(uint64_t)ValueSites[IPVK_MemOPSize].size() << 40 |
672 (uint64_t)MST.AllEdges.size() << 32 | JC.getCRC();
673 // Reserve bit 60-63 for other information purpose.
674 FunctionHash &= 0x0FFFFFFFFFFFFFFF;
675 if (IsCS)
676 NamedInstrProfRecord::setCSFlagInHash(FunctionHash);
677 LLVM_DEBUG(dbgs() << "Function Hash Computation for " << F.getName() << ":\n"
678 << " CRC = " << JC.getCRC()
679 << ", Selects = " << SIVisitor.getNumOfSelectInsts()
680 << ", Edges = " << MST.AllEdges.size() << ", ICSites = "
681 << ValueSites[IPVK_IndirectCallTarget].size()
682 << ", Hash = " << FunctionHash << "\n";);
685 // Check if we can safely rename this Comdat function.
686 static bool canRenameComdat(
687 Function &F,
688 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers) {
689 if (!DoComdatRenaming || !canRenameComdatFunc(F, true))
690 return false;
692 // FIXME: Current only handle those Comdat groups that only containing one
693 // function and function aliases.
694 // (1) For a Comdat group containing multiple functions, we need to have a
695 // unique postfix based on the hashes for each function. There is a
696 // non-trivial code refactoring to do this efficiently.
697 // (2) Variables can not be renamed, so we can not rename Comdat function in a
698 // group including global vars.
699 Comdat *C = F.getComdat();
700 for (auto &&CM : make_range(ComdatMembers.equal_range(C))) {
701 if (dyn_cast<GlobalAlias>(CM.second))
702 continue;
703 Function *FM = dyn_cast<Function>(CM.second);
704 if (FM != &F)
705 return false;
707 return true;
710 // Append the CFGHash to the Comdat function name.
711 template <class Edge, class BBInfo>
712 void FuncPGOInstrumentation<Edge, BBInfo>::renameComdatFunction() {
713 if (!canRenameComdat(F, ComdatMembers))
714 return;
715 std::string OrigName = F.getName().str();
716 std::string NewFuncName =
717 Twine(F.getName() + "." + Twine(FunctionHash)).str();
718 F.setName(Twine(NewFuncName));
719 GlobalAlias::create(GlobalValue::WeakAnyLinkage, OrigName, &F);
720 FuncName = Twine(FuncName + "." + Twine(FunctionHash)).str();
721 Comdat *NewComdat;
722 Module *M = F.getParent();
723 // For AvailableExternallyLinkage functions, change the linkage to
724 // LinkOnceODR and put them into comdat. This is because after renaming, there
725 // is no backup external copy available for the function.
726 if (!F.hasComdat()) {
727 assert(F.getLinkage() == GlobalValue::AvailableExternallyLinkage);
728 NewComdat = M->getOrInsertComdat(StringRef(NewFuncName));
729 F.setLinkage(GlobalValue::LinkOnceODRLinkage);
730 F.setComdat(NewComdat);
731 return;
734 // This function belongs to a single function Comdat group.
735 Comdat *OrigComdat = F.getComdat();
736 std::string NewComdatName =
737 Twine(OrigComdat->getName() + "." + Twine(FunctionHash)).str();
738 NewComdat = M->getOrInsertComdat(StringRef(NewComdatName));
739 NewComdat->setSelectionKind(OrigComdat->getSelectionKind());
741 for (auto &&CM : make_range(ComdatMembers.equal_range(OrigComdat))) {
742 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(CM.second)) {
743 // For aliases, change the name directly.
744 assert(dyn_cast<Function>(GA->getAliasee()->stripPointerCasts()) == &F);
745 std::string OrigGAName = GA->getName().str();
746 GA->setName(Twine(GA->getName() + "." + Twine(FunctionHash)));
747 GlobalAlias::create(GlobalValue::WeakAnyLinkage, OrigGAName, GA);
748 continue;
750 // Must be a function.
751 Function *CF = dyn_cast<Function>(CM.second);
752 assert(CF);
753 CF->setComdat(NewComdat);
757 // Collect all the BBs that will be instruments and return them in
758 // InstrumentBBs and setup InEdges/OutEdge for UseBBInfo.
759 template <class Edge, class BBInfo>
760 void FuncPGOInstrumentation<Edge, BBInfo>::getInstrumentBBs(
761 std::vector<BasicBlock *> &InstrumentBBs) {
762 // Use a worklist as we will update the vector during the iteration.
763 std::vector<Edge *> EdgeList;
764 EdgeList.reserve(MST.AllEdges.size());
765 for (auto &E : MST.AllEdges)
766 EdgeList.push_back(E.get());
768 for (auto &E : EdgeList) {
769 BasicBlock *InstrBB = getInstrBB(E);
770 if (InstrBB)
771 InstrumentBBs.push_back(InstrBB);
774 // Set up InEdges/OutEdges for all BBs.
775 for (auto &E : MST.AllEdges) {
776 if (E->Removed)
777 continue;
778 const BasicBlock *SrcBB = E->SrcBB;
779 const BasicBlock *DestBB = E->DestBB;
780 BBInfo &SrcInfo = getBBInfo(SrcBB);
781 BBInfo &DestInfo = getBBInfo(DestBB);
782 SrcInfo.addOutEdge(E.get());
783 DestInfo.addInEdge(E.get());
787 // Given a CFG E to be instrumented, find which BB to place the instrumented
788 // code. The function will split the critical edge if necessary.
789 template <class Edge, class BBInfo>
790 BasicBlock *FuncPGOInstrumentation<Edge, BBInfo>::getInstrBB(Edge *E) {
791 if (E->InMST || E->Removed)
792 return nullptr;
794 BasicBlock *SrcBB = const_cast<BasicBlock *>(E->SrcBB);
795 BasicBlock *DestBB = const_cast<BasicBlock *>(E->DestBB);
796 // For a fake edge, instrument the real BB.
797 if (SrcBB == nullptr)
798 return DestBB;
799 if (DestBB == nullptr)
800 return SrcBB;
802 auto canInstrument = [](BasicBlock *BB) -> BasicBlock * {
803 // There are basic blocks (such as catchswitch) cannot be instrumented.
804 // If the returned first insertion point is the end of BB, skip this BB.
805 if (BB->getFirstInsertionPt() == BB->end())
806 return nullptr;
807 return BB;
810 // Instrument the SrcBB if it has a single successor,
811 // otherwise, the DestBB if this is not a critical edge.
