[AMDGPU] Test codegen'ing True16 additions.
[llvm-project.git] / llvm / tools / llvm-profgen / ProfiledBinary.h
bloba6d78c661cc1c3162624cb19273d1a0afba61cc8
1 //===-- ProfiledBinary.h - Binary decoder -----------------------*- C++ -*-===//
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 //===----------------------------------------------------------------------===//
9 #ifndef LLVM_TOOLS_LLVM_PROFGEN_PROFILEDBINARY_H
10 #define LLVM_TOOLS_LLVM_PROFGEN_PROFILEDBINARY_H
12 #include "CallContext.h"
13 #include "ErrorHandling.h"
14 #include "llvm/ADT/DenseMap.h"
15 #include "llvm/ADT/StringRef.h"
16 #include "llvm/ADT/StringSet.h"
17 #include "llvm/DebugInfo/DWARF/DWARFContext.h"
18 #include "llvm/DebugInfo/Symbolize/Symbolize.h"
19 #include "llvm/MC/MCAsmInfo.h"
20 #include "llvm/MC/MCContext.h"
21 #include "llvm/MC/MCDisassembler/MCDisassembler.h"
22 #include "llvm/MC/MCInst.h"
23 #include "llvm/MC/MCInstPrinter.h"
24 #include "llvm/MC/MCInstrAnalysis.h"
25 #include "llvm/MC/MCInstrInfo.h"
26 #include "llvm/MC/MCObjectFileInfo.h"
27 #include "llvm/MC/MCPseudoProbe.h"
28 #include "llvm/MC/MCRegisterInfo.h"
29 #include "llvm/MC/MCSubtargetInfo.h"
30 #include "llvm/MC/MCTargetOptions.h"
31 #include "llvm/Object/ELFObjectFile.h"
32 #include "llvm/ProfileData/SampleProf.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Support/Path.h"
35 #include "llvm/Transforms/IPO/SampleContextTracker.h"
36 #include <list>
37 #include <map>
38 #include <set>
39 #include <sstream>
40 #include <string>
41 #include <unordered_map>
42 #include <unordered_set>
43 #include <vector>
45 namespace llvm {
46 extern cl::opt<bool> EnableCSPreInliner;
47 extern cl::opt<bool> UseContextCostForPreInliner;
48 } // namespace llvm
50 using namespace llvm;
51 using namespace sampleprof;
52 using namespace llvm::object;
54 namespace llvm {
55 namespace sampleprof {
57 class ProfiledBinary;
58 class MissingFrameInferrer;
60 struct InstructionPointer {
61 const ProfiledBinary *Binary;
62 // Address of the executable segment of the binary.
63 uint64_t Address;
64 // Index to the sorted code address array of the binary.
65 uint64_t Index = 0;
66 InstructionPointer(const ProfiledBinary *Binary, uint64_t Address,
67 bool RoundToNext = false);
68 bool advance();
69 bool backward();
70 void update(uint64_t Addr);
73 // The special frame addresses.
74 enum SpecialFrameAddr {
75 // Dummy root of frame trie.
76 DummyRoot = 0,
77 // Represent all the addresses outside of current binary.
78 // This's also used to indicate the call stack should be truncated since this
79 // isn't a real call context the compiler will see.
80 ExternalAddr = 1,
83 using RangesTy = std::vector<std::pair<uint64_t, uint64_t>>;
85 struct BinaryFunction {
86 StringRef FuncName;
87 // End of range is an exclusive bound.
88 RangesTy Ranges;
90 uint64_t getFuncSize() {
91 uint64_t Sum = 0;
92 for (auto &R : Ranges) {
93 Sum += R.second - R.first;
95 return Sum;
99 // Info about function range. A function can be split into multiple
100 // non-continuous ranges, each range corresponds to one FuncRange.
101 struct FuncRange {
102 uint64_t StartAddress;
103 // EndAddress is an exclusive bound.
104 uint64_t EndAddress;
105 // Function the range belongs to
106 BinaryFunction *Func;
107 // Whether the start address is the real entry of the function.
