[AMDGPU] Test codegen'ing True16 additions.
[llvm-project.git] / llvm / lib / Transforms / Instrumentation / MemProfiler.cpp
blob453dd00488b8ffef9656fe51e9132e86ea902abf
1 //===- MemProfiler.cpp - memory allocation and access profiler ------------===//
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 is a part of MemProfiler. Memory accesses are instrumented
10 // to increment the access count held in a shadow memory location, or
11 // alternatively to call into the runtime. Memory intrinsic calls (memmove,
12 // memcpy, memset) are changed to call the memory profiling runtime version
13 // instead.
15 //===----------------------------------------------------------------------===//
17 #include "llvm/Transforms/Instrumentation/MemProfiler.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/ADT/StringRef.h"
21 #include "llvm/Analysis/MemoryBuiltins.h"
22 #include "llvm/Analysis/MemoryProfileInfo.h"
23 #include "llvm/Analysis/ValueTracking.h"
24 #include "llvm/IR/Constant.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DiagnosticInfo.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/GlobalValue.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/Instruction.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/Module.h"
33 #include "llvm/IR/Type.h"
34 #include "llvm/IR/Value.h"
35 #include "llvm/ProfileData/InstrProf.h"
36 #include "llvm/ProfileData/InstrProfReader.h"
37 #include "llvm/Support/BLAKE3.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/HashBuilder.h"
41 #include "llvm/Support/VirtualFileSystem.h"
42 #include "llvm/TargetParser/Triple.h"
43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
44 #include "llvm/Transforms/Utils/ModuleUtils.h"
45 #include <map>
46 #include <set>
48 using namespace llvm;
49 using namespace llvm::memprof;
51 #define DEBUG_TYPE "memprof"
53 namespace llvm {
54 extern cl::opt<bool> PGOWarnMissing;
55 extern cl::opt<bool> NoPGOWarnMismatch;
56 extern cl::opt<bool> NoPGOWarnMismatchComdatWeak;
57 } // namespace llvm
59 constexpr int LLVM_MEM_PROFILER_VERSION = 1;
61 // Size of memory mapped to a single shadow location.
62 constexpr uint64_t DefaultShadowGranularity = 64;
64 // Scale from granularity down to shadow size.
65 constexpr uint64_t DefaultShadowScale = 3;
67 constexpr char MemProfModuleCtorName[] = "memprof.module_ctor";
68 constexpr uint64_t MemProfCtorAndDtorPriority = 1;
69 // On Emscripten, the system needs more than one priorities for constructors.
70 constexpr uint64_t MemProfEmscriptenCtorAndDtorPriority = 50;
71 constexpr char MemProfInitName[] = "__memprof_init";
72 constexpr char MemProfVersionCheckNamePrefix[] =
73 "__memprof_version_mismatch_check_v";
75 constexpr char MemProfShadowMemoryDynamicAddress[] =
76 "__memprof_shadow_memory_dynamic_address";
78 constexpr char MemProfFilenameVar[] = "__memprof_profile_filename";
80 // Command-line flags.
82 static cl::opt<bool> ClInsertVersionCheck(
83 "memprof-guard-against-version-mismatch",
84 cl::desc("Guard against compiler/runtime version mismatch."), cl::Hidden,
85 cl::init(true));
87 // This flag may need to be replaced with -f[no-]memprof-reads.
88 static cl::opt<bool> ClInstrumentReads("memprof-instrument-reads",
89 cl::desc("instrument read instructions"),
90 cl::Hidden, cl::init(true));
92 static cl::opt<bool>
93 ClInstrumentWrites("memprof-instrument-writes",
94 cl::desc("instrument write instructions"), cl::Hidden,
95 cl::init(true));
97 static cl::opt<bool> ClInstrumentAtomics(
98 "memprof-instrument-atomics",
99 cl::desc("instrument atomic instructions (rmw, cmpxchg)"), cl::Hidden,
100 cl::init(true));
102 static cl::opt<bool> ClUseCalls(
103 "memprof-use-callbacks",
104 cl::desc("Use callbacks instead of inline instrumentation sequences."),
105 cl::Hidden, cl::init(false));
107 static cl::opt<std::string>
108 ClMemoryAccessCallbackPrefix("memprof-memory-access-callback-prefix",
109 cl::desc("Prefix for memory access callbacks"),
110 cl::Hidden, cl::init("__memprof_"));
112 // These flags allow to change the shadow mapping.
113 // The shadow mapping looks like
114 // Shadow = ((Mem & mask) >> scale) + offset
116 static cl::opt<int> ClMappingScale("memprof-mapping-scale",
117 cl::desc("scale of memprof shadow mapping"),
118 cl::Hidden, cl::init(DefaultShadowScale));
120 static cl::opt<int>
121 ClMappingGranularity("memprof-mapping-granularity",
122 cl::desc("granularity of memprof shadow mapping"),
123 cl::Hidden, cl::init(DefaultShadowGranularity));
125 static cl::opt<bool> ClStack("memprof-instrument-stack",
126 cl::desc("Instrument scalar stack variables"),
127 cl::Hidden, cl::init(false));
129 // Debug flags.