812 Instruction *TI = SrcBB->getTerminator();
813 if (TI->getNumSuccessors() <= 1)
814 return canInstrument(SrcBB);
815 if (!E->IsCritical)
816 return canInstrument(DestBB);
818 unsigned SuccNum = GetSuccessorNumber(SrcBB, DestBB);
819 BasicBlock *InstrBB = SplitCriticalEdge(TI, SuccNum);
820 if (!InstrBB) {
821 LLVM_DEBUG(
822 dbgs() << "Fail to split critical edge: not instrument this edge.\n");
823 return nullptr;
825 // For a critical edge, we have to split. Instrument the newly
826 // created BB.
827 IsCS ? NumOfCSPGOSplit++ : NumOfPGOSplit++;
828 LLVM_DEBUG(dbgs() << "Split critical edge: " << getBBInfo(SrcBB).Index
829 << " --> " << getBBInfo(DestBB).Index << "\n");
830 // Need to add two new edges. First one: Add new edge of SrcBB->InstrBB.
831 MST.addEdge(SrcBB, InstrBB, 0);
832 // Second one: Add new edge of InstrBB->DestBB.
833 Edge &NewEdge1 = MST.addEdge(InstrBB, DestBB, 0);
834 NewEdge1.InMST = true;
835 E->Removed = true;
837 return canInstrument(InstrBB);
840 // Visit all edge and instrument the edges not in MST, and do value profiling.
841 // Critical edges will be split.
842 static void instrumentOneFunc(
843 Function &F, Module *M, BranchProbabilityInfo *BPI, BlockFrequencyInfo *BFI,
844 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers,
845 bool IsCS) {
846 // Split indirectbr critical edges here before computing the MST rather than
847 // later in getInstrBB() to avoid invalidating it.
848 SplitIndirectBrCriticalEdges(F, BPI, BFI);
850 FuncPGOInstrumentation<PGOEdge, BBInfo> FuncInfo(F, ComdatMembers, true, BPI,
851 BFI, IsCS);
852 std::vector<BasicBlock *> InstrumentBBs;
853 FuncInfo.getInstrumentBBs(InstrumentBBs);
854 unsigned NumCounters =
855 InstrumentBBs.size() + FuncInfo.SIVisitor.getNumOfSelectInsts();
857 uint32_t I = 0;
858 Type *I8PtrTy = Type::getInt8PtrTy(M->getContext());
859 for (auto *InstrBB : InstrumentBBs) {
860 IRBuilder<> Builder(InstrBB, InstrBB->getFirstInsertionPt());
861 assert(Builder.GetInsertPoint() != InstrBB->end() &&
862 "Cannot get the Instrumentation point");
863 Builder.CreateCall(
864 Intrinsic::getDeclaration(M, Intrinsic::instrprof_increment),
865 {ConstantExpr::getBitCast(FuncInfo.FuncNameVar, I8PtrTy),
866 Builder.getInt64(FuncInfo.FunctionHash), Builder.getInt32(NumCounters),
867 Builder.getInt32(I++)});
870 // Now instrument select instructions:
871 FuncInfo.SIVisitor.instrumentSelects(F, &I, NumCounters, FuncInfo.FuncNameVar,
872 FuncInfo.FunctionHash);
873 assert(I == NumCounters);
875 if (DisableValueProfiling)
876 return;
878 unsigned NumIndirectCalls = 0;
879 for (auto &I : FuncInfo.ValueSites[IPVK_IndirectCallTarget]) {
880 CallSite CS(I);
881 Value *Callee = CS.getCalledValue();
882 LLVM_DEBUG(dbgs() << "Instrument one indirect call: CallSite Index = "
883 << NumIndirectCalls << "\n");
884 IRBuilder<> Builder(I);
885 assert(Builder.GetInsertPoint() != I->getParent()->end() &&
886 "Cannot get the Instrumentation point");
887 Builder.CreateCall(
888 Intrinsic::getDeclaration(M, Intrinsic::instrprof_value_profile),
889 {ConstantExpr::getBitCast(FuncInfo.FuncNameVar, I8PtrTy),
890 Builder.getInt64(FuncInfo.FunctionHash),
891 Builder.CreatePtrToInt(Callee, Builder.getInt64Ty()),
892 Builder.getInt32(IPVK_IndirectCallTarget),
893 Builder.getInt32(NumIndirectCalls++)});
895 NumOfPGOICall += NumIndirectCalls;
897 // Now instrument memop intrinsic calls.
898 FuncInfo.MIVisitor.instrumentMemIntrinsics(
899 F, NumCounters, FuncInfo.FuncNameVar, FuncInfo.FunctionHash);
902 namespace {
904 // This class represents a CFG edge in profile use compilation.
905 struct PGOUseEdge : public PGOEdge {
906 bool CountValid = false;
907 uint64_t CountValue = 0;
909 PGOUseEdge(const BasicBlock *Src, const BasicBlock *Dest, uint64_t W = 1)
910 : PGOEdge(Src, Dest, W) {}
912 // Set edge count value
913 void setEdgeCount(uint64_t Value) {
914 CountValue = Value;
915 CountValid = true;
918 // Return the information string for this object.
919 const std::string infoString() const {
920 if (!CountValid)
921 return PGOEdge::infoString();
922 return (Twine(PGOEdge::infoString()) + " Count=" + Twine(CountValue))
923 .str();
927 using DirectEdges = SmallVector<PGOUseEdge *, 2>;
929 // This class stores the auxiliary information for each BB.
930 struct UseBBInfo : public BBInfo {
931 uint64_t CountValue = 0;
932 bool CountValid;
933 int32_t UnknownCountInEdge = 0;
934 int32_t UnknownCountOutEdge = 0;
935 DirectEdges InEdges;
936 DirectEdges OutEdges;
938 UseBBInfo(unsigned IX) : BBInfo(IX), CountValid(false) {}
940 UseBBInfo(unsigned IX, uint64_t C)
941 : BBInfo(IX), CountValue(C), CountValid(true) {}
943 // Set the profile count value for this BB.
944 void setBBInfoCount(uint64_t Value) {
945 CountValue = Value;
946 CountValid = true;
949 // Return the information string of this object.
950 const std::string infoString() const {
951 if (!CountValid)
952 return BBInfo::infoString();
953 return (Twine(BBInfo::infoString()) + " Count=" + Twine(CountValue)).str();
956 // Add an OutEdge and update the edge count.
957 void addOutEdge(PGOUseEdge *E) {
958 OutEdges.push_back(E);
959 UnknownCountOutEdge++;
962 // Add an InEdge and update the edge count.
963 void addInEdge(PGOUseEdge *E) {
964 InEdges.push_back(E);
965 UnknownCountInEdge++;
969 } // end anonymous namespace
971 // Sum up the count values for all the edges.