108 bool IsFuncEntry = false;
110 StringRef getFuncName() { return Func->FuncName; }
113 // PrologEpilog address tracker, used to filter out broken stack samples
114 // Currently we use a heuristic size (two) to infer prolog and epilog
115 // based on the start address and return address. In the future,
116 // we will switch to Dwarf CFI based tracker
117 struct PrologEpilogTracker {
118 // A set of prolog and epilog addresses. Used by virtual unwinding.
119 std::unordered_set<uint64_t> PrologEpilogSet;
120 ProfiledBinary *Binary;
121 PrologEpilogTracker(ProfiledBinary *Bin) : Binary(Bin){};
123 // Take the two addresses from the start of function as prolog
124 void
125 inferPrologAddresses(std::map<uint64_t, FuncRange> &FuncStartAddressMap) {
126 for (auto I : FuncStartAddressMap) {
127 PrologEpilogSet.insert(I.first);
128 InstructionPointer IP(Binary, I.first);
129 if (!IP.advance())
130 break;
131 PrologEpilogSet.insert(IP.Address);
135 // Take the last two addresses before the return address as epilog
136 void inferEpilogAddresses(std::unordered_set<uint64_t> &RetAddrs) {
137 for (auto Addr : RetAddrs) {
138 PrologEpilogSet.insert(Addr);
139 InstructionPointer IP(Binary, Addr);
140 if (!IP.backward())
141 break;
142 PrologEpilogSet.insert(IP.Address);
147 // Track function byte size under different context (outlined version as well as
148 // various inlined versions). It also provides query support to get function
149 // size with the best matching context, which is used to help pre-inliner use
150 // accurate post-optimization size to make decisions.
151 // TODO: If an inlinee is completely optimized away, ideally we should have zero
152 // for its context size, currently we would misss such context since it doesn't
153 // have instructions. To fix this, we need to mark all inlinee with entry probe
154 // but without instructions as having zero size.
155 class BinarySizeContextTracker {
156 public:
157 // Add instruction with given size to a context
158 void addInstructionForContext(const SampleContextFrameVector &Context,
159 uint32_t InstrSize);
161 // Get function size with a specific context. When there's no exact match
162 // for the given context, try to retrieve the size of that function from
163 // closest matching context.
164 uint32_t getFuncSizeForContext(const ContextTrieNode *Context);
166 // For inlinees that are full optimized away, we can establish zero size using
167 // their remaining probes.
168 void trackInlineesOptimizedAway(MCPseudoProbeDecoder &ProbeDecoder);
170 using ProbeFrameStack = SmallVector<std::pair<StringRef, uint32_t>>;
171 void trackInlineesOptimizedAway(MCPseudoProbeDecoder &ProbeDecoder,
172 MCDecodedPseudoProbeInlineTree &ProbeNode,
173 ProbeFrameStack &Context);
175 void dump() { RootContext.dumpTree(); }
177 private:
178 // Root node for context trie tree, node that this is a reverse context trie
179 // with callee as parent and caller as child. This way we can traverse from
180 // root to find the best/longest matching context if an exact match does not
181 // exist. It gives us the best possible estimate for function's post-inline,
182 // post-optimization byte size.
183 ContextTrieNode RootContext;
186 using AddressRange = std::pair<uint64_t, uint64_t>;
188 class ProfiledBinary {
189 // Absolute path of the executable binary.
190 std::string Path;
191 // Path of the debug info binary.
192 std::string DebugBinaryPath;
193 // The target triple.
194 Triple TheTriple;
195 // Path of symbolizer path which should be pointed to binary with debug info.
196 StringRef SymbolizerPath;
197 // Options used to configure the symbolizer
198 symbolize::LLVMSymbolizer::Options SymbolizerOpts;
199 // The runtime base address that the first executable segment is loaded at.
200 uint64_t BaseAddress = 0;
201 // The runtime base address that the first loadabe segment is loaded at.