131 static cl::opt<int> ClDebug("memprof-debug", cl::desc("debug"), cl::Hidden,
132 cl::init(0));
134 static cl::opt<std::string> ClDebugFunc("memprof-debug-func", cl::Hidden,
135 cl::desc("Debug func"));
137 static cl::opt<int> ClDebugMin("memprof-debug-min", cl::desc("Debug min inst"),
138 cl::Hidden, cl::init(-1));
140 static cl::opt<int> ClDebugMax("memprof-debug-max", cl::desc("Debug max inst"),
141 cl::Hidden, cl::init(-1));
143 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
144 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
145 STATISTIC(NumSkippedStackReads, "Number of non-instrumented stack reads");
146 STATISTIC(NumSkippedStackWrites, "Number of non-instrumented stack writes");
147 STATISTIC(NumOfMemProfMissing, "Number of functions without memory profile.");
149 namespace {
151 /// This struct defines the shadow mapping using the rule:
152 /// shadow = ((mem & mask) >> Scale) ADD DynamicShadowOffset.
153 struct ShadowMapping {
154 ShadowMapping() {
155 Scale = ClMappingScale;
156 Granularity = ClMappingGranularity;
157 Mask = ~(Granularity - 1);
160 int Scale;
161 int Granularity;
162 uint64_t Mask; // Computed as ~(Granularity-1)
165 static uint64_t getCtorAndDtorPriority(Triple &TargetTriple) {
166 return TargetTriple.isOSEmscripten() ? MemProfEmscriptenCtorAndDtorPriority
167 : MemProfCtorAndDtorPriority;
170 struct InterestingMemoryAccess {
171 Value *Addr = nullptr;
172 bool IsWrite;
173 Type *AccessTy;
174 uint64_t TypeSize;
175 Value *MaybeMask = nullptr;
178 /// Instrument the code in module to profile memory accesses.
179 class MemProfiler {
180 public:
181 MemProfiler(Module &M) {
182 C = &(M.getContext());
183 LongSize = M.getDataLayout().getPointerSizeInBits();
184 IntptrTy = Type::getIntNTy(*C, LongSize);
187 /// If it is an interesting memory access, populate information
188 /// about the access and return a InterestingMemoryAccess struct.
189 /// Otherwise return std::nullopt.
190 std::optional<InterestingMemoryAccess>
191 isInterestingMemoryAccess(Instruction *I) const;
193 void instrumentMop(Instruction *I, const DataLayout &DL,
194 InterestingMemoryAccess &Access);
195 void instrumentAddress(Instruction *OrigIns, Instruction *InsertBefore,
196 Value *Addr, uint32_t TypeSize, bool IsWrite);
197 void instrumentMaskedLoadOrStore(const DataLayout &DL, Value *Mask,
198 Instruction *I, Value *Addr, Type *AccessTy,
199 bool IsWrite);
200 void instrumentMemIntrinsic(MemIntrinsic *MI);
201 Value *memToShadow(Value *Shadow, IRBuilder<> &IRB);
202 bool instrumentFunction(Function &F);
203 bool maybeInsertMemProfInitAtFunctionEntry(Function &F);
204 bool insertDynamicShadowAtFunctionEntry(Function &F);
206 private:
207 void initializeCallbacks(Module &M);
209 LLVMContext *C;
210 int LongSize;
211 Type *IntptrTy;
212 ShadowMapping Mapping;
214 // These arrays is indexed by AccessIsWrite
215 FunctionCallee MemProfMemoryAccessCallback[2];
216 FunctionCallee MemProfMemoryAccessCallbackSized[2];
218 FunctionCallee MemProfMemmove, MemProfMemcpy, MemProfMemset;
219 Value *DynamicShadowOffset = nullptr;
222 class ModuleMemProfiler {
223 public:
224 ModuleMemProfiler(Module &M) { TargetTriple = Triple(M.getTargetTriple()); }
226 bool instrumentModule(Module &);
228 private:
229 Triple TargetTriple;
230 ShadowMapping Mapping;
231 Function *MemProfCtorFunction = nullptr;
234 } // end anonymous namespace
236 MemProfilerPass::MemProfilerPass() = default;
238 PreservedAnalyses MemProfilerPass::run(Function &F,
239 AnalysisManager<Function> &AM) {
240 Module &M = *F.getParent();
241 MemProfiler Profiler(M);
242 if (Profiler.instrumentFunction(F))
243 return PreservedAnalyses::none();
244 return PreservedAnalyses::all();
247 ModuleMemProfilerPass::ModuleMemProfilerPass() = default;
249 PreservedAnalyses ModuleMemProfilerPass::run(Module &M,
250 AnalysisManager<Module> &AM) {
251 ModuleMemProfiler Profiler(M);
252 if (Profiler.instrumentModule(M))
253 return PreservedAnalyses::none();
254 return PreservedAnalyses::all();
257 Value *MemProfiler::memToShadow(Value *Shadow, IRBuilder<> &IRB) {
258 // (Shadow & mask) >> scale
