//===-- MemoryProfileInfo.cpp - memory profile info ------------------------==// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===----------------------------------------------------------------------===// // // This file contains utilities to analyze memory profile information. // //===----------------------------------------------------------------------===// #include "llvm/Analysis/MemoryProfileInfo.h" #include "llvm/Analysis/OptimizationRemarkEmitter.h" #include "llvm/IR/Constants.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Compiler.h" #include "llvm/Support/Format.h" using namespace llvm; using namespace llvm::memprof; #define DEBUG_TYPE "memory-profile-info" cl::opt MemProfReportHintedSizes( "memprof-report-hinted-sizes", cl::init(false), cl::Hidden, cl::desc("Report total allocation sizes of hinted allocations")); // This is useful if we have enabled reporting of hinted sizes, and want to get // information from the indexing step for all contexts (especially for testing), // or have specified a value less than 100% for -memprof-cloning-cold-threshold. LLVM_ABI cl::opt MemProfKeepAllNotColdContexts( "memprof-keep-all-not-cold-contexts", cl::init(false), cl::Hidden, cl::desc("Keep all non-cold contexts (increases cloning overheads)")); cl::opt MinClonedColdBytePercent( "memprof-cloning-cold-threshold", cl::init(100), cl::Hidden, cl::desc("Min percent of cold bytes to hint alloc cold during cloning")); // Discard non-cold contexts if they overlap with much larger cold contexts, // specifically, if all contexts reaching a given callsite are at least this // percent cold byte allocations. This reduces the amount of cloning required // to expose the cold contexts when they greatly dominate non-cold contexts. cl::opt MinCallsiteColdBytePercent( "memprof-callsite-cold-threshold", cl::init(100), cl::Hidden, cl::desc("Min percent of cold bytes at a callsite to discard non-cold " "contexts")); // Enable saving context size information for largest cold contexts, which can // be used to flag contexts for more aggressive cloning and reporting. cl::opt MinPercentMaxColdSize( "memprof-min-percent-max-cold-size", cl::init(100), cl::Hidden, cl::desc("Min percent of max cold bytes for critical cold context")); bool llvm::memprof::metadataIncludesAllContextSizeInfo() { return MemProfReportHintedSizes || MinClonedColdBytePercent < 100; } bool llvm::memprof::metadataMayIncludeContextSizeInfo() { return metadataIncludesAllContextSizeInfo() || MinPercentMaxColdSize < 100; } bool llvm::memprof::recordContextSizeInfoForAnalysis() { return metadataMayIncludeContextSizeInfo() || MinCallsiteColdBytePercent < 100; } MDNode *llvm::memprof::buildCallstackMetadata(ArrayRef CallStack, LLVMContext &Ctx) { SmallVector StackVals; StackVals.reserve(CallStack.size()); for (auto Id : CallStack) { auto *StackValMD = ValueAsMetadata::get(ConstantInt::get(Type::getInt64Ty(Ctx), Id)); StackVals.push_back(StackValMD); } return MDNode::get(Ctx, StackVals); } MDNode *llvm::memprof::getMIBStackNode(const MDNode *MIB) { assert(MIB->getNumOperands() >= 2); // The stack metadata is the first operand of each memprof MIB metadata. return cast(MIB->getOperand(0)); } AllocationType llvm::memprof::getMIBAllocType(const MDNode *MIB) { assert(MIB->getNumOperands() >= 2); // The allocation type is currently the second operand of each memprof // MIB metadata. This will need to change as we add additional allocation // types that can be applied based on the allocation profile data. auto *MDS = dyn_cast(MIB->getOperand(1)); assert(MDS); if (MDS->getString() == "cold") { return AllocationType::Cold; } else if (MDS->getString() == "hot") { return AllocationType::Hot; } return AllocationType::NotCold; } std::string llvm::memprof::getAllocTypeAttributeString(AllocationType Type) { switch (Type) { case AllocationType::NotCold: return "notcold"; break; case AllocationType::Cold: return "cold"; break; case AllocationType::Hot: return "hot"; break; default: assert(false && "Unexpected alloc type"); } llvm_unreachable("invalid alloc