202 uint64_t FirstLoadableAddress = 0;
203 // The preferred load address of each executable segment.
204 std::vector<uint64_t> PreferredTextSegmentAddresses;
205 // The file offset of each executable segment.
206 std::vector<uint64_t> TextSegmentOffsets;
208 // Mutiple MC component info
209 std::unique_ptr<const MCRegisterInfo> MRI;
210 std::unique_ptr<const MCAsmInfo> AsmInfo;
211 std::unique_ptr<const MCSubtargetInfo> STI;
212 std::unique_ptr<const MCInstrInfo> MII;
213 std::unique_ptr<MCDisassembler> DisAsm;
214 std::unique_ptr<const MCInstrAnalysis> MIA;
215 std::unique_ptr<MCInstPrinter> IPrinter;
216 // A list of text sections sorted by start RVA and size. Used to check
217 // if a given RVA is a valid code address.
218 std::set<std::pair<uint64_t, uint64_t>> TextSections;
220 // A map of mapping function name to BinaryFunction info.
221 std::unordered_map<std::string, BinaryFunction> BinaryFunctions;
223 // A list of binary functions that have samples.
224 std::unordered_set<const BinaryFunction *> ProfiledFunctions;
226 // GUID to Elf symbol start address map
227 DenseMap<uint64_t, uint64_t> SymbolStartAddrs;
229 // Start address to Elf symbol GUID map
230 std::unordered_multimap<uint64_t, uint64_t> StartAddrToSymMap;
232 // An ordered map of mapping function's start address to function range
233 // relevant info. Currently to determine if the offset of ELF is the start of
234 // a real function, we leverage the function range info from DWARF.
235 std::map<uint64_t, FuncRange> StartAddrToFuncRangeMap;
237 // Address to context location map. Used to expand the context.
238 std::unordered_map<uint64_t, SampleContextFrameVector> AddressToLocStackMap;
240 // Address to instruction size map. Also used for quick Address lookup.
241 std::unordered_map<uint64_t, uint64_t> AddressToInstSizeMap;
243 // An array of Addresses of all instructions sorted in increasing order. The
244 // sorting is needed to fast advance to the next forward/backward instruction.
245 std::vector<uint64_t> CodeAddressVec;
246 // A set of call instruction addresses. Used by virtual unwinding.
247 std::unordered_set<uint64_t> CallAddressSet;
248 // A set of return instruction addresses. Used by virtual unwinding.
249 std::unordered_set<uint64_t> RetAddressSet;
250 // An ordered set of unconditional branch instruction addresses.
251 std::set<uint64_t> UncondBranchAddrSet;
252 // A set of branch instruction addresses.
253 std::unordered_set<uint64_t> BranchAddressSet;
255 // Estimate and track function prolog and epilog ranges.
256 PrologEpilogTracker ProEpilogTracker;
258 // Infer missing frames due to compiler optimizations such as tail call
259 // elimination.
260 std::unique_ptr<MissingFrameInferrer> MissingContextInferrer;
262 // Track function sizes under different context
263 BinarySizeContextTracker FuncSizeTracker;
265 // The symbolizer used to get inline context for an instruction.
266 std::unique_ptr<symbolize::LLVMSymbolizer> Symbolizer;
268 // String table owning function name strings created from the symbolizer.
269 std::unordered_set<std::string> NameStrings;
271 // A collection of functions to print disassembly for.
272 StringSet<> DisassembleFunctionSet;
274 // Pseudo probe decoder
275 MCPseudoProbeDecoder ProbeDecoder;
277 // Function name to probe frame map for top-level outlined functions.
278 StringMap<MCDecodedPseudoProbeInlineTree *> TopLevelProbeFrameMap;
280 bool UsePseudoProbes = false;
282 bool UseFSDiscriminator = false;
284 // Whether we need to symbolize all instructions to get function context size.
285 bool TrackFuncContextSize = false;
287 // Indicate if the base loading address is parsed from the mmap event or uses
288 // the preferred address
289 bool IsLoadedByMMap = false;
290 // Use to avoid redundant warning.