972 static uint64_t sumEdgeCount(const ArrayRef<PGOUseEdge *> Edges) {
973 uint64_t Total = 0;
974 for (auto &E : Edges) {
975 if (E->Removed)
976 continue;
977 Total += E->CountValue;
979 return Total;
982 namespace {
984 class PGOUseFunc {
985 public:
986 PGOUseFunc(Function &Func, Module *Modu,
987 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers,
988 BranchProbabilityInfo *BPI, BlockFrequencyInfo *BFIin,
989 ProfileSummaryInfo *PSI, bool IsCS)
990 : F(Func), M(Modu), BFI(BFIin), PSI(PSI),
991 FuncInfo(Func, ComdatMembers, false, BPI, BFIin, IsCS),
992 FreqAttr(FFA_Normal), IsCS(IsCS) {}
994 // Read counts for the instrumented BB from profile.
995 bool readCounters(IndexedInstrProfReader *PGOReader, bool &AllZeros);
997 // Populate the counts for all BBs.
998 void populateCounters();
1000 // Set the branch weights based on the count values.
1001 void setBranchWeights();
1003 // Annotate the value profile call sites for all value kind.
1004 void annotateValueSites();
1006 // Annotate the value profile call sites for one value kind.
1007 void annotateValueSites(uint32_t Kind);
1009 // Annotate the irreducible loop header weights.
1010 void annotateIrrLoopHeaderWeights();
1012 // The hotness of the function from the profile count.
1013 enum FuncFreqAttr { FFA_Normal, FFA_Cold, FFA_Hot };
1015 // Return the function hotness from the profile.
1016 FuncFreqAttr getFuncFreqAttr() const { return FreqAttr; }
1018 // Return the function hash.
1019 uint64_t getFuncHash() const { return FuncInfo.FunctionHash; }
1021 // Return the profile record for this function;
1022 InstrProfRecord &getProfileRecord() { return ProfileRecord; }
1024 // Return the auxiliary BB information.
1025 UseBBInfo &getBBInfo(const BasicBlock *BB) const {
1026 return FuncInfo.getBBInfo(BB);
1029 // Return the auxiliary BB information if available.
1030 UseBBInfo *findBBInfo(const BasicBlock *BB) const {
1031 return FuncInfo.findBBInfo(BB);
1034 Function &getFunc() const { return F; }
1036 void dumpInfo(std::string Str = "") const {
1037 FuncInfo.dumpInfo(Str);
1040 uint64_t getProgramMaxCount() const { return ProgramMaxCount; }
1041 private:
1042 Function &F;
1043 Module *M;
1044 BlockFrequencyInfo *BFI;
1045 ProfileSummaryInfo *PSI;
1047 // This member stores the shared information with class PGOGenFunc.
1048 FuncPGOInstrumentation<PGOUseEdge, UseBBInfo> FuncInfo;
1050 // The maximum count value in the profile. This is only used in PGO use
1051 // compilation.
1052 uint64_t ProgramMaxCount;
1054 // Position of counter that remains to be read.
1055 uint32_t CountPosition = 0;
1057 // Total size of the profile count for this function.
1058 uint32_t ProfileCountSize = 0;
1060 // ProfileRecord for this function.
1061 InstrProfRecord ProfileRecord;
1063 // Function hotness info derived from profile.
1064 FuncFreqAttr FreqAttr;
1066 // Is to use the context sensitive profile.
1067 bool IsCS;
1069 // Find the Instrumented BB and set the value. Return false on error.
1070 bool setInstrumentedCounts(const std::vector<uint64_t> &CountFromProfile);
1072 // Set the edge counter value for the unknown edge -- there should be only
1073 // one unknown edge.
1074 void setEdgeCount(DirectEdges &Edges, uint64_t Value);
1076 // Return FuncName string;
1077 const std::string getFuncName() const { return FuncInfo.FuncName; }
1079 // Set the hot/cold inline hints based on the count values.
1080 // FIXME: This function should be removed once the functionality in
1081 // the inliner is implemented.
1082 void markFunctionAttributes(uint64_t EntryCount, uint64_t MaxCount) {
1083 if (PSI->isHotCount(EntryCount))
1084 FreqAttr = FFA_Hot;
1085 else if (PSI->isColdCount(MaxCount))
1086 FreqAttr = FFA_Cold;
1090 } // end anonymous namespace
1092 // Visit all the edges and assign the count value for the instrumented
1093 // edges and the BB. Return false on error.
1094 bool PGOUseFunc::setInstrumentedCounts(
1095 const std::vector<uint64_t> &CountFromProfile) {
1097 std::vector<BasicBlock *> InstrumentBBs;
1098 FuncInfo.getInstrumentBBs(InstrumentBBs);
1099 unsigned NumCounters =
1100 InstrumentBBs.size() + FuncInfo.SIVisitor.getNumOfSelectInsts();
1101 // The number of counters here should match the number of counters
1102 // in profile. Return if they mismatch.
1103 if (NumCounters != CountFromProfile.size()) {
1104 return false;
1106 // Set the profile count to the Instrumented BBs.
1107 uint32_t I = 0;
1108 for (BasicBlock *InstrBB : InstrumentBBs) {
1109 uint64_t CountValue = CountFromProfile[I++];
1110 UseBBInfo &Info = getBBInfo(InstrBB);
1111 Info.setBBInfoCount(CountValue);
1113 ProfileCountSize = CountFromProfile.size();
1114 CountPosition = I;
1116 // Set the edge count and update the count of unknown edges for BBs.
1117 auto setEdgeCount = [this](PGOUseEdge *E, uint64_t Value) -> void {
1118 E->setEdgeCount(Value);
1119 this->getBBInfo(E->SrcBB).UnknownCountOutEdge--;
1120 this->getBBInfo(E->DestBB).UnknownCountInEdge--;
1123 // Set the profile count the Instrumented edges. There are BBs that not in
1124 // MST but not instrumented. Need to set the edge count value so that we can
1125 // populate the profile counts later.
1126 for (auto &E : FuncInfo.MST.AllEdges) {
1127 if (E->Removed || E->InMST)
1128 continue;
1129 const BasicBlock *SrcBB = E->SrcBB;
1130 UseBBInfo &SrcInfo = getBBInfo(SrcBB);
1132 // If only one out-edge, the edge profile count should be the same as BB
1133 // profile count.
1134 if (SrcInfo.CountValid && SrcInfo.OutEdges.size() == 1)
1135 setEdgeCount(E.get(), SrcInfo.CountValue);
1136 else {
1137 const BasicBlock *DestBB = E->DestBB;
1138 UseBBInfo &DestInfo = getBBInfo(DestBB);
1139 // If only one in-edge, the edge profile count should be the same as BB
1140 // profile count.
