[MemProf] Optionally save context size info on largest cold allocations (llvm#142507)

teresajohnson · rorth · commit bcea6e2bea24 · 2025-06-11T14:12:05.000+02:00
In order to allow selective reporting of context hinting during the LTO
link, and in the future to allow selective more aggressive cloning, add
an option to specify a minimum percent of the max cold size in the
profile summary. Contexts that meet that threshold will get context size
info metadata (and ThinLTO summary information) on the associated
allocations.

Specifying -memprof-report-hinted-sizes during the pre-LTO compile step
will continue to cause all contexts to receive this metadata. But
specifying -memprof-report-hinted-sizes only during the LTO link will
cause only those that meet the new threshold and have the metadata to
get reported.

To support this, because the alloc info summary and associated bitcode
requires the context size information to be in the same order as the
other context information, 0s are inserted for contexts without this
metadata. The bitcode writer uses a more compact format for the context
ids to allow better compression of the 0s.

As part of this change several helper methods are added to query whether
metadata contains context size info on any or all contexts.
diff --git a/llvm/include/llvm/Analysis/MemoryProfileInfo.h b/llvm/include/llvm/Analysis/MemoryProfileInfo.h
@@ -24,6 +24,18 @@ class OptimizationRemarkEmitter;
 
 namespace memprof {
 
+/// Whether the alloc memeprof metadata will include context size info for all
+/// MIBs.
+LLVM_ABI bool metadataIncludesAllContextSizeInfo();
+
+/// Whether the alloc memprof metadata may include context size info for some
+/// MIBs (but possibly not all).
+LLVM_ABI bool metadataMayIncludeContextSizeInfo();
+
+/// Whether we need to record the context size info in the alloc trie used to
+/// build metadata.
+LLVM_ABI bool recordContextSizeInfoForAnalysis();
+
 /// Build callstack metadata from the provided list of call stack ids. Returns
 /// the resulting metadata node.
 LLVM_ABI MDNode *buildCallstackMetadata(ArrayRef<uint64_t> CallStack,
@@ -87,6 +99,9 @@ class CallStackTrie {
   // allocations for which we apply non-context sensitive allocation hints.
   OptimizationRemarkEmitter *ORE;
 
+  // The maximum size of a cold allocation context, from the profile summary.
+  uint64_t MaxColdSize;
+
   void deleteTrieNode(CallStackTrieNode *Node) {
     if (!Node)
       return;
@@ -113,7 +128,9 @@ class CallStackTrie {
                      uint64_t &ColdBytes);
 
 public:
-  CallStackTrie(OptimizationRemarkEmitter *ORE = nullptr) : ORE(ORE) {}
+  CallStackTrie(OptimizationRemarkEmitter *ORE = nullptr,
+                uint64_t MaxColdSize = 0)
+      : ORE(ORE), MaxColdSize(MaxColdSize) {}
   ~CallStackTrie() { deleteTrieNode(Alloc); }
 
   bool empty() const { return Alloc == nullptr; }
diff --git a/llvm/lib/Analysis/MemoryProfileInfo.cpp b/llvm/lib/Analysis/MemoryProfileInfo.cpp
@@ -46,6 +46,25 @@ cl::opt<unsigned> MinCallsiteColdBytePercent(
     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<unsigned> 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<uint64_t> CallStack,
                                               LLVMContext &Ctx) {
   SmallVector<Metadata *, 8> StackVals;
@@ -168,7 +187,8 @@ void CallStackTrie::addCallStack(MDNode *MIB) {
 static MDNode *createMIBNode(LLVMContext &Ctx, ArrayRef<uint64_t> MIBCallStack,
                              AllocationType AllocType,
                              ArrayRef<ContextTotalSize> ContextSizeInfo,
-                             uint64_t &TotalBytes, uint64_t &ColdBytes) {
+                             const uint64_t MaxColdSize, uint64_t &TotalBytes,
+                             uint64_t &ColdBytes) {
   SmallVector<Metadata *> MIBPayload(
       {buildCallstackMetadata(MIBCallStack, Ctx)});
   MIBPayload.push_back(
@@ -184,12 +204,21 @@ static MDNode *createMIBNode(LLVMContext &Ctx, ArrayRef<uint64_t> MIBCallStack,
 