291 bool MissingMMapWarned = false;
293 void setPreferredTextSegmentAddresses(const ELFObjectFileBase *O);
295 template <class ELFT>
296 void setPreferredTextSegmentAddresses(const ELFFile<ELFT> &Obj,
297 StringRef FileName);
299 void checkPseudoProbe(const ELFObjectFileBase *Obj);
301 void decodePseudoProbe(const ELFObjectFileBase *Obj);
303 void
304 checkUseFSDiscriminator(const ELFObjectFileBase *Obj,
305 std::map<SectionRef, SectionSymbolsTy> &AllSymbols);
307 // Set up disassembler and related components.
308 void setUpDisassembler(const ELFObjectFileBase *Obj);
309 symbolize::LLVMSymbolizer::Options getSymbolizerOpts() const;
311 // Load debug info of subprograms from DWARF section.
312 void loadSymbolsFromDWARF(ObjectFile &Obj);
314 // Load debug info from DWARF unit.
315 void loadSymbolsFromDWARFUnit(DWARFUnit &CompilationUnit);
317 // Create elf symbol to its start address mapping.
318 void populateElfSymbolAddressList(const ELFObjectFileBase *O);
320 // A function may be spilt into multiple non-continuous address ranges. We use
321 // this to set whether start a function range is the real entry of the
322 // function and also set false to the non-function label.
323 void setIsFuncEntry(FuncRange *FRange, StringRef RangeSymName);
325 // Warn if no entry range exists in the function.
326 void warnNoFuncEntry();
328 /// Dissassemble the text section and build various address maps.
329 void disassemble(const ELFObjectFileBase *O);
331 /// Helper function to dissassemble the symbol and extract info for unwinding
332 bool dissassembleSymbol(std::size_t SI, ArrayRef<uint8_t> Bytes,
333 SectionSymbolsTy &Symbols, const SectionRef &Section);
334 /// Symbolize a given instruction pointer and return a full call context.
335 SampleContextFrameVector symbolize(const InstructionPointer &IP,
336 bool UseCanonicalFnName = false,
337 bool UseProbeDiscriminator = false);
338 /// Decode the interesting parts of the binary and build internal data
339 /// structures. On high level, the parts of interest are:
340 /// 1. Text sections, including the main code section and the PLT
341 /// entries that will be used to handle cross-module call transitions.
342 /// 2. The .debug_line section, used by Dwarf-based profile generation.
343 /// 3. Pseudo probe related sections, used by probe-based profile
344 /// generation.
345 void load();
347 public:
348 ProfiledBinary(const StringRef ExeBinPath, const StringRef DebugBinPath);
349 ~ProfiledBinary();
351 void decodePseudoProbe();
353 StringRef getPath() const { return Path; }
354 StringRef getName() const { return llvm::sys::path::filename(Path); }
355 uint64_t getBaseAddress() const { return BaseAddress; }
356 void setBaseAddress(uint64_t Address) { BaseAddress = Address; }
358 // Canonicalize to use preferred load address as base address.
359 uint64_t canonicalizeVirtualAddress(uint64_t Address) {
360 return Address - BaseAddress + getPreferredBaseAddress();
362 // Return the preferred load address for the first executable segment.
363 uint64_t getPreferredBaseAddress() const {
364 return PreferredTextSegmentAddresses[0];
366 // Return the preferred load address for the first loadable segment.
367 uint64_t getFirstLoadableAddress() const { return FirstLoadableAddress; }
368 // Return the file offset for the first executable segment.