1141 if (DestInfo.CountValid && DestInfo.InEdges.size() == 1)
1142 setEdgeCount(E.get(), DestInfo.CountValue);
1144 if (E->CountValid)
1145 continue;
1146 // E's count should have been set from profile. If not, this meenas E skips
1147 // the instrumentation. We set the count to 0.
1148 setEdgeCount(E.get(), 0);
1150 return true;
1153 // Set the count value for the unknown edge. There should be one and only one
1154 // unknown edge in Edges vector.
1155 void PGOUseFunc::setEdgeCount(DirectEdges &Edges, uint64_t Value) {
1156 for (auto &E : Edges) {
1157 if (E->CountValid)
1158 continue;
1159 E->setEdgeCount(Value);
1161 getBBInfo(E->SrcBB).UnknownCountOutEdge--;
1162 getBBInfo(E->DestBB).UnknownCountInEdge--;
1163 return;
1165 llvm_unreachable("Cannot find the unknown count edge");
1168 // Read the profile from ProfileFileName and assign the value to the
1169 // instrumented BB and the edges. This function also updates ProgramMaxCount.
1170 // Return true if the profile are successfully read, and false on errors.
1171 bool PGOUseFunc::readCounters(IndexedInstrProfReader *PGOReader, bool &AllZeros) {
1172 auto &Ctx = M->getContext();
1173 Expected<InstrProfRecord> Result =
1174 PGOReader->getInstrProfRecord(FuncInfo.FuncName, FuncInfo.FunctionHash);
1175 if (Error E = Result.takeError()) {
1176 handleAllErrors(std::move(E), [&](const InstrProfError &IPE) {
1177 auto Err = IPE.get();
1178 bool SkipWarning = false;
1179 LLVM_DEBUG(dbgs() << "Error in reading profile for Func "
1180 << FuncInfo.FuncName << ": ");
1181 if (Err == instrprof_error::unknown_function) {
1182 IsCS ? NumOfCSPGOMissing++ : NumOfPGOMissing++;
1183 SkipWarning = !PGOWarnMissing;
1184 LLVM_DEBUG(dbgs() << "unknown function");
1185 } else if (Err == instrprof_error::hash_mismatch ||
1186 Err == instrprof_error::malformed) {
1187 IsCS ? NumOfCSPGOMismatch++ : NumOfPGOMismatch++;
1188 SkipWarning =
1189 NoPGOWarnMismatch ||
1190 (NoPGOWarnMismatchComdat &&
1191 (F.hasComdat() ||
1192 F.getLinkage() == GlobalValue::AvailableExternallyLinkage));
1193 LLVM_DEBUG(dbgs() << "hash mismatch (skip=" << SkipWarning << ")");
1196 LLVM_DEBUG(dbgs() << " IsCS=" << IsCS << "\n");
1197 if (SkipWarning)
1198 return;
1200 std::string Msg = IPE.message() + std::string(" ") + F.getName().str() +
1201 std::string(" Hash = ") +
1202 std::to_string(FuncInfo.FunctionHash);
1204 Ctx.diagnose(
1205 DiagnosticInfoPGOProfile(M->getName().data(), Msg, DS_Warning));
1207 return false;
1209 ProfileRecord = std::move(Result.get());
1210 std::vector<uint64_t> &CountFromProfile = ProfileRecord.Counts;
1212 IsCS ? NumOfCSPGOFunc++ : NumOfPGOFunc++;
1213 LLVM_DEBUG(dbgs() << CountFromProfile.size() << " counts\n");
1214 uint64_t ValueSum = 0;
1215 for (unsigned I = 0, S = CountFromProfile.size(); I < S; I++) {
1216 LLVM_DEBUG(dbgs() << " " << I << ": " << CountFromProfile[I] << "\n");
1217 ValueSum += CountFromProfile[I];
1219 AllZeros = (ValueSum == 0);
1221 LLVM_DEBUG(dbgs() << "SUM = " << ValueSum << "\n");
1223 getBBInfo(nullptr).UnknownCountOutEdge = 2;
1224 getBBInfo(nullptr).UnknownCountInEdge = 2;
1226 if (!setInstrumentedCounts(CountFromProfile)) {
1227 LLVM_DEBUG(
1228 dbgs() << "Inconsistent number of counts, skipping this function");
1229 Ctx.diagnose(DiagnosticInfoPGOProfile(
1230 M->getName().data(),
1231 Twine("Inconsistent number of counts in ") + F.getName().str()
1232 + Twine(": the profile may be stale or there is a function name collision."),
1233 DS_Warning));
1234 return false;
1236 ProgramMaxCount = PGOReader->getMaximumFunctionCount(IsCS);
1237 return true;
1240 // Populate the counters from instrumented BBs to all BBs.
1241 // In the end of this operation, all BBs should have a valid count value.
1242 void PGOUseFunc::populateCounters() {
1243 bool Changes = true;
1244 unsigned NumPasses = 0;
1245 while (Changes) {
1246 NumPasses++;
1247 Changes = false;
1249 // For efficient traversal, it's better to start from the end as most
1250 // of the instrumented edges are at the end.
1251 for (auto &BB : reverse(F)) {
1252 UseBBInfo *Count = findBBInfo(&BB);
1253 if (Count == nullptr)
1254 continue;
1255 if (!Count->CountValid) {
1256 if (Count->UnknownCountOutEdge == 0) {
1257 Count->CountValue = sumEdgeCount(Count->OutEdges);
1258 Count->CountValid = true;
1259 Changes = true;
1260 } else if (Count->UnknownCountInEdge == 0) {
1261 Count->CountValue = sumEdgeCount(Count->InEdges);
1262 Count->CountValid = true;
1263 Changes = true;
1266 if (Count->CountValid) {
1267 if (Count->UnknownCountOutEdge == 1) {
1268 uint64_t Total = 0;
1269 uint64_t OutSum = sumEdgeCount(Count->OutEdges);
1270 // If the one of the successor block can early terminate (no-return),
1271 // we can end up with situation where out edge sum count is larger as
1272 // the source BB's count is collected by a post-dominated block.