   for (const auto &[FullStackId, TotalSize] : ContextSizeInfo) {
     TotalBytes += TotalSize;
-    if (AllocType == AllocationType::Cold)
+    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 or we have
-    // specified a threshold for marking allocations cold after cloning).
-    if (MemProfReportHintedSizes || MinClonedColdBytePercent < 100) {
+    // 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(
@@ -357,9 +386,9 @@ bool CallStackTrie::buildMIBNodes(CallStackTrieNode *Node, LLVMContext &Ctx,
   if (hasSingleAllocType(Node->AllocTypes)) {
     std::vector<ContextTotalSize> ContextSizeInfo;
     collectContextSizeInfo(Node, ContextSizeInfo);
-    MIBNodes.push_back(createMIBNode(Ctx, MIBCallStack,
-                                     (AllocationType)Node->AllocTypes,
-                                     ContextSizeInfo, TotalBytes, ColdBytes));
+    MIBNodes.push_back(
+        createMIBNode(Ctx, MIBCallStack, (AllocationType)Node->AllocTypes,
+                      ContextSizeInfo, MaxColdSize, TotalBytes, ColdBytes));
     return true;
   }
 
@@ -413,7 +442,8 @@ bool CallStackTrie::buildMIBNodes(CallStackTrieNode *Node, LLVMContext &Ctx,
   std::vector<ContextTotalSize> ContextSizeInfo;
   collectContextSizeInfo(Node, ContextSizeInfo);
   MIBNodes.push_back(createMIBNode(Ctx, MIBCallStack, AllocationType::NotCold,
-                                   ContextSizeInfo, TotalBytes, ColdBytes));
+                                   ContextSizeInfo, MaxColdSize, TotalBytes,
+                                   ColdBytes));
   return true;
 }
 
diff --git a/llvm/lib/Analysis/ModuleSummaryAnalysis.cpp b/llvm/lib/Analysis/ModuleSummaryAnalysis.cpp
@@ -525,6 +525,7 @@ static void computeFunctionSummary(
       if (MemProfMD) {
         std::vector<MIBInfo> MIBs;
         std::vector<std::vector<ContextTotalSize>> ContextSizeInfos;
+        bool HasNonZeroContextSizeInfos = false;
         for (auto &MDOp : MemProfMD->operands()) {
           auto *MIBMD = cast<const MDNode>(MDOp);
           MDNode *StackNode = getMIBStackNode(MIBMD);
@@ -544,7 +545,8 @@ static void computeFunctionSummary(
           }
           // If we have context size information, collect it for inclusion in
           // the summary.
-          assert(MIBMD->getNumOperands() > 2 || !MemProfReportHintedSizes);
+          assert(MIBMD->getNumOperands() > 2 ||
+                 !metadataIncludesAllContextSizeInfo());
           if (MIBMD->getNumOperands() > 2) {
             std::vector<ContextTotalSize> ContextSizes;
             for (unsigned I = 2; I < MIBMD->getNumOperands(); I++) {
@@ -558,14 +560,31 @@ static void computeFunctionSummary(
                                 ->getZExtValue();
               ContextSizes.push_back({FullStackId, TS});
             }
+            // Flag that we need to keep the ContextSizeInfos array for this
+            // alloc as it now contains non-zero context info sizes.
+            HasNonZeroContextSizeInfos = true;
             ContextSizeInfos.push_back(std::move(ContextSizes));
+          } else {
+            // The ContextSizeInfos must be in the same relative position as the
+            // associated MIB. In some cases we only include a ContextSizeInfo
+            // for a subset of MIBs in an allocation. To handle that, eagerly
+            // fill any MIB entries that don't have context size info metadata
+            // with a pair of 0s. Later on we will only use this array if it
+            // ends up containing any non-zero entries (see where we set
+            // HasNonZeroContextSizeInfos above).
+            ContextSizeInfos.push_back({{0, 0}});
           }
           MIBs.push_back(
               MIBInfo(getMIBAllocType(MIBMD), std::move(StackIdIndices)));
         }
         Allocs.push_back(AllocInfo(std::move(MIBs)));
-        assert(!ContextSizeInfos.empty() || !MemProfReportHintedSizes);
-        if (!ContextSizeInfos.empty()) {
+        assert(HasNonZeroContextSizeInfos ||
+               !metadataIncludesAllContextSizeInfo());
+        // We eagerly build the ContextSizeInfos array, but it will be filled
+        // with sub arrays of pairs of 0s if no MIBs on this alloc actually
+        // contained context size info metadata. Only save it if any MIBs had
+        // any such metadata.
+        if (HasNonZeroContextSizeInfos) {
           assert(Allocs.back().MIBs.size() == ContextSizeInfos.size());
           Allocs.back().ContextSizeInfos = std::move(ContextSizeInfos);
         }
diff --git a/llvm/lib/Bitcode/Reader/BitcodeReader.cpp b/llvm/lib/Bitcode/Reader/BitcodeReader.cpp
@@ -8164,6 +8164,14 @@ Error ModuleSummaryIndexBitcodeReader::parseEntireSummary(unsigned ID) {
           ContextSizes.reserve(NumContextSizeInfoEntries);
           for (unsigned J = 0; J < NumContextSizeInfoEntries; J++) {
             assert(ContextIdIndex < PendingContextIds.size());
+            // Skip any 0 entries for MIBs without the context size info.
+            if (PendingContextIds[ContextIdIndex] == 0) {
+              // The size should also be 0 if the context was 0.
+              assert(!Record[I]);
+              ContextIdIndex++;
+              I++;
+              continue;
+            }
             // PendingContextIds read from the preceding FS_ALLOC_CONTEXT_IDS
             // should be in the same order as the total sizes.
             ContextSizes.push_back(
diff --git a/llvm/lib/Bitcode/Writer/BitcodeWriter.cpp b/llvm/lib/Bitcode/Writer/BitcodeWriter.cpp
@@ -23,6 +23,7 @@
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringMap.h"
 #include "llvm/ADT/StringRef.h"
+#include "llvm/Analysis/MemoryProfileInfo.h"
 #include "llvm/BinaryFormat/Dwarf.h"
 #include "llvm/Bitcode/BitcodeCommon.h"
 #include "llvm/Bitcode/BitcodeReader.h"
@@ -4585,14 +4586,23 @@ void ModuleBitcodeWriterBase::writePerModuleGlobalValueSummary() {
     Stream.EmitRecord(bitc::FS_STACK_IDS, Vals, StackIdAbbvId);
   }
 