type"); } bool llvm::memprof::hasSingleAllocType(uint8_t AllocTypes) { const unsigned NumAllocTypes = llvm::popcount(AllocTypes); assert(NumAllocTypes != 0); return NumAllocTypes == 1; } void CallStackTrie::addCallStack( AllocationType AllocType, ArrayRef StackIds, std::vector ContextSizeInfo) { bool First = true; CallStackTrieNode *Curr = nullptr; for (auto StackId : StackIds) { // If this is the first stack frame, add or update alloc node. if (First) { First = false; if (Alloc) { assert(AllocStackId == StackId); Alloc->addAllocType(AllocType); } else { AllocStackId = StackId; Alloc = new CallStackTrieNode(AllocType); } Curr = Alloc; continue; } // Update existing caller node if it exists. auto [Next, Inserted] = Curr->Callers.try_emplace(StackId); if (!Inserted) { Curr = Next->second; Curr->addAllocType(AllocType); continue; } // Otherwise add a new caller node. auto *New = new CallStackTrieNode(AllocType); Next->second = New; Curr = New; } assert(Curr); llvm::append_range(Curr->ContextSizeInfo, ContextSizeInfo); } void CallStackTrie::addCallStack(MDNode *MIB) { MDNode *StackMD = getMIBStackNode(MIB); assert(StackMD); std::vector CallStack; CallStack.reserve(StackMD->getNumOperands()); for (const auto &MIBStackIter : StackMD->operands()) { auto *StackId = mdconst::dyn_extract(MIBStackIter); assert(StackId); CallStack.push_back(StackId->getZExtValue()); } std::vector ContextSizeInfo; // Collect the context size information if it exists. if (MIB->getNumOperands() > 2) { for (unsigned I = 2; I < MIB->getNumOperands(); I++) { MDNode *ContextSizePair = dyn_cast(MIB->getOperand(I)); assert(ContextSizePair->getNumOperands() == 2); uint64_t FullStackId = mdconst::dyn_extract(ContextSizePair->getOperand(0)) ->getZExtValue(); uint64_t TotalSize = mdconst::dyn_extract(ContextSizePair->getOperand(1)) ->getZExtValue(); ContextSizeInfo.push_back({FullStackId, TotalSize}); } } addCallStack(getMIBAllocType(MIB), CallStack, std::move(ContextSizeInfo)); } static MDNode *createMIBNode(LLVMContext &Ctx, ArrayRef MIBCallStack, AllocationType AllocType, ArrayRef ContextSizeInfo, const uint64_t MaxColdSize, uint64_t &TotalBytes, uint64_t &ColdBytes) { SmallVector MIBPayload( {buildCallstackMetadata(MIBCallStack, Ctx)}); MIBPayload.push_back( MDString::get(Ctx, getAllocTypeAttributeString(AllocType))); if (ContextSizeInfo.empty()) { // The profile matcher should have provided context size info if there was a // MinCallsiteColdBytePercent < 100. Here we check >=100 to gracefully // handle a user-provided percent larger than 100. assert(MinCallsiteColdBytePercent >= 100); return MDNode::get(Ctx, MIBPayload); } for (const auto &[FullStackId, TotalSize] : ContextSizeInfo) { TotalBytes += TotalSize; bool LargeColdContext = false; if (AllocType == AllocationType::Cold) { ColdBytes += TotalSize; // If we have the max cold context size from summary information and have // requested identification of contexts above a percentage of the max, see // if this context qualifies. if (MaxColdSize > 0 && MinPercentMaxColdSize < 100 && TotalSize * 100 >= MaxColdSize * MinPercentMaxColdSize) LargeColdContext = true; } // Only add the context size info as metadata if we need it in the thin // link (currently if reporting of hinted sizes is enabled, we have // specified a threshold for marking allocations cold after cloning, or we // have identified this as a large cold context of interest above). if (metadataIncludesAllContextSizeInfo() || LargeColdContext) { auto *FullStackIdMD = ValueAsMetadata::get( ConstantInt::get(Type::getInt64Ty(Ctx), FullStackId)); auto *TotalSizeMD = ValueAsMetadata::get( ConstantInt::get(Type::getInt64Ty(Ctx), TotalSize)); auto *ContextSizeMD = MDNode::get(Ctx, {FullStackIdMD, TotalSizeMD}); MIBPayload.push_back(ContextSizeMD); } } assert(TotalBytes > 0); return MDNode::get(Ctx, MIBPayload); } void CallStackTrie::collectContextSizeInfo( CallStackTrieNode *Node, std::vector &ContextSizeInfo) { llvm::append_range(ContextSizeInfo, Node->ContextSizeInfo); for (auto &Caller : Node->Callers) collectContextSizeInfo(Caller.second, ContextSizeInfo); } void CallStackTrie::convertHotToNotCold(CallStackTrieNode *Node) { if (Node->hasAllocType(AllocationType::Hot)) { Node->removeAllocType(AllocationType::Hot); Node->addAllocType(AllocationType::NotCold); } for (auto &Caller : Node->Callers) convertHotToNotCold(Caller.second); } // Copy over some or all of NewMIBNodes to the SavedMIBNodes vector, depending // on options that enable filtering out some NotCold contexts. static void saveFilteredNewMIBNodes(std::vector &NewMIBNodes, std::vector &SavedMIBNodes, unsigned CallerContextLength, uint64_t TotalBytes, uint64_t ColdBytes) { const bool MostlyCold = MinCallsiteColdBytePercent < 100 && ColdBytes * 100 >= MinCallsiteColdBytePercent * TotalBytes; // In the simplest case, with pruning disabled, keep all the new MIB nodes. if (MemProfKeepAllNotColdContexts && !MostlyCold) { append_range(SavedMIBNodes, NewMIBNodes); return; } auto EmitMessageForRemovedContexts = [](const MDNode *MIBMD, StringRef Tag, StringRef Extra) { assert(MIBMD->getNumOperands() > 2); for (unsigned I = 2; I < MIBMD->getNumOperands(); I++) { MDNode *ContextSizePair = dyn_cast(MIBMD->getOperand(I)); assert(ContextSizePair->getNumOperands() == 2); uint64_t FullStackId = mdconst::dyn_extract(ContextSizePair->getOperand(0)) ->getZExtValue(); uint64_t TS = mdconst::dyn_extract(ContextSizePair->getOperand(1)) ->getZExtValue(); errs() << "MemProf hinting: Total size for " << Tag << " non-cold full allocation context hash " << FullStackId << Extra << ": " << TS << "\n"; } }; // If the cold bytes at the current callsite exceed the given threshold, we // discard all non-cold contexts so do not need any of the later pruning // handling. We can simply copy over all the cold contexts and return early. if (MostlyCold) { auto NewColdMIBNodes = make_filter_range(NewMIBNodes, [&](const Metadata *M) { auto MIBMD = cast(M); // Only append cold contexts. if (getMIBAllocType(MIBMD) == AllocationType::Cold) return true; if (MemProfReportHintedSizes) { const float PercentCold = ColdBytes * 100.0 / TotalBytes; std::string PercentStr; llvm::raw_string_ostream OS(PercentStr); OS << format(" for %5.2f%% cold bytes", PercentCold); EmitMessageForRemovedContexts(MIBMD, "discarded", OS.str()); } return false; }); for (auto *M : NewColdMIBNodes) SavedMIBNodes.push_back(M); return; } // Prune unneeded NotCold contexts, taking advantage of the fact // that we later will only clone Cold contexts, as NotCold is the allocation // default. We only need to keep as metadata the NotCold contexts that // overlap the longest with Cold allocations, so that we know how deeply we // need to clone. For example, assume we add the following contexts to the // trie: // 1 3 (notcold) // 1 2 4 (cold) // 1 2 5 (notcold) // 1 2 6 (notcold) // the trie looks like: // 1 // / \ // 2 3 // /|\ // 4 5 6 // // It is sufficient to prune all but one not-cold contexts (either 1,2,5 or // 1,2,6, we arbitrarily keep the first one we encounter which will be // 1,2,5). // // To do this pruning, we first check if there were any not-cold // contexts kept for a deeper caller, which will have a context length larger // than the CallerContextLength being handled here (i.e. kept by a deeper // recursion step). If so, none of the not-cold MIB nodes added for the // immediate callers need to be kept. If not, we keep the first (created // for the immediate caller) not-cold MIB node. bool LongerNotColdContextKept = false; for (auto *MIB : NewMIBNodes) { auto MIBMD = cast(MIB); if (getMIBAllocType(MIBMD) == AllocationType::Cold) continue; MDNode *StackMD = getMIBStackNode(MIBMD); assert(StackMD); if (StackMD->getNumOperands() > CallerContextLength) { LongerNotColdContextKept = true; break; } } // Don't need to emit any for the immediate caller if we already have // longer overlapping contexts; bool KeepFirstNewNotCold = !LongerNotColdContextKept; auto NewColdMIBNodes = make_filter_range(NewMIBNodes, [&](const Metadata *M) { auto MIBMD = cast(M); // Only keep cold contexts and first (longest non-cold