1273 if (Count->CountValue > OutSum)
1274 Total = Count->CountValue - OutSum;
1275 setEdgeCount(Count->OutEdges, Total);
1276 Changes = true;
1278 if (Count->UnknownCountInEdge == 1) {
1279 uint64_t Total = 0;
1280 uint64_t InSum = sumEdgeCount(Count->InEdges);
1281 if (Count->CountValue > InSum)
1282 Total = Count->CountValue - InSum;
1283 setEdgeCount(Count->InEdges, Total);
1284 Changes = true;
1290 LLVM_DEBUG(dbgs() << "Populate counts in " << NumPasses << " passes.\n");
1291 #ifndef NDEBUG
1292 // Assert every BB has a valid counter.
1293 for (auto &BB : F) {
1294 auto BI = findBBInfo(&BB);
1295 if (BI == nullptr)
1296 continue;
1297 assert(BI->CountValid && "BB count is not valid");
1299 #endif
1300 uint64_t FuncEntryCount = getBBInfo(&*F.begin()).CountValue;
1301 F.setEntryCount(ProfileCount(FuncEntryCount, Function::PCT_Real));
1302 uint64_t FuncMaxCount = FuncEntryCount;
1303 for (auto &BB : F) {
1304 auto BI = findBBInfo(&BB);
1305 if (BI == nullptr)
1306 continue;
1307 FuncMaxCount = std::max(FuncMaxCount, BI->CountValue);
1309 markFunctionAttributes(FuncEntryCount, FuncMaxCount);
1311 // Now annotate select instructions
1312 FuncInfo.SIVisitor.annotateSelects(F, this, &CountPosition);
1313 assert(CountPosition == ProfileCountSize);
1315 LLVM_DEBUG(FuncInfo.dumpInfo("after reading profile."));
1318 // Assign the scaled count values to the BB with multiple out edges.
1319 void PGOUseFunc::setBranchWeights() {
1320 // Generate MD_prof metadata for every branch instruction.
1321 LLVM_DEBUG(dbgs() << "\nSetting branch weights for func " << F.getName()
1322 << " IsCS=" << IsCS << "\n");
1323 for (auto &BB : F) {
1324 Instruction *TI = BB.getTerminator();
1325 if (TI->getNumSuccessors() < 2)
1326 continue;
1327 if (!(isa<BranchInst>(TI) || isa<SwitchInst>(TI) ||
1328 isa<IndirectBrInst>(TI)))
1329 continue;
1331 if (getBBInfo(&BB).CountValue == 0)
1332 continue;
1334 // We have a non-zero Branch BB.
1335 const UseBBInfo &BBCountInfo = getBBInfo(&BB);
1336 unsigned Size = BBCountInfo.OutEdges.size();
1337 SmallVector<uint64_t, 2> EdgeCounts(Size, 0);
1338 uint64_t MaxCount = 0;
1339 for (unsigned s = 0; s < Size; s++) {
1340 const PGOUseEdge *E = BBCountInfo.OutEdges[s];
1341 const BasicBlock *SrcBB = E->SrcBB;
1342 const BasicBlock *DestBB = E->DestBB;
1343 if (DestBB == nullptr)
1344 continue;
1345 unsigned SuccNum = GetSuccessorNumber(SrcBB, DestBB);
1346 uint64_t EdgeCount = E->CountValue;
1347 if (EdgeCount > MaxCount)
1348 MaxCount = EdgeCount;
1349 EdgeCounts[SuccNum] = EdgeCount;
1351 setProfMetadata(M, TI, EdgeCounts, MaxCount);
1355 static bool isIndirectBrTarget(BasicBlock *BB) {
1356 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1357 if (isa<IndirectBrInst>((*PI)->getTerminator()))
1358 return true;
1360 return false;
1363 void PGOUseFunc::annotateIrrLoopHeaderWeights() {
1364 LLVM_DEBUG(dbgs() << "\nAnnotating irreducible loop header weights.\n");
1365 // Find irr loop headers
1366 for (auto &BB : F) {
1367 // As a heuristic also annotate indrectbr targets as they have a high chance
1368 // to become an irreducible loop header after the indirectbr tail
1369 // duplication.
1370 if (BFI->isIrrLoopHeader(&BB) || isIndirectBrTarget(&BB)) {
1371 Instruction *TI = BB.getTerminator();
1372 const UseBBInfo &BBCountInfo = getBBInfo(&BB);
1373 setIrrLoopHeaderMetadata(M, TI, BBCountInfo.CountValue);
1378 void SelectInstVisitor::instrumentOneSelectInst(SelectInst &SI) {
1379 Module *M = F.getParent();
1380 IRBuilder<> Builder(&SI);
1381 Type *Int64Ty = Builder.getInt64Ty();
1382 Type *I8PtrTy = Builder.getInt8PtrTy();
1383 auto *Step = Builder.CreateZExt(SI.getCondition(), Int64Ty);
1384 Builder.CreateCall(
1385 Intrinsic::getDeclaration(M, Intrinsic::instrprof_increment_step),
1386 {ConstantExpr::getBitCast(FuncNameVar, I8PtrTy),
1387 Builder.getInt64(FuncHash), Builder.getInt32(TotalNumCtrs),
1388 Builder.getInt32(*CurCtrIdx), Step});
1389 ++(*CurCtrIdx);
1392 void SelectInstVisitor::annotateOneSelectInst(SelectInst &SI) {
1393 std::vector<uint64_t> &CountFromProfile = UseFunc->getProfileRecord().Counts;
1394 assert(*CurCtrIdx < CountFromProfile.size() &&
1395 "Out of bound access of counters");
1396 uint64_t SCounts[2];
1397 SCounts[0] = CountFromProfile[*CurCtrIdx]; // True count
1398 ++(*CurCtrIdx);
1399 uint64_t TotalCount = 0;
1400 auto BI = UseFunc->findBBInfo(SI.getParent());
1401 if (BI != nullptr)
1402 TotalCount = BI->CountValue;
1403 // False Count
1404 SCounts[1] = (TotalCount > SCounts[0] ? TotalCount - SCounts[0] : 0);
1405 uint64_t MaxCount = std::max(SCounts[0], SCounts[1]);
1406 if (MaxCount)
1407 setProfMetadata(F.getParent(), &SI, SCounts, MaxCount);
1410 void SelectInstVisitor::visitSelectInst(SelectInst &SI) {
1411 if (!PGOInstrSelect)
1412 return;
1413 // FIXME: do not handle this yet.