-  // n x context id
-  auto ContextIdAbbv = std::make_shared<BitCodeAbbrev>();
-  ContextIdAbbv->Add(BitCodeAbbrevOp(bitc::FS_ALLOC_CONTEXT_IDS));
-  ContextIdAbbv->Add(BitCodeAbbrevOp(BitCodeAbbrevOp::Array));
-  // The context ids are hashes that are close to 64 bits in size, so emitting
-  // as a pair of 32-bit fixed-width values is more efficient than a VBR.
-  ContextIdAbbv->Add(BitCodeAbbrevOp(BitCodeAbbrevOp::Fixed, 32));
-  unsigned ContextIdAbbvId = Stream.EmitAbbrev(std::move(ContextIdAbbv));
+  unsigned ContextIdAbbvId = 0;
+  if (metadataMayIncludeContextSizeInfo()) {
+    // n x context id
+    auto ContextIdAbbv = std::make_shared<BitCodeAbbrev>();
+    ContextIdAbbv->Add(BitCodeAbbrevOp(bitc::FS_ALLOC_CONTEXT_IDS));
+    ContextIdAbbv->Add(BitCodeAbbrevOp(BitCodeAbbrevOp::Array));
+    // The context ids are hashes that are close to 64 bits in size, so emitting
+    // as a pair of 32-bit fixed-width values is more efficient than a VBR if we
+    // are emitting them for all MIBs. Otherwise we use VBR to better compress 0
+    // values that are expected to more frequently occur in an alloc's memprof
+    // summary.
+    if (metadataIncludesAllContextSizeInfo())
+      ContextIdAbbv->Add(BitCodeAbbrevOp(BitCodeAbbrevOp::Fixed, 32));
+    else
+      ContextIdAbbv->Add(BitCodeAbbrevOp(BitCodeAbbrevOp::VBR, 8));
+    ContextIdAbbvId = Stream.EmitAbbrev(std::move(ContextIdAbbv));
+  }
 