259 Shadow = IRB.CreateAnd(Shadow, Mapping.Mask);
260 Shadow = IRB.CreateLShr(Shadow, Mapping.Scale);
261 // (Shadow >> scale) | offset
262 assert(DynamicShadowOffset);
263 return IRB.CreateAdd(Shadow, DynamicShadowOffset);
266 // Instrument memset/memmove/memcpy
267 void MemProfiler::instrumentMemIntrinsic(MemIntrinsic *MI) {
268 IRBuilder<> IRB(MI);
269 if (isa<MemTransferInst>(MI)) {
270 IRB.CreateCall(
271 isa<MemMoveInst>(MI) ? MemProfMemmove : MemProfMemcpy,
272 {IRB.CreatePointerCast(MI->getOperand(0), IRB.getInt8PtrTy()),
273 IRB.CreatePointerCast(MI->getOperand(1), IRB.getInt8PtrTy()),
274 IRB.CreateIntCast(MI->getOperand(2), IntptrTy, false)});
275 } else if (isa<MemSetInst>(MI)) {
276 IRB.CreateCall(
277 MemProfMemset,
278 {IRB.CreatePointerCast(MI->getOperand(0), IRB.getInt8PtrTy()),
279 IRB.CreateIntCast(MI->getOperand(1), IRB.getInt32Ty(), false),
280 IRB.CreateIntCast(MI->getOperand(2), IntptrTy, false)});
282 MI->eraseFromParent();
285 std::optional<InterestingMemoryAccess>
286 MemProfiler::isInterestingMemoryAccess(Instruction *I) const {
287 // Do not instrument the load fetching the dynamic shadow address.
288 if (DynamicShadowOffset == I)
289 return std::nullopt;
291 InterestingMemoryAccess Access;
293 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
294 if (!ClInstrumentReads)
295 return std::nullopt;
296 Access.IsWrite = false;
297 Access.AccessTy = LI->getType();
298 Access.Addr = LI->getPointerOperand();
299 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
300 if (!ClInstrumentWrites)
301 return std::nullopt;
302 Access.IsWrite = true;
303 Access.AccessTy = SI->getValueOperand()->getType();
304 Access.Addr = SI->getPointerOperand();
305 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) {
306 if (!ClInstrumentAtomics)
307 return std::nullopt;
308 Access.IsWrite = true;
309 Access.AccessTy = RMW->getValOperand()->getType();
310 Access.Addr = RMW->getPointerOperand();
311 } else if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(I)) {
312 if (!ClInstrumentAtomics)
313 return std::nullopt;
314 Access.IsWrite = true;
315 Access.AccessTy = XCHG->getCompareOperand()->getType();
316 Access.Addr = XCHG->getPointerOperand();
317 } else if (auto *CI = dyn_cast<CallInst>(I)) {
318 auto *F = CI->getCalledFunction();
319 if (F && (F->getIntrinsicID() == Intrinsic::masked_load ||
320 F->getIntrinsicID() == Intrinsic::masked_store)) {
321 unsigned OpOffset = 0;
322 if (F->getIntrinsicID() == Intrinsic::masked_store) {
323 if (!ClInstrumentWrites)
324 return std::nullopt;
325 // Masked store has an initial operand for the value.
326 OpOffset = 1;
327 Access.AccessTy = CI->getArgOperand(0)->getType();
328 Access.IsWrite = true;
329 } else {
330 if (!ClInstrumentReads)
331 return std::nullopt;
332 Access.AccessTy = CI->getType();
333 Access.IsWrite = false;
336 auto *BasePtr = CI->getOperand(0 + OpOffset);
337 Access.MaybeMask = CI->getOperand(2 + OpOffset);
338 Access.Addr = BasePtr;
342 if (!Access.Addr)
343 return std::nullopt;
345 // Do not instrument accesses from different address spaces; we cannot deal
346 // with them.
347 Type *PtrTy = cast<PointerType>(Access.Addr->getType()->getScalarType());
348 if (PtrTy->getPointerAddressSpace() != 0)
349 return std::nullopt;
351 // Ignore swifterror addresses.
352 // swifterror memory addresses are mem2reg promoted by instruction
353 // selection. As such they cannot have regular uses like an instrumentation
354 // function and it makes no sense to track them as memory.
355 if (Access.Addr->isSwiftError())
356 return std::nullopt;
358 // Peel off GEPs and BitCasts.
359 auto *Addr = Access.Addr->stripInBoundsOffsets();
361 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
362 // Do not instrument PGO counter updates.
363 if (GV->hasSection()) {
364 StringRef SectionName = GV->getSection();
365 // Check if the global is in the PGO counters section.
366 auto OF = Triple(I->getModule()->getTargetTriple()).getObjectFormat();
367 if (SectionName.endswith(
368 getInstrProfSectionName(IPSK_cnts, OF, /*AddSegmentInfo=*/false)))
369 return std::nullopt;
372 // Do not instrument accesses to LLVM internal variables.