1414 if (SI.getCondition()->getType()->isVectorTy())
1415 return;
1417 switch (Mode) {
1418 case VM_counting:
1419 NSIs++;
1420 return;
1421 case VM_instrument:
1422 instrumentOneSelectInst(SI);
1423 return;
1424 case VM_annotate:
1425 annotateOneSelectInst(SI);
1426 return;
1429 llvm_unreachable("Unknown visiting mode");
1432 void MemIntrinsicVisitor::instrumentOneMemIntrinsic(MemIntrinsic &MI) {
1433 Module *M = F.getParent();
1434 IRBuilder<> Builder(&MI);
1435 Type *Int64Ty = Builder.getInt64Ty();
1436 Type *I8PtrTy = Builder.getInt8PtrTy();
1437 Value *Length = MI.getLength();
1438 assert(!isa<ConstantInt>(Length));
1439 Builder.CreateCall(
1440 Intrinsic::getDeclaration(M, Intrinsic::instrprof_value_profile),
1441 {ConstantExpr::getBitCast(FuncNameVar, I8PtrTy),
1442 Builder.getInt64(FuncHash), Builder.CreateZExtOrTrunc(Length, Int64Ty),
1443 Builder.getInt32(IPVK_MemOPSize), Builder.getInt32(CurCtrId)});
1444 ++CurCtrId;
1447 void MemIntrinsicVisitor::visitMemIntrinsic(MemIntrinsic &MI) {
1448 if (!PGOInstrMemOP)
1449 return;
1450 Value *Length = MI.getLength();
1451 // Not instrument constant length calls.
1452 if (dyn_cast<ConstantInt>(Length))
1453 return;
1455 switch (Mode) {
1456 case VM_counting:
1457 NMemIs++;
1458 return;
1459 case VM_instrument:
1460 instrumentOneMemIntrinsic(MI);
1461 return;
1462 case VM_annotate:
1463 Candidates.push_back(&MI);
1464 return;
1466 llvm_unreachable("Unknown visiting mode");
1469 // Traverse all valuesites and annotate the instructions for all value kind.
1470 void PGOUseFunc::annotateValueSites() {
1471 if (DisableValueProfiling)
1472 return;
1474 // Create the PGOFuncName meta data.
1475 createPGOFuncNameMetadata(F, FuncInfo.FuncName);
1477 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
1478 annotateValueSites(Kind);
1481 static const char *ValueProfKindDescr[] = {
1482 #define VALUE_PROF_KIND(Enumerator, Value, Descr) Descr,
1483 #include "llvm/ProfileData/InstrProfData.inc"
1486 // Annotate the instructions for a specific value kind.
1487 void PGOUseFunc::annotateValueSites(uint32_t Kind) {
1488 assert(Kind <= IPVK_Last);
1489 unsigned ValueSiteIndex = 0;
1490 auto &ValueSites = FuncInfo.ValueSites[Kind];
1491 unsigned NumValueSites = ProfileRecord.getNumValueSites(Kind);
1492 if (NumValueSites != ValueSites.size()) {
1493 auto &Ctx = M->getContext();
1494 Ctx.diagnose(DiagnosticInfoPGOProfile(
1495 M->getName().data(),
1496 Twine("Inconsistent number of value sites for ") +
1497 Twine(ValueProfKindDescr[Kind]) +
1498 Twine(" profiling in \"") + F.getName().str() +
1499 Twine("\", possibly due to the use of a stale profile."),
1500 DS_Warning));
1501 return;
1504 for (auto &I : ValueSites) {
1505 LLVM_DEBUG(dbgs() << "Read one value site profile (kind = " << Kind
1506 << "): Index = " << ValueSiteIndex << " out of "
1507 << NumValueSites << "\n");
1508 annotateValueSite(*M, *I, ProfileRecord,
1509 static_cast<InstrProfValueKind>(Kind), ValueSiteIndex,
1510 Kind == IPVK_MemOPSize ? MaxNumMemOPAnnotations
1511 : MaxNumAnnotations);
1512 ValueSiteIndex++;
1516 // Collect the set of members for each Comdat in module M and store
1517 // in ComdatMembers.
1518 static void collectComdatMembers(
1519 Module &M,
1520 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers) {
1521 if (!DoComdatRenaming)
1522 return;
1523 for (Function &F : M)
1524 if (Comdat *C = F.getComdat())
1525 ComdatMembers.insert(std::make_pair(C, &F));
1526 for (GlobalVariable &GV : M.globals())
1527 if (Comdat *C = GV.getComdat())
1528 ComdatMembers.insert(std::make_pair(C, &GV));
1529 for (GlobalAlias &GA : M.aliases())
1530 if (Comdat *C = GA.getComdat())
1531 ComdatMembers.insert(std::make_pair(C, &GA));
1534 static bool InstrumentAllFunctions(
1535 Module &M, function_ref<BranchProbabilityInfo *(Function &)> LookupBPI,
1536 function_ref<BlockFrequencyInfo *(Function &)> LookupBFI, bool IsCS) {
1537 // For the context-sensitve instrumentation, we should have a separated pass
1538 // (before LTO/ThinLTO linking) to create these variables.
1539 if (!IsCS)
1540 createIRLevelProfileFlagVar(M, /* IsCS */ false);
1541 std::unordered_multimap<Comdat *, GlobalValue *> ComdatMembers;
1542 collectComdatMembers(M, ComdatMembers);
1544 for (auto &F : M) {
1545 if (F.isDeclaration())
1546 continue;
1547 auto *BPI = LookupBPI(F);
1548 auto *BFI = LookupBFI(F);
1549 instrumentOneFunc(F, &M, BPI, BFI, ComdatMembers, IsCS);
1551 return true;
1554 PreservedAnalyses
1555 PGOInstrumentationGenCreateVar::run(Module &M, ModuleAnalysisManager &AM) {
1556 createProfileFileNameVar(M, CSInstrName);
1557 createIRLevelProfileFlagVar(M, /* IsCS */ true);
1558 return PreservedAnalyses::all();
1561 bool PGOInstrumentationGenLegacyPass::runOnModule(Module &M) {
1562 if (skipModule(M))
1563 return false;
1565 auto LookupBPI = [this](Function &F) {
1566 return &this->getAnalysis<BranchProbabilityInfoWrapperPass>(F).getBPI();
1568 auto LookupBFI = [this](Function &F) {
1569 return &this->getAnalysis<BlockFrequencyInfoWrapperPass>(F).getBFI();
1571 return InstrumentAllFunctions(M, LookupBPI, LookupBFI, IsCS);
1574 PreservedAnalyses PGOInstrumentationGen::run(Module &M,
1575 ModuleAnalysisManager &AM) {
1576 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
1577 auto LookupBPI = [&FAM](Function &F) {
1578 return &FAM.getResult<BranchProbabilityAnalysis>(F);
1581 auto LookupBFI = [&FAM](Function &F) {
1582 return &FAM.getResult<BlockFrequencyAnalysis>(F);
1585 if (!InstrumentAllFunctions(M, LookupBPI, LookupBFI, IsCS))
1586 return PreservedAnalyses::all();
1588 return PreservedAnalyses::none();
1591 static bool annotateAllFunctions(
1592 Module &M, StringRef ProfileFileName, StringRef ProfileRemappingFileName,
1593 function_ref<BranchProbabilityInfo *(Function &)> LookupBPI,
1594 function_ref<BlockFrequencyInfo *(Function &)> LookupBFI,
1595 ProfileSummaryInfo *PSI, bool IsCS) {
1596 LLVM_DEBUG(dbgs() << "Read in profile counters: ");
1597 auto &Ctx = M.getContext();
1598 // Read the counter array from file.