   // Abbrev for FS_PERMODULE_PROFILE.
   Abbv = std::make_shared<BitCodeAbbrev>();
diff --git a/llvm/lib/Transforms/IPO/MemProfContextDisambiguation.cpp b/llvm/lib/Transforms/IPO/MemProfContextDisambiguation.cpp
@@ -2232,9 +2232,8 @@ IndexCallsiteContextGraph::IndexCallsiteContextGraph(
           CallStack<MIBInfo, SmallVector<unsigned>::const_iterator>
               EmptyContext;
           unsigned I = 0;
-          assert(
-              (!MemProfReportHintedSizes && MinClonedColdBytePercent >= 100) ||
-              AN.ContextSizeInfos.size() == AN.MIBs.size());
+          assert(!metadataMayIncludeContextSizeInfo() ||
+                 AN.ContextSizeInfos.size() == AN.MIBs.size());
           // Now add all of the MIBs and their stack nodes.
           for (auto &MIB : AN.MIBs) {
             CallStack<MIBInfo, SmallVector<unsigned>::const_iterator>
diff --git a/llvm/lib/Transforms/Instrumentation/MemProfiler.cpp b/llvm/lib/Transforms/Instrumentation/MemProfiler.cpp
@@ -184,10 +184,6 @@ static cl::opt<bool> ClMemProfAttachCalleeGuids(
         "Attach calleeguids as value profile metadata for indirect calls."),
     cl::init(true), cl::Hidden);
 
-extern cl::opt<bool> MemProfReportHintedSizes;
-extern cl::opt<unsigned> MinClonedColdBytePercent;
-extern cl::opt<unsigned> MinCallsiteColdBytePercent;
-
 static cl::opt<unsigned> MinMatchedColdBytePercent(
     "memprof-matching-cold-threshold", cl::init(100), cl::Hidden,
     cl::desc("Min percent of cold bytes matched to hint allocation cold"));
@@ -299,13 +295,6 @@ class ModuleMemProfiler {
   Function *MemProfCtorFunction = nullptr;
 };
 
-// Options under which we need to record the context size info in the alloc trie
-// used to build metadata.
-bool recordContextSizeInfo() {
-  return MemProfReportHintedSizes || MinClonedColdBytePercent < 100 ||
-         MinCallsiteColdBytePercent < 100;
-}
-
 } // end anonymous namespace
 
 MemProfilerPass::MemProfilerPass() = default;
@@ -758,7 +747,7 @@ static AllocationType addCallStack(CallStackTrie &AllocTrie,
                                 AllocInfo->Info.getAllocCount(),
                                 AllocInfo->Info.getTotalLifetime());
   std::vector<ContextTotalSize> ContextSizeInfo;
-  if (recordContextSizeInfo()) {
+  if (recordContextSizeInfoForAnalysis()) {
     auto TotalSize = AllocInfo->Info.getTotalSize();
     assert(TotalSize);
     assert(FullStackId != 0);
@@ -1003,7 +992,7 @@ static void readMemprof(Module &M, Function &F,
                             &FullStackIdToAllocMatchInfo,
                         std::set<std::vector<uint64_t>> &MatchedCallSites,
                         DenseMap<uint64_t, LocToLocMap> &UndriftMaps,
-                        OptimizationRemarkEmitter &ORE) {
+                        OptimizationRemarkEmitter &ORE, uint64_t MaxColdSize) {
   auto &Ctx = M.getContext();
   // Previously we used getIRPGOFuncName() here. If F is local linkage,
   // getIRPGOFuncName() returns FuncName with prefix 'FileName;'. But
@@ -1192,7 +1181,7 @@ static void readMemprof(Module &M, Function &F,
         // We may match this instruction's location list to multiple MIB
         // contexts. Add them to a Trie specialized for trimming the contexts to
         // the minimal needed to disambiguate contexts with unique behavior.
-        CallStackTrie AllocTrie(&ORE);
+        CallStackTrie AllocTrie(&ORE, MaxColdSize);
         uint64_t TotalSize = 0;
         uint64_t TotalColdSize = 0;
         for (auto *AllocInfo : AllocInfoIter->second) {
@@ -1203,7 +1192,7 @@ static void readMemprof(Module &M, Function &F,
                                                  InlinedCallStack)) {
             NumOfMemProfMatchedAllocContexts++;
             uint64_t FullStackId = 0;
-            if (ClPrintMemProfMatchInfo || recordContextSizeInfo())
+            if (ClPrintMemProfMatchInfo || recordContextSizeInfoForAnalysis())
               FullStackId = computeFullStackId(AllocInfo->CallStack);
             auto AllocType = addCallStack(AllocTrie, AllocInfo, FullStackId);
             TotalSize += AllocInfo->Info.getTotalSize();
@@ -1340,14 +1329,18 @@ PreservedAnalyses MemProfUsePass::run(Module &M, ModuleAnalysisManager &AM) {
   // call stack.
   std::set<std::vector<uint64_t>> MatchedCallSites;
 