373 if (GV->getName().startswith("__llvm"))
374 return std::nullopt;
377 const DataLayout &DL = I->getModule()->getDataLayout();
378 Access.TypeSize = DL.getTypeStoreSizeInBits(Access.AccessTy);
379 return Access;
382 void MemProfiler::instrumentMaskedLoadOrStore(const DataLayout &DL, Value *Mask,
383 Instruction *I, Value *Addr,
384 Type *AccessTy, bool IsWrite) {
385 auto *VTy = cast<FixedVectorType>(AccessTy);
386 uint64_t ElemTypeSize = DL.getTypeStoreSizeInBits(VTy->getScalarType());
387 unsigned Num = VTy->getNumElements();
388 auto *Zero = ConstantInt::get(IntptrTy, 0);
389 for (unsigned Idx = 0; Idx < Num; ++Idx) {
390 Value *InstrumentedAddress = nullptr;
391 Instruction *InsertBefore = I;
392 if (auto *Vector = dyn_cast<ConstantVector>(Mask)) {
393 // dyn_cast as we might get UndefValue
394 if (auto *Masked = dyn_cast<ConstantInt>(Vector->getOperand(Idx))) {
395 if (Masked->isZero())
396 // Mask is constant false, so no instrumentation needed.
397 continue;
398 // If we have a true or undef value, fall through to instrumentAddress.
399 // with InsertBefore == I
401 } else {
402 IRBuilder<> IRB(I);
403 Value *MaskElem = IRB.CreateExtractElement(Mask, Idx);
404 Instruction *ThenTerm = SplitBlockAndInsertIfThen(MaskElem, I, false);
405 InsertBefore = ThenTerm;
408 IRBuilder<> IRB(InsertBefore);
409 InstrumentedAddress =
410 IRB.CreateGEP(VTy, Addr, {Zero, ConstantInt::get(IntptrTy, Idx)});
411 instrumentAddress(I, InsertBefore, InstrumentedAddress, ElemTypeSize,
412 IsWrite);
416 void MemProfiler::instrumentMop(Instruction *I, const DataLayout &DL,
417 InterestingMemoryAccess &Access) {
418 // Skip instrumentation of stack accesses unless requested.
419 if (!ClStack && isa<AllocaInst>(getUnderlyingObject(Access.Addr))) {
420 if (Access.IsWrite)
421 ++NumSkippedStackWrites;
422 else
423 ++NumSkippedStackReads;
424 return;
427 if (Access.IsWrite)
428 NumInstrumentedWrites++;
429 else
430 NumInstrumentedReads++;
432 if (Access.MaybeMask) {
433 instrumentMaskedLoadOrStore(DL, Access.MaybeMask, I, Access.Addr,
434 Access.AccessTy, Access.IsWrite);
435 } else {
436 // Since the access counts will be accumulated across the entire allocation,
437 // we only update the shadow access count for the first location and thus
438 // don't need to worry about alignment and type size.
439 instrumentAddress(I, I, Access.Addr, Access.TypeSize, Access.IsWrite);
443 void MemProfiler::instrumentAddress(Instruction *OrigIns,
444 Instruction *InsertBefore, Value *Addr,
445 uint32_t TypeSize, bool IsWrite) {
446 IRBuilder<> IRB(InsertBefore);
447 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
449 if (ClUseCalls) {
450 IRB.CreateCall(MemProfMemoryAccessCallback[IsWrite], AddrLong);
451 return;
454 // Create an inline sequence to compute shadow location, and increment the
455 // value by one.
456 Type *ShadowTy = Type::getInt64Ty(*C);
457 Type *ShadowPtrTy = PointerType::get(ShadowTy, 0);
458 Value *ShadowPtr = memToShadow(AddrLong, IRB);
459 Value *ShadowAddr = IRB.CreateIntToPtr(ShadowPtr, ShadowPtrTy);
460 Value *ShadowValue = IRB.CreateLoad(ShadowTy, ShadowAddr);
461 Value *Inc = ConstantInt::get(Type::getInt64Ty(*C), 1);
462 ShadowValue = IRB.CreateAdd(ShadowValue, Inc);
463 IRB.CreateStore(ShadowValue, ShadowAddr);
466 // Create the variable for the profile file name.
467 void createProfileFileNameVar(Module &M) {
468 const MDString *MemProfFilename =
469 dyn_cast_or_null<MDString>(M.getModuleFlag("MemProfProfileFilename"));
470 if (!MemProfFilename)
471 return;
472 assert(!MemProfFilename->getString().empty() &&
473 "Unexpected MemProfProfileFilename metadata with empty string");
474 Constant *ProfileNameConst = ConstantDataArray::getString(
475 M.getContext(), MemProfFilename->getString(), true);
476 GlobalVariable *ProfileNameVar = new GlobalVariable(
477 M, ProfileNameConst->getType(), /*isConstant=*/true,
478 GlobalValue::WeakAnyLinkage, ProfileNameConst, MemProfFilenameVar);
479 Triple TT(M.getTargetTriple());
480 if (TT.supportsCOMDAT()) {
481 ProfileNameVar->setLinkage(GlobalValue::ExternalLinkage);
482 ProfileNameVar->setComdat(M.getOrInsertComdat(MemProfFilenameVar));
486 bool ModuleMemProfiler::instrumentModule(Module &M) {
487 // Create a module constructor.