1599 auto ReaderOrErr =
1600 IndexedInstrProfReader::create(ProfileFileName, ProfileRemappingFileName);
1601 if (Error E = ReaderOrErr.takeError()) {
1602 handleAllErrors(std::move(E), [&](const ErrorInfoBase &EI) {
1603 Ctx.diagnose(
1604 DiagnosticInfoPGOProfile(ProfileFileName.data(), EI.message()));
1606 return false;
1609 std::unique_ptr<IndexedInstrProfReader> PGOReader =
1610 std::move(ReaderOrErr.get());
1611 if (!PGOReader) {
1612 Ctx.diagnose(DiagnosticInfoPGOProfile(ProfileFileName.data(),
1613 StringRef("Cannot get PGOReader")));
1614 return false;
1616 if (!PGOReader->hasCSIRLevelProfile() && IsCS)
1617 return false;
1619 // TODO: might need to change the warning once the clang option is finalized.
1620 if (!PGOReader->isIRLevelProfile()) {
1621 Ctx.diagnose(DiagnosticInfoPGOProfile(
1622 ProfileFileName.data(), "Not an IR level instrumentation profile"));
1623 return false;
1626 // Add the profile summary (read from the header of the indexed summary) here
1627 // so that we can use it below when reading counters (which checks if the
1628 // function should be marked with a cold or inlinehint attribute).
1629 M.setProfileSummary(PGOReader->getSummary(IsCS).getMD(M.getContext()),
1630 IsCS ? ProfileSummary::PSK_CSInstr
1631 : ProfileSummary::PSK_Instr);
1633 std::unordered_multimap<Comdat *, GlobalValue *> ComdatMembers;
1634 collectComdatMembers(M, ComdatMembers);
1635 std::vector<Function *> HotFunctions;
1636 std::vector<Function *> ColdFunctions;
1637 for (auto &F : M) {
1638 if (F.isDeclaration())
1639 continue;
1640 auto *BPI = LookupBPI(F);
1641 auto *BFI = LookupBFI(F);
1642 // Split indirectbr critical edges here before computing the MST rather than
1643 // later in getInstrBB() to avoid invalidating it.
1644 SplitIndirectBrCriticalEdges(F, BPI, BFI);
1645 PGOUseFunc Func(F, &M, ComdatMembers, BPI, BFI, PSI, IsCS);
1646 bool AllZeros = false;
1647 if (!Func.readCounters(PGOReader.get(), AllZeros))
1648 continue;
1649 if (AllZeros) {
1650 F.setEntryCount(ProfileCount(0, Function::PCT_Real));
1651 if (Func.getProgramMaxCount() != 0)
1652 ColdFunctions.push_back(&F);
1653 continue;
1655 Func.populateCounters();
1656 Func.setBranchWeights();
1657 Func.annotateValueSites();
1658 Func.annotateIrrLoopHeaderWeights();
1659 PGOUseFunc::FuncFreqAttr FreqAttr = Func.getFuncFreqAttr();
1660 if (FreqAttr == PGOUseFunc::FFA_Cold)
1661 ColdFunctions.push_back(&F);
1662 else if (FreqAttr == PGOUseFunc::FFA_Hot)
1663 HotFunctions.push_back(&F);
1664 if (PGOViewCounts != PGOVCT_None &&
1665 (ViewBlockFreqFuncName.empty() ||
1666 F.getName().equals(ViewBlockFreqFuncName))) {
1667 LoopInfo LI{DominatorTree(F)};
1668 std::unique_ptr<BranchProbabilityInfo> NewBPI =
1669 std::make_unique<BranchProbabilityInfo>(F, LI);
1670 std::unique_ptr<BlockFrequencyInfo> NewBFI =
1671 std::make_unique<BlockFrequencyInfo>(F, *NewBPI, LI);
1672 if (PGOViewCounts == PGOVCT_Graph)
1673 NewBFI->view();
1674 else if (PGOViewCounts == PGOVCT_Text) {
1675 dbgs() << "pgo-view-counts: " << Func.getFunc().getName() << "\n";
1676 NewBFI->print(dbgs());
1679 if (PGOViewRawCounts != PGOVCT_None &&
1680 (ViewBlockFreqFuncName.empty() ||
1681 F.getName().equals(ViewBlockFreqFuncName))) {
1682 if (PGOViewRawCounts == PGOVCT_Graph)
1683 if (ViewBlockFreqFuncName.empty())
1684 WriteGraph(&Func, Twine("PGORawCounts_") + Func.getFunc().getName());
1685 else
1686 ViewGraph(&Func, Twine("PGORawCounts_") + Func.getFunc().getName());
1687 else if (PGOViewRawCounts == PGOVCT_Text) {
1688 dbgs() << "pgo-view-raw-counts: " << Func.getFunc().getName() << "\n";
1689 Func.dumpInfo();
1694 // Set function hotness attribute from the profile.
1695 // We have to apply these attributes at the end because their presence
1696 // can affect the BranchProbabilityInfo of any callers, resulting in an
1697 // inconsistent MST between prof-gen and prof-use.