context). if (getMIBAllocType(MIBMD) != AllocationType::Cold) { MDNode *StackMD = getMIBStackNode(MIBMD); assert(StackMD); // Keep any already kept for longer contexts. if (StackMD->getNumOperands() > CallerContextLength) return true; // Otherwise keep the first one added by the immediate caller if there // were no longer contexts. if (KeepFirstNewNotCold) { KeepFirstNewNotCold = false; return true; } if (MemProfReportHintedSizes) EmitMessageForRemovedContexts(MIBMD, "pruned", ""); return false; } return true; }); for (auto *M : NewColdMIBNodes) SavedMIBNodes.push_back(M); } // Recursive helper to trim contexts and create metadata nodes. // Caller should have pushed Node's loc to MIBCallStack. Doing this in the // caller makes it simpler to handle the many early returns in this method. // Updates the total and cold profiled bytes in the subtrie rooted at this node. bool CallStackTrie::buildMIBNodes(CallStackTrieNode *Node, LLVMContext &Ctx, std::vector &MIBCallStack, std::vector &MIBNodes, bool CalleeHasAmbiguousCallerContext, uint64_t &TotalBytes, uint64_t &ColdBytes) { // Trim context below the first node in a prefix with a single alloc type. // Add an MIB record for the current call stack prefix. if (hasSingleAllocType(Node->AllocTypes)) { std::vector ContextSizeInfo; collectContextSizeInfo(Node, ContextSizeInfo); MIBNodes.push_back( createMIBNode(Ctx, MIBCallStack, (AllocationType)Node->AllocTypes, ContextSizeInfo, MaxColdSize, TotalBytes, ColdBytes)); return true; } // We don't have a single allocation for all the contexts sharing this prefix, // so recursively descend into callers in trie. if (!Node->Callers.empty()) { bool NodeHasAmbiguousCallerContext = Node->Callers.size() > 1; bool AddedMIBNodesForAllCallerContexts = true; // Accumulate all new MIB nodes by the recursive calls below into a vector // that will later be filtered before adding to the caller's MIBNodes // vector. std::vector NewMIBNodes; // Determine the total and cold byte counts for all callers, then add to the // caller's counts further below. uint64_t CallerTotalBytes = 0; uint64_t CallerColdBytes = 0; for (auto &Caller : Node->Callers) { MIBCallStack.push_back(Caller.first); AddedMIBNodesForAllCallerContexts &= buildMIBNodes( Caller.second, Ctx, MIBCallStack, NewMIBNodes, NodeHasAmbiguousCallerContext, CallerTotalBytes, CallerColdBytes); // Remove Caller. MIBCallStack.pop_back(); } // Pass in the stack length of the MIB nodes added for the immediate caller, // which is the current stack length plus 1. saveFilteredNewMIBNodes(NewMIBNodes, MIBNodes, MIBCallStack.size() + 1, CallerTotalBytes, CallerColdBytes); TotalBytes += CallerTotalBytes; ColdBytes += CallerColdBytes; if (AddedMIBNodesForAllCallerContexts) return true; // We expect that the callers should be forced to add MIBs to disambiguate // the context in this case (see below). assert(!NodeHasAmbiguousCallerContext); } // If we reached here, then this node does not have a single allocation type, // and we didn't add metadata for a longer call stack prefix including any of // Node's callers. That means we never hit a single allocation type along all // call stacks with this prefix. This can happen due to recursion collapsing // or the stack being deeper than tracked by the profiler runtime, leading to // contexts with different allocation types being merged. In that case, we // trim the context just below the deepest context split, which is this // node if the callee has an ambiguous caller context (multiple callers), // since the recursive calls above returned false. Conservatively give it // non-cold allocation type. if (!CalleeHasAmbiguousCallerContext) return false; std::vector ContextSizeInfo; collectContextSizeInfo(Node, ContextSizeInfo); MIBNodes.push_back(createMIBNode(Ctx, MIBCallStack, AllocationType::NotCold, ContextSizeInfo, MaxColdSize, TotalBytes, ColdBytes)); return true; } void CallStackTrie::addSingleAllocTypeAttribute(CallBase *CI, AllocationType AT, StringRef Descriptor) { auto AllocTypeString = getAllocTypeAttributeString(AT); auto A = llvm::Attribute::get(CI->getContext(), "memprof", AllocTypeString); CI->addFnAttr(A); if (MemProfReportHintedSizes) { std::vector ContextSizeInfo; collectContextSizeInfo(Alloc, ContextSizeInfo); for (const auto &[FullStackId, TotalSize] : ContextSizeInfo) { errs() << "MemProf hinting: Total size for full allocation context hash " << FullStackId << " and " << Descriptor << " alloc type " << getAllocTypeAttributeString(AT) << ": " << TotalSize << "\n"; } } if (ORE) ORE->emit(OptimizationRemark(DEBUG_TYPE, "MemprofAttribute", CI) << ore::NV("AllocationCall", CI) << " in function " << ore::NV("Caller", CI->getFunction()) << " marked with memprof allocation attribute " << ore::NV("Attribute", AllocTypeString)); } // Build and attach the minimal necessary MIB metadata. If the alloc has a // single allocation type, add a function attribute instead. Returns true if // memprof metadata attached, false if not (attribute added). bool CallStackTrie::buildAndAttachMIBMetadata(CallBase *CI) { if (hasSingleAllocType(Alloc->AllocTypes)) { addSingleAllocTypeAttribute(CI, (AllocationType)Alloc->AllocTypes, "single"); return false; } // If there were any hot allocation contexts, the Alloc trie node would have // the Hot type set. If so, because we don't currently support cloning for hot // contexts, they should be converted to NotCold. This happens in the cloning // support anyway, however, doing this now enables more aggressive context // trimming when building the MIB metadata (and possibly may make the // allocation have a single NotCold allocation type), greatly reducing // overheads in bitcode, cloning memory and cloning time. if (Alloc->hasAllocType(AllocationType::Hot)) { convertHotToNotCold(Alloc); // Check whether we now have a single alloc type. if (hasSingleAllocType(Alloc->AllocTypes)) { addSingleAllocTypeAttribute(CI, (AllocationType)Alloc->AllocTypes, "single"); return false; } } auto &Ctx = CI->getContext(); std::vector MIBCallStack; MIBCallStack.push_back(AllocStackId); std::vector MIBNodes; uint64_t TotalBytes = 0; uint64_t ColdBytes = 0; assert(!Alloc->Callers.empty() && "addCallStack has not been called yet"); // The CalleeHasAmbiguousCallerContext flag is meant to say whether the // callee of the given node has more than one caller. Here the node being // passed in is the alloc and it has no callees. So it's false. if (buildMIBNodes(Alloc, Ctx, MIBCallStack, MIBNodes, /*CalleeHasAmbiguousCallerContext=*/false, TotalBytes, ColdBytes)) { assert(MIBCallStack.size() == 1 && "Should only be left with Alloc's location in stack"); CI->setMetadata(LLVMContext::MD_memprof, MDNode::get(Ctx, MIBNodes)); return true; } // If there exists corner case that CallStackTrie has one chain to leaf // and all node in the chain have multi alloc type, conservatively give // it non-cold allocation type. // FIXME: Avoid this case before memory profile created. Alternatively, select // hint based on fraction cold. addSingleAllocTypeAttribute(CI, AllocationType::NotCold, "indistinguishable"); return false; } template <> CallStack::CallStackIterator::CallStackIterator( const MDNode *N, bool End) : N(N) { if (!N) return; Iter = End ? N->op_end() : N->op_begin(); } template <> uint64_t CallStack::CallStackIterator::operator*() { assert(Iter != N->op_end()); ConstantInt *StackIdCInt = mdconst::dyn_extract(*Iter); assert(StackIdCInt); return StackIdCInt->getZExtValue(); } template <> uint64_t CallStack::back() const { assert(N); return mdconst::dyn_extract(N->operands().back()) ->getZExtValue(); } MDNode *MDNode::getMergedMemProfMetadata(MDNode *A, MDNode *B) { // TODO: Support more sophisticated merging, such as selecting the one with // more bytes allocated, or implement support for carrying multiple allocation // leaf contexts. For now, keep the first one. if (A) return A; return B; } MDNode *MDNode::getMergedCallsiteMetadata(MDNode *A, MDNode *B) { // TODO: Support more sophisticated merging, which will require support for // carrying multiple contexts. For now, keep the first one. if (A) return A; return B; }