488 std::string MemProfVersion = std::to_string(LLVM_MEM_PROFILER_VERSION);
489 std::string VersionCheckName =
490 ClInsertVersionCheck ? (MemProfVersionCheckNamePrefix + MemProfVersion)
491 : "";
492 std::tie(MemProfCtorFunction, std::ignore) =
493 createSanitizerCtorAndInitFunctions(M, MemProfModuleCtorName,
494 MemProfInitName, /*InitArgTypes=*/{},
495 /*InitArgs=*/{}, VersionCheckName);
497 const uint64_t Priority = getCtorAndDtorPriority(TargetTriple);
498 appendToGlobalCtors(M, MemProfCtorFunction, Priority);
500 createProfileFileNameVar(M);
502 return true;
505 void MemProfiler::initializeCallbacks(Module &M) {
506 IRBuilder<> IRB(*C);
508 for (size_t AccessIsWrite = 0; AccessIsWrite <= 1; AccessIsWrite++) {
509 const std::string TypeStr = AccessIsWrite ? "store" : "load";
511 SmallVector<Type *, 3> Args2 = {IntptrTy, IntptrTy};
512 SmallVector<Type *, 2> Args1{1, IntptrTy};
513 MemProfMemoryAccessCallbackSized[AccessIsWrite] =
514 M.getOrInsertFunction(ClMemoryAccessCallbackPrefix + TypeStr + "N",
515 FunctionType::get(IRB.getVoidTy(), Args2, false));
517 MemProfMemoryAccessCallback[AccessIsWrite] =
518 M.getOrInsertFunction(ClMemoryAccessCallbackPrefix + TypeStr,
519 FunctionType::get(IRB.getVoidTy(), Args1, false));
521 MemProfMemmove = M.getOrInsertFunction(
522 ClMemoryAccessCallbackPrefix + "memmove", IRB.getInt8PtrTy(),
523 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IntptrTy);
524 MemProfMemcpy = M.getOrInsertFunction(ClMemoryAccessCallbackPrefix + "memcpy",
525 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
526 IRB.getInt8PtrTy(), IntptrTy);
527 MemProfMemset = M.getOrInsertFunction(ClMemoryAccessCallbackPrefix + "memset",
528 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
529 IRB.getInt32Ty(), IntptrTy);
532 bool MemProfiler::maybeInsertMemProfInitAtFunctionEntry(Function &F) {
533 // For each NSObject descendant having a +load method, this method is invoked
534 // by the ObjC runtime before any of the static constructors is called.
535 // Therefore we need to instrument such methods with a call to __memprof_init
536 // at the beginning in order to initialize our runtime before any access to
537 // the shadow memory.
538 // We cannot just ignore these methods, because they may call other
539 // instrumented functions.
540 if (F.getName().find(" load]") != std::string::npos) {
541 FunctionCallee MemProfInitFunction =
542 declareSanitizerInitFunction(*F.getParent(), MemProfInitName, {});
543 IRBuilder<> IRB(&F.front(), F.front().begin());
544 IRB.CreateCall(MemProfInitFunction, {});
545 return true;
547 return false;
550 bool MemProfiler::insertDynamicShadowAtFunctionEntry(Function &F) {
551 IRBuilder<> IRB(&F.front().front());
552 Value *GlobalDynamicAddress = F.getParent()->getOrInsertGlobal(
553 MemProfShadowMemoryDynamicAddress, IntptrTy);
554 if (F.getParent()->getPICLevel() == PICLevel::NotPIC)
555 cast<GlobalVariable>(GlobalDynamicAddress)->setDSOLocal(true);
556 DynamicShadowOffset = IRB.CreateLoad(IntptrTy, GlobalDynamicAddress);
557 return true;
560 bool MemProfiler::instrumentFunction(Function &F) {
561 if (F.getLinkage() == GlobalValue::AvailableExternallyLinkage)
562 return false;
563 if (ClDebugFunc == F.getName())
564 return false;
565 if (F.getName().startswith("__memprof_"))
566 return false;
568 bool FunctionModified = false;
570 // If needed, insert __memprof_init.
571 // This function needs to be called even if the function body is not
572 // instrumented.
573 if (maybeInsertMemProfInitAtFunctionEntry(F))
574 FunctionModified = true;
576 LLVM_DEBUG(dbgs() << "MEMPROF instrumenting:\n" << F << "\n");
578 initializeCallbacks(*F.getParent());
580 SmallVector<Instruction *, 16> ToInstrument;
582 // Fill the set of memory operations to instrument.