1698 for (auto &F : HotFunctions) {
1699 F->addFnAttr(Attribute::InlineHint);
1700 LLVM_DEBUG(dbgs() << "Set inline attribute to function: " << F->getName()
1701 << "\n");
1703 for (auto &F : ColdFunctions) {
1704 F->addFnAttr(Attribute::Cold);
1705 LLVM_DEBUG(dbgs() << "Set cold attribute to function: " << F->getName()
1706 << "\n");
1708 return true;
1711 PGOInstrumentationUse::PGOInstrumentationUse(std::string Filename,
1712 std::string RemappingFilename,
1713 bool IsCS)
1714 : ProfileFileName(std::move(Filename)),
1715 ProfileRemappingFileName(std::move(RemappingFilename)), IsCS(IsCS) {
1716 if (!PGOTestProfileFile.empty())
1717 ProfileFileName = PGOTestProfileFile;
1718 if (!PGOTestProfileRemappingFile.empty())
1719 ProfileRemappingFileName = PGOTestProfileRemappingFile;
1722 PreservedAnalyses PGOInstrumentationUse::run(Module &M,
1723 ModuleAnalysisManager &AM) {
1725 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
1726 auto LookupBPI = [&FAM](Function &F) {
1727 return &FAM.getResult<BranchProbabilityAnalysis>(F);
1730 auto LookupBFI = [&FAM](Function &F) {
1731 return &FAM.getResult<BlockFrequencyAnalysis>(F);
1734 auto *PSI = &AM.getResult<ProfileSummaryAnalysis>(M);
1736 if (!annotateAllFunctions(M, ProfileFileName, ProfileRemappingFileName,
1737 LookupBPI, LookupBFI, PSI, IsCS))
1738 return PreservedAnalyses::all();
1740 return PreservedAnalyses::none();
1743 bool PGOInstrumentationUseLegacyPass::runOnModule(Module &M) {
1744 if (skipModule(M))
1745 return false;
1747 auto LookupBPI = [this](Function &F) {
1748 return &this->getAnalysis<BranchProbabilityInfoWrapperPass>(F).getBPI();
1750 auto LookupBFI = [this](Function &F) {
1751 return &this->getAnalysis<BlockFrequencyInfoWrapperPass>(F).getBFI();
1754 auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
1755 return annotateAllFunctions(M, ProfileFileName, "", LookupBPI, LookupBFI, PSI,
1756 IsCS);
1759 static std::string getSimpleNodeName(const BasicBlock *Node) {
1760 if (!Node->getName().empty())
1761 return Node->getName();
1763 std::string SimpleNodeName;
1764 raw_string_ostream OS(SimpleNodeName);
1765 Node->printAsOperand(OS, false);
1766 return OS.str();
1769 void llvm::setProfMetadata(Module *M, Instruction *TI,
1770 ArrayRef<uint64_t> EdgeCounts,
1771 uint64_t MaxCount) {
1772 MDBuilder MDB(M->getContext());
1773 assert(MaxCount > 0 && "Bad max count");
1774 uint64_t Scale = calculateCountScale(MaxCount);
1775 SmallVector<unsigned, 4> Weights;
1776 for (const auto &ECI : EdgeCounts)
1777 Weights.push_back(scaleBranchCount(ECI, Scale));
1779 LLVM_DEBUG(dbgs() << "Weight is: "; for (const auto &W
1780 : Weights) {
1781 dbgs() << W << " ";
1782 } dbgs() << "\n";);
1784 misexpect::verifyMisExpect(TI, Weights, TI->getContext());
1786 TI->setMetadata(LLVMContext::MD_prof, MDB.createBranchWeights(Weights));
1787 if (EmitBranchProbability) {
1788 std::string BrCondStr = getBranchCondString(TI);
1789 if (BrCondStr.empty())
1790 return;
1792 uint64_t WSum =
1793 std::accumulate(Weights.begin(), Weights.end(), (uint64_t)0,
1794 [](uint64_t w1, uint64_t w2) { return w1 + w2; });
1795 uint64_t TotalCount =
1796 std::accumulate(EdgeCounts.begin(), EdgeCounts.end(), (uint64_t)0,
1797 [](uint64_t c1, uint64_t c2) { return c1 + c2; });
1798 Scale = calculateCountScale(WSum);
1799 BranchProbability BP(scaleBranchCount(Weights[0], Scale),
1800 scaleBranchCount(WSum, Scale));
1801 std::string BranchProbStr;
1802 raw_string_ostream OS(BranchProbStr);
1803 OS << BP;
1804 OS << " (total count : " << TotalCount << ")";
1805 OS.flush();
1806 Function *F = TI->getParent()->getParent();
1807 OptimizationRemarkEmitter ORE(F);
1808 ORE.emit([&]() {
1809 return OptimizationRemark(DEBUG_TYPE, "pgo-instrumentation", TI)
1810 << BrCondStr << " is true with probability : " << BranchProbStr;
1815 namespace llvm {
1817 void setIrrLoopHeaderMetadata(Module *M, Instruction *TI, uint64_t Count) {
1818 MDBuilder MDB(M->getContext());
1819 TI->setMetadata(llvm::LLVMContext::MD_irr_loop,
1820 MDB.createIrrLoopHeaderWeight(Count));
1823 template <> struct GraphTraits<PGOUseFunc *> {
1824 using NodeRef = const BasicBlock *;
1825 using ChildIteratorType = succ_const_iterator;
1826 using nodes_iterator = pointer_iterator<Function::const_iterator>;
1828 static NodeRef getEntryNode(const PGOUseFunc *G) {
1829 return &G->getFunc().front();
1832 static ChildIteratorType child_begin(const NodeRef N) {
1833 return succ_begin(N);
1836 static ChildIteratorType child_end(const NodeRef N) { return succ_end(N); }
1838 static nodes_iterator nodes_begin(const PGOUseFunc *G) {
1839 return nodes_iterator(G->getFunc().begin());
1842 static nodes_iterator nodes_end(const PGOUseFunc *G) {
1843 return nodes_iterator(G->getFunc().end());
1847 template <> struct DOTGraphTraits<PGOUseFunc *> : DefaultDOTGraphTraits {
1848 explicit DOTGraphTraits(bool isSimple = false)
1849 : DefaultDOTGraphTraits(isSimple) {}
1851 static std::string getGraphName(const PGOUseFunc *G) {
1852 return G->getFunc().getName();
1855 std::string getNodeLabel(const BasicBlock *Node, const PGOUseFunc *Graph) {
1856 std::string Result;
1857 raw_string_ostream OS(Result);
1859 OS << getSimpleNodeName(Node) << ":\\l";
1860 UseBBInfo *BI = Graph->findBBInfo(Node);
1861 OS << "Count : ";
1862 if (BI && BI->CountValid)
1863 OS << BI->CountValue << "\\l";
1864 else
1865 OS << "Unknown\\l";
1867 if (!PGOInstrSelect)
1868 return Result;
1870 for (auto BI = Node->begin(); BI != Node->end(); ++BI) {
1871 auto *I = &*BI;
1872 if (!isa<SelectInst>(I))
1873 continue;
1874 // Display scaled counts for SELECT instruction:
1875 OS << "SELECT : { T = ";
1876 uint64_t TC, FC;
1877 bool HasProf = I->extractProfMetadata(TC, FC);
1878 if (!HasProf)
1879 OS << "Unknown, F = Unknown }\\l";
1880 else
1881 OS << TC << ", F = " << FC << " }\\l";
1883 return Result;
1887 } // end namespace llvm