369 uint64_t getTextSegmentOffset() const { return TextSegmentOffsets[0]; }
370 const std::vector<uint64_t> &getPreferredTextSegmentAddresses() const {
371 return PreferredTextSegmentAddresses;
373 const std::vector<uint64_t> &getTextSegmentOffsets() const {
374 return TextSegmentOffsets;
377 uint64_t getInstSize(uint64_t Address) const {
378 auto I = AddressToInstSizeMap.find(Address);
379 if (I == AddressToInstSizeMap.end())
380 return 0;
381 return I->second;
384 bool addressIsCode(uint64_t Address) const {
385 return AddressToInstSizeMap.find(Address) != AddressToInstSizeMap.end();
388 bool addressIsCall(uint64_t Address) const {
389 return CallAddressSet.count(Address);
391 bool addressIsReturn(uint64_t Address) const {
392 return RetAddressSet.count(Address);
394 bool addressInPrologEpilog(uint64_t Address) const {
395 return ProEpilogTracker.PrologEpilogSet.count(Address);
398 bool addressIsTransfer(uint64_t Address) {
399 return BranchAddressSet.count(Address) || RetAddressSet.count(Address) ||
400 CallAddressSet.count(Address);
403 bool rangeCrossUncondBranch(uint64_t Start, uint64_t End) {
404 if (Start >= End)
405 return false;
406 auto R = UncondBranchAddrSet.lower_bound(Start);
407 return R != UncondBranchAddrSet.end() && *R < End;
410 uint64_t getAddressforIndex(uint64_t Index) const {
411 return CodeAddressVec[Index];
414 size_t getCodeAddrVecSize() const { return CodeAddressVec.size(); }
416 bool usePseudoProbes() const { return UsePseudoProbes; }
417 bool useFSDiscriminator() const { return UseFSDiscriminator; }
418 // Get the index in CodeAddressVec for the address
419 // As we might get an address which is not the code
420 // here it would round to the next valid code address by
421 // using lower bound operation
422 uint32_t getIndexForAddr(uint64_t Address) const {
423 auto Low = llvm::lower_bound(CodeAddressVec, Address);
424 return Low - CodeAddressVec.begin();
427 uint64_t getCallAddrFromFrameAddr(uint64_t FrameAddr) const {
428 if (FrameAddr == ExternalAddr)
429 return ExternalAddr;
430 auto I = getIndexForAddr(FrameAddr);
431 FrameAddr = I ? getAddressforIndex(I - 1) : 0;
432 if (FrameAddr && addressIsCall(FrameAddr))
433 return FrameAddr;
434 return 0;
437 FuncRange *findFuncRangeForStartAddr(uint64_t Address) {
438 auto I = StartAddrToFuncRangeMap.find(Address);
439 if (I == StartAddrToFuncRangeMap.end())
440 return nullptr;
441 return &I->second;
444 // Binary search the function range which includes the input address.
445 FuncRange *findFuncRange(uint64_t Address) {
446 auto I = StartAddrToFuncRangeMap.upper_bound(Address);
447 if (I == StartAddrToFuncRangeMap.begin())
448 return nullptr;
449 I--;
451 if (Address >= I->second.EndAddress)
452 return nullptr;
454 return &I->second;
457 // Get all ranges of one function.
458 RangesTy getRanges(uint64_t Address) {
459 auto *FRange = findFuncRange(Address);
460 // Ignore the range which falls into plt section or system lib.
461 if (!FRange)
462 return RangesTy();
464 return FRange->Func->Ranges;
467 const std::unordered_map<std::string, BinaryFunction> &
468 getAllBinaryFunctions() {
469 return BinaryFunctions;
472 std::unordered_set<const BinaryFunction *> &getProfiledFunctions() {
473 return ProfiledFunctions;
476 void setProfiledFunctions(std::unordered_set<const BinaryFunction *> &Funcs) {
477 ProfiledFunctions = Funcs;
480 BinaryFunction *getBinaryFunction(StringRef FName) {
481 auto I = BinaryFunctions.find(FName.str());
482 if (I == BinaryFunctions.end())
483 return nullptr;
484 return &I->second;
487 uint32_t getFuncSizeForContext(const ContextTrieNode *ContextNode) {
488 return FuncSizeTracker.getFuncSizeForContext(ContextNode);
491 void inferMissingFrames(const SmallVectorImpl<uint64_t> &Context,
492 SmallVectorImpl<uint64_t> &NewContext);
494 // Load the symbols from debug table and populate into symbol list.