583 for (auto &BB : F) {
584 for (auto &Inst : BB) {
585 if (isInterestingMemoryAccess(&Inst) || isa<MemIntrinsic>(Inst))
586 ToInstrument.push_back(&Inst);
590 if (ToInstrument.empty()) {
591 LLVM_DEBUG(dbgs() << "MEMPROF done instrumenting: " << FunctionModified
592 << " " << F << "\n");
594 return FunctionModified;
597 FunctionModified |= insertDynamicShadowAtFunctionEntry(F);
599 int NumInstrumented = 0;
600 for (auto *Inst : ToInstrument) {
601 if (ClDebugMin < 0 || ClDebugMax < 0 ||
602 (NumInstrumented >= ClDebugMin && NumInstrumented <= ClDebugMax)) {
603 std::optional<InterestingMemoryAccess> Access =
604 isInterestingMemoryAccess(Inst);
605 if (Access)
606 instrumentMop(Inst, F.getParent()->getDataLayout(), *Access);
607 else
608 instrumentMemIntrinsic(cast<MemIntrinsic>(Inst));
610 NumInstrumented++;
613 if (NumInstrumented > 0)
614 FunctionModified = true;
616 LLVM_DEBUG(dbgs() << "MEMPROF done instrumenting: " << FunctionModified << " "
617 << F << "\n");
619 return FunctionModified;
622 static void addCallsiteMetadata(Instruction &I,
623 std::vector<uint64_t> &InlinedCallStack,
624 LLVMContext &Ctx) {
625 I.setMetadata(LLVMContext::MD_callsite,
626 buildCallstackMetadata(InlinedCallStack, Ctx));
629 static uint64_t computeStackId(GlobalValue::GUID Function, uint32_t LineOffset,
630 uint32_t Column) {
631 llvm::HashBuilder<llvm::TruncatedBLAKE3<8>, llvm::support::endianness::little>
632 HashBuilder;
633 HashBuilder.add(Function, LineOffset, Column);
634 llvm::BLAKE3Result<8> Hash = HashBuilder.final();
635 uint64_t Id;
636 std::memcpy(&Id, Hash.data(), sizeof(Hash));
637 return Id;
640 static uint64_t computeStackId(const memprof::Frame &Frame) {
641 return computeStackId(Frame.Function, Frame.LineOffset, Frame.Column);
644 static void addCallStack(CallStackTrie &AllocTrie,
645 const AllocationInfo *AllocInfo) {
646 SmallVector<uint64_t> StackIds;
647 for (const auto &StackFrame : AllocInfo->CallStack)
648 StackIds.push_back(computeStackId(StackFrame));
649 auto AllocType = getAllocType(AllocInfo->Info.getTotalLifetimeAccessDensity(),
650 AllocInfo->Info.getAllocCount(),
651 AllocInfo->Info.getTotalLifetime());
652 AllocTrie.addCallStack(AllocType, StackIds);
655 // Helper to compare the InlinedCallStack computed from an instruction's debug
656 // info to a list of Frames from profile data (either the allocation data or a
657 // callsite). For callsites, the StartIndex to use in the Frame array may be
658 // non-zero.
659 static bool
660 stackFrameIncludesInlinedCallStack(ArrayRef<Frame> ProfileCallStack,
661 ArrayRef<uint64_t> InlinedCallStack,
662 unsigned StartIndex = 0) {
663 auto StackFrame = ProfileCallStack.begin() + StartIndex;
664 auto InlCallStackIter = InlinedCallStack.begin();
665 for (; StackFrame != ProfileCallStack.end() &&
666 InlCallStackIter != InlinedCallStack.end();
667 ++StackFrame, ++InlCallStackIter) {
668 uint64_t StackId = computeStackId(*StackFrame);
669 if (StackId != *InlCallStackIter)
670 return false;
672 // Return true if we found and matched all stack ids from the call
673 // instruction.