+  uint64_t MaxColdSize = 0;
+  if (auto *MemProfSum = MemProfReader->getMemProfSummary())
+    MaxColdSize = MemProfSum->getMaxColdTotalSize();
+
   for (auto &F : M) {
     if (F.isDeclaration())
       continue;
 
     const TargetLibraryInfo &TLI = FAM.getResult<TargetLibraryAnalysis>(F);
     auto &ORE = FAM.getResult<OptimizationRemarkEmitterAnalysis>(F);
     readMemprof(M, F, MemProfReader.get(), TLI, FullStackIdToAllocMatchInfo,
-                MatchedCallSites, UndriftMaps, ORE);
+                MatchedCallSites, UndriftMaps, ORE, MaxColdSize);
   }
 
   if (ClPrintMemProfMatchInfo) {
diff --git a/llvm/test/ThinLTO/X86/memprof-report-hinted-partial.ll b/llvm/test/ThinLTO/X86/memprof-report-hinted-partial.ll
@@ -0,0 +1,73 @@
+;; Test that we get hinted size reporting for just the subset of MIBs that
+;; contain context size info in the metadata.
+
+;; Generate the bitcode including ThinLTO summary. Specify
+;; -memprof-min-percent-max-cold-size (value doesn't matter) to indicate to
+;; the bitcode writer that it should expect and optimize for partial context
+;; size info.
+; RUN: opt -thinlto-bc -memprof-min-percent-max-cold-size=50 %s >%t.o
+
+; RUN: llvm-lto2 run %t.o -enable-memprof-context-disambiguation \
+; RUN:	-supports-hot-cold-new \
+; RUN:	-r=%t.o,main,plx \
+; RUN:	-r=%t.o,_Znam, \
+; RUN:	-memprof-report-hinted-sizes \
+; RUN:	-o %t.out 2>&1 | FileCheck %s --check-prefix=SIZES
+
+;; We should only get these two messages from -memprof-report-hinted-sizes
+;; as they are the only MIBs with recorded context size info.
+; SIZES-NOT: full allocation context
+; SIZES: Cold full allocation context 456 with total size 200 is Cold after cloning (context id 2)
+; SIZES: Cold full allocation context 789 with total size 300 is Cold after cloning (context id 2)
+; SIZES-NOT: full allocation context
+
+source_filename = "memprof-report-hinted-partial.ll"
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define i32 @main() #0 {
+entry:
+  %call = call ptr @_Z3foov(), !callsite !0
+  %call1 = call ptr @_Z3foov(), !callsite !1
+  ret i32 0
+}
+
+define internal ptr @_Z3barv() #0 {
+entry:
+  %call = call ptr @_Znam(i64 0), !memprof !2, !callsite !7
+  ret ptr null
+}
+
+declare ptr @_Znam(i64)
+
+define internal ptr @_Z3bazv() #0 {
+entry:
+  %call = call ptr @_Z3barv(), !callsite !8
+  ret ptr null
+}
+
+define internal ptr @_Z3foov() #0 {
+entry:
+  %call = call ptr @_Z3bazv(), !callsite !9
+  ret ptr null
+}
+
+; uselistorder directives
+uselistorder ptr @_Z3foov, { 1, 0 }
+
+attributes #0 = { noinline optnone }
+
+!0 = !{i64 8632435727821051414}
+!1 = !{i64 -3421689549917153178}
+!2 = !{!3, !5, !13}
+!3 = !{!4, !"notcold"}
+!4 = !{i64 9086428284934609951, i64 -5964873800580613432, i64 2732490490862098848, i64 8632435727821051414}
+!5 = !{!6, !"cold", !11, !12}
+!6 = !{i64 9086428284934609951, i64 -5964873800580613432, i64 2732490490862098848, i64 -3421689549917153178}
+!7 = !{i64 9086428284934609951}
+!8 = !{i64 -5964873800580613432}
+!9 = !{i64 2732490490862098848}
+!11 = !{i64 456, i64 200}
+!12 = !{i64 789, i64 300}
+!13 = !{!14, !"cold"}
+!14 = !{i64 9086428284934609951, i64 12345}
diff --git a/llvm/test/Transforms/PGOProfile/memprof_max_cold_threshold.test b/llvm/test/Transforms/PGOProfile/memprof_max_cold_threshold.test