495 void populateSymbolListFromDWARF(ProfileSymbolList &SymbolList);
497 SampleContextFrameVector
498 getFrameLocationStack(uint64_t Address, bool UseProbeDiscriminator = false) {
499 InstructionPointer IP(this, Address);
500 return symbolize(IP, SymbolizerOpts.UseSymbolTable, UseProbeDiscriminator);
503 const SampleContextFrameVector &
504 getCachedFrameLocationStack(uint64_t Address,
505 bool UseProbeDiscriminator = false) {
506 auto I = AddressToLocStackMap.emplace(Address, SampleContextFrameVector());
507 if (I.second) {
508 I.first->second = getFrameLocationStack(Address, UseProbeDiscriminator);
510 return I.first->second;
513 std::optional<SampleContextFrame> getInlineLeafFrameLoc(uint64_t Address) {
514 const auto &Stack = getCachedFrameLocationStack(Address);
515 if (Stack.empty())
516 return {};
517 return Stack.back();
520 void flushSymbolizer() { Symbolizer.reset(); }
522 MissingFrameInferrer* getMissingContextInferrer() {
523 return MissingContextInferrer.get();
526 // Compare two addresses' inline context
527 bool inlineContextEqual(uint64_t Add1, uint64_t Add2);
529 // Get the full context of the current stack with inline context filled in.
530 // It will search the disassembling info stored in AddressToLocStackMap. This
531 // is used as the key of function sample map
532 SampleContextFrameVector
533 getExpandedContext(const SmallVectorImpl<uint64_t> &Stack,
534 bool &WasLeafInlined);
535 // Go through instructions among the given range and record its size for the
536 // inline context.
537 void computeInlinedContextSizeForRange(uint64_t StartAddress,
538 uint64_t EndAddress);
540 void computeInlinedContextSizeForFunc(const BinaryFunction *Func);
542 const MCDecodedPseudoProbe *getCallProbeForAddr(uint64_t Address) const {
543 return ProbeDecoder.getCallProbeForAddr(Address);
546 void getInlineContextForProbe(const MCDecodedPseudoProbe *Probe,
547 SampleContextFrameVector &InlineContextStack,
548 bool IncludeLeaf = false) const {
549 SmallVector<MCPseduoProbeFrameLocation, 16> ProbeInlineContext;
550 ProbeDecoder.getInlineContextForProbe(Probe, ProbeInlineContext,
551 IncludeLeaf);
552 for (uint32_t I = 0; I < ProbeInlineContext.size(); I++) {
553 auto &Callsite = ProbeInlineContext[I];
554 // Clear the current context for an unknown probe.
555 if (Callsite.second == 0 && I != ProbeInlineContext.size() - 1) {
556 InlineContextStack.clear();
557 continue;
559 InlineContextStack.emplace_back(Callsite.first,
560 LineLocation(Callsite.second, 0));
563 const AddressProbesMap &getAddress2ProbesMap() const {
564 return ProbeDecoder.getAddress2ProbesMap();
566 const MCPseudoProbeFuncDesc *getFuncDescForGUID(uint64_t GUID) {
567 return ProbeDecoder.getFuncDescForGUID(GUID);
570 const MCPseudoProbeFuncDesc *
571 getInlinerDescForProbe(const MCDecodedPseudoProbe *Probe) {
572 return ProbeDecoder.getInlinerDescForProbe(Probe);
575 bool getTrackFuncContextSize() { return TrackFuncContextSize; }
577 bool getIsLoadedByMMap() { return IsLoadedByMMap; }
579 void setIsLoadedByMMap(bool Value) { IsLoadedByMMap = Value; }
581 bool getMissingMMapWarned() { return MissingMMapWarned; }
583 void setMissingMMapWarned(bool Value) { MissingMMapWarned = Value; }
586 } // end namespace sampleprof
587 } // end namespace llvm
589 #endif