674 return InlCallStackIter == InlinedCallStack.end();
677 static void readMemprof(Module &M, Function &F,
678 IndexedInstrProfReader *MemProfReader,
679 const TargetLibraryInfo &TLI) {
680 auto &Ctx = M.getContext();
682 auto FuncName = getIRPGOFuncName(F);
683 auto FuncGUID = Function::getGUID(FuncName);
684 std::optional<memprof::MemProfRecord> MemProfRec;
685 auto Err = MemProfReader->getMemProfRecord(FuncGUID).moveInto(MemProfRec);
686 if (Err) {
687 // If we don't find getIRPGOFuncName(), try getPGOFuncName() to handle
688 // profiles built by older compilers
689 Err = handleErrors(std::move(Err), [&](const InstrProfError &IE) -> Error {
690 if (IE.get() != instrprof_error::unknown_function)
691 return make_error<InstrProfError>(IE);
692 auto FuncName = getPGOFuncName(F);
693 auto FuncGUID = Function::getGUID(FuncName);
694 if (auto Err =
695 MemProfReader->getMemProfRecord(FuncGUID).moveInto(MemProfRec))
696 return Err;
697 return Error::success();
700 if (Err) {
701 handleAllErrors(std::move(Err), [&](const InstrProfError &IPE) {
702 auto Err = IPE.get();
703 bool SkipWarning = false;
704 LLVM_DEBUG(dbgs() << "Error in reading profile for Func " << FuncName
705 << ": ");
706 if (Err == instrprof_error::unknown_function) {
707 NumOfMemProfMissing++;
708 SkipWarning = !PGOWarnMissing;
709 LLVM_DEBUG(dbgs() << "unknown function");
710 } else if (Err == instrprof_error::hash_mismatch) {
711 SkipWarning =
712 NoPGOWarnMismatch ||
713 (NoPGOWarnMismatchComdatWeak &&
714 (F.hasComdat() ||
715 F.getLinkage() == GlobalValue::AvailableExternallyLinkage));
716 LLVM_DEBUG(dbgs() << "hash mismatch (skip=" << SkipWarning << ")");
719 if (SkipWarning)
720 return;
722 std::string Msg = (IPE.message() + Twine(" ") + F.getName().str() +
723 Twine(" Hash = ") + std::to_string(FuncGUID))
724 .str();
726 Ctx.diagnose(
727 DiagnosticInfoPGOProfile(M.getName().data(), Msg, DS_Warning));
729 return;
732 // Build maps of the location hash to all profile data with that leaf location
733 // (allocation info and the callsites).
734 std::map<uint64_t, std::set<const AllocationInfo *>> LocHashToAllocInfo;
735 // For the callsites we need to record the index of the associated frame in
736 // the frame array (see comments below where the map entries are added).
737 std::map<uint64_t, std::set<std::pair<const SmallVector<Frame> *, unsigned>>>
738 LocHashToCallSites;
739 for (auto &AI : MemProfRec->AllocSites) {
740 // Associate the allocation info with the leaf frame. The later matching
741 // code will match any inlined call sequences in the IR with a longer prefix
742 // of call stack frames.
743 uint64_t StackId = computeStackId(AI.CallStack[0]);
744 LocHashToAllocInfo[StackId].insert(&AI);
746 for (auto &CS : MemProfRec->CallSites) {
747 // Need to record all frames from leaf up to and including this function,
748 // as any of these may or may not have been inlined at this point.
749 unsigned Idx = 0;
750 for (auto &StackFrame : CS) {
751 uint64_t StackId = computeStackId(StackFrame);
752 LocHashToCallSites[StackId].insert(std::make_pair(&CS, Idx++));
753 // Once we find this function, we can stop recording.
754 if (StackFrame.Function == FuncGUID)
755 break;
757 assert(Idx <= CS.size() && CS[Idx - 1].Function == FuncGUID);
760 auto GetOffset = [](const DILocation *DIL) {
761 return (DIL->getLine() - DIL->getScope()->getSubprogram()->getLine()) &
762 0xffff;
765 // Now walk the instructions, looking up the associated profile data using
766 // dbug locations.
767 for (auto &BB : F) {
768 for (auto &I : BB) {
769 if (I.isDebugOrPseudoInst())
770 continue;
771 // We are only interested in calls (allocation or interior call stack
772 // context calls).
773 auto *CI = dyn_cast<CallBase>(&I);
774 if (!CI)
775 continue;
776 auto *CalledFunction = CI->getCalledFunction();
777 if (CalledFunction && CalledFunction->isIntrinsic())
778 continue;
779 // List of call stack ids computed from the location hashes on debug
780 // locations (leaf to inlined at root).
781 std::vector<uint64_t> InlinedCallStack;
782 // Was the leaf location found in one of the profile maps?
783 bool LeafFound = false;
784 // If leaf was found in a map, iterators pointing to its location in both
785 // of the maps. It might exist in neither, one, or both (the latter case
786 // can happen because we don't currently have discriminators to
787 // distinguish the case when a single line/col maps to both an allocation
788 // and another callsite).
789 std::map<uint64_t, std::set<const AllocationInfo *>>::iterator
790 AllocInfoIter;
791 std::map<uint64_t, std::set<std::pair<const SmallVector<Frame> *,
792 unsigned>>>::iterator CallSitesIter;
793 for (const DILocation *DIL = I.getDebugLoc(); DIL != nullptr;
794 DIL = DIL->getInlinedAt()) {
795 // Use C++ linkage name if possible. Need to compile with
796 // -fdebug-info-for-profiling to get linkage name.
797 StringRef Name = DIL->getScope()->getSubprogram()->getLinkageName();
798 if (Name.empty())
799 Name = DIL->getScope()->getSubprogram()->getName();
800 auto CalleeGUID = Function::getGUID(Name);
801 auto StackId =
802 computeStackId(CalleeGUID, GetOffset(DIL), DIL->getColumn());
803 // LeafFound will only be false on the first iteration, since we either
804 // set it true or break out of the loop below.
805 if (!LeafFound) {
806 AllocInfoIter = LocHashToAllocInfo.find(StackId);
807 CallSitesIter = LocHashToCallSites.find(StackId);
808 // Check if the leaf is in one of the maps. If not, no need to look
809 // further at this call.
810 if (AllocInfoIter == LocHashToAllocInfo.end() &&
811 CallSitesIter == LocHashToCallSites.end())
812 break;
813 LeafFound = true;
815 InlinedCallStack.push_back(StackId);
817 // If leaf not in either of the maps, skip inst.
818 if (!LeafFound)
819 continue;
821 // First add !memprof metadata from allocation info, if we found the
822 // instruction's leaf location in that map, and if the rest of the
823 // instruction's locations match the prefix Frame locations on an
824 // allocation context with the same leaf.
825 if (AllocInfoIter != LocHashToAllocInfo.end()) {
826 // Only consider allocations via new, to reduce unnecessary metadata,
827 // since those are the only allocations that will be targeted initially.
828 if (!isNewLikeFn(CI, &TLI))
829 continue;
830 // We may match this instruction's location list to multiple MIB
831 // contexts. Add them to a Trie specialized for trimming the contexts to
832 // the minimal needed to disambiguate contexts with unique behavior.
833 CallStackTrie AllocTrie;
834 for (auto *AllocInfo : AllocInfoIter->second) {
835 // Check the full inlined call stack against this one.
836 // If we found and thus matched all frames on the call, include
837 // this MIB.
838 if (stackFrameIncludesInlinedCallStack(AllocInfo->CallStack,
839 InlinedCallStack))
840 addCallStack(AllocTrie, AllocInfo);
842 // We might not have matched any to the full inlined call stack.
843 // But if we did, create and attach metadata, or a function attribute if
844 // all contexts have identical profiled behavior.
845 if (!AllocTrie.empty()) {
846 // MemprofMDAttached will be false if a function attribute was
847 // attached.
848 bool MemprofMDAttached = AllocTrie.buildAndAttachMIBMetadata(CI);
849 assert(MemprofMDAttached == I.hasMetadata(LLVMContext::MD_memprof));
850 if (MemprofMDAttached) {
851 // Add callsite metadata for the instruction's location list so that
852 // it simpler later on to identify which part of the MIB contexts
853 // are from this particular instruction (including during inlining,
854 // when the callsite metdata will be updated appropriately).
855 // FIXME: can this be changed to strip out the matching stack
856 // context ids from the MIB contexts and not add any callsite
857 // metadata here to save space?
858 addCallsiteMetadata(I, InlinedCallStack, Ctx);
861 continue;
864 // Otherwise, add callsite metadata. If we reach here then we found the
865 // instruction's leaf location in the callsites map and not the allocation
866 // map.
867 assert(CallSitesIter != LocHashToCallSites.end());
868 for (auto CallStackIdx : CallSitesIter->second) {
869 // If we found and thus matched all frames on the call, create and
870 // attach call stack metadata.
871 if (stackFrameIncludesInlinedCallStack(
872 *CallStackIdx.first, InlinedCallStack, CallStackIdx.second)) {
873 addCallsiteMetadata(I, InlinedCallStack, Ctx);
874 // Only need to find one with a matching call stack and add a single
875 // callsite metadata.
876 break;
883 MemProfUsePass::MemProfUsePass(std::string MemoryProfileFile,
884 IntrusiveRefCntPtr<vfs::FileSystem> FS)
885 : MemoryProfileFileName(MemoryProfileFile), FS(FS) {
886 if (!FS)
887 this->FS = vfs::getRealFileSystem();
890 PreservedAnalyses MemProfUsePass::run(Module &M, ModuleAnalysisManager &AM) {
891 LLVM_DEBUG(dbgs() << "Read in memory profile:");
892 auto &Ctx = M.getContext();
893 auto ReaderOrErr = IndexedInstrProfReader::create(MemoryProfileFileName, *FS);
894 if (Error E = ReaderOrErr.takeError()) {
895 handleAllErrors(std::move(E), [&](const ErrorInfoBase &EI) {
896 Ctx.diagnose(
897 DiagnosticInfoPGOProfile(MemoryProfileFileName.data(), EI.message()));
899 return PreservedAnalyses::all();
902 std::unique_ptr<IndexedInstrProfReader> MemProfReader =
903 std::move(ReaderOrErr.get());
904 if (!MemProfReader) {
905 Ctx.diagnose(DiagnosticInfoPGOProfile(
906 MemoryProfileFileName.data(), StringRef("Cannot get MemProfReader")));
907 return PreservedAnalyses::all();
910 if (!MemProfReader->hasMemoryProfile()) {
911 Ctx.diagnose(DiagnosticInfoPGOProfile(MemoryProfileFileName.data(),
912 "Not a memory profile"));
913 return PreservedAnalyses::all();
916 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
918 for (auto &F : M) {
919 if (F.isDeclaration())
920 continue;
922 const TargetLibraryInfo &TLI = FAM.getResult<TargetLibraryAnalysis>(F);
923 readMemprof(M, F, MemProfReader.get(), TLI);
926 return PreservedAnalyses::none();