| 1 | //==-- MemProfContextDisambiguation.cpp - Disambiguate contexts -------------=// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file implements support for context disambiguation of allocation |
| 10 | // calls for profile-guided memory optimization (MemPGO). Specifically, it uses |
| 11 | // MemProf profiles which indicate context specific allocation behavior |
| 12 | // (currently distinguishing cold vs hot memory allocations). Cloning is |
| 13 | // performed to expose the cold allocation call contexts, and the allocation |
| 14 | // calls are subsequently annotated with an attribute for later transformation. |
| 15 | // |
| 16 | // The transformations can be performed either directly on IR (regular LTO), or |
| 17 | // on a ThinLTO index (and later applied to the IR during the ThinLTO backend). |
| 18 | // Both types of LTO operate on a the same base graph representation, which |
| 19 | // uses CRTP to support either IR or Index formats. |
| 20 | // |
| 21 | //===----------------------------------------------------------------------===// |
| 22 | |
| 23 | #include "llvm/Transforms/IPO/MemProfContextDisambiguation.h" |
| 24 | #include "llvm/ADT/DenseMap.h" |
| 25 | #include "llvm/ADT/DenseSet.h" |
| 26 | #include "llvm/ADT/MapVector.h" |
| 27 | #include "llvm/ADT/SetOperations.h" |
| 28 | #include "llvm/ADT/SmallPtrSet.h" |
| 29 | #include "llvm/ADT/SmallSet.h" |
| 30 | #include "llvm/ADT/SmallVector.h" |
| 31 | #include "llvm/ADT/Statistic.h" |
| 32 | #include "llvm/ADT/StringExtras.h" |
| 33 | #include "llvm/Analysis/MemoryProfileInfo.h" |
| 34 | #include "llvm/Analysis/ModuleSummaryAnalysis.h" |
| 35 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
| 36 | #include "llvm/Bitcode/BitcodeReader.h" |
| 37 | #include "llvm/IR/Instructions.h" |
| 38 | #include "llvm/IR/Module.h" |
| 39 | #include "llvm/IR/ModuleSummaryIndex.h" |
| 40 | #include "llvm/Pass.h" |
| 41 | #include "llvm/Support/CommandLine.h" |
| 42 | #include "llvm/Support/GraphWriter.h" |
| 43 | #include "llvm/Support/SHA1.h" |
| 44 | #include "llvm/Support/raw_ostream.h" |
| 45 | #include "llvm/Transforms/IPO.h" |
| 46 | #include "llvm/Transforms/Utils/CallPromotionUtils.h" |
| 47 | #include "llvm/Transforms/Utils/Cloning.h" |
| 48 | #include "llvm/Transforms/Utils/Instrumentation.h" |
| 49 | #include <deque> |
| 50 | #include <sstream> |
| 51 | #include <vector> |
| 52 | using namespace llvm; |
| 53 | using namespace llvm::memprof; |
| 54 | |
| 55 | #define DEBUG_TYPE "memprof-context-disambiguation" |
| 56 | |
| 57 | STATISTIC(FunctionClonesAnalysis, |
| 58 | "Number of function clones created during whole program analysis" ); |
| 59 | STATISTIC(FunctionClonesThinBackend, |
| 60 | "Number of function clones created during ThinLTO backend" ); |
| 61 | STATISTIC(FunctionsClonedThinBackend, |
| 62 | "Number of functions that had clones created during ThinLTO backend" ); |
| 63 | STATISTIC( |
| 64 | FunctionCloneDuplicatesThinBackend, |
| 65 | "Number of function clone duplicates detected during ThinLTO backend" ); |
| 66 | STATISTIC(AllocTypeNotCold, "Number of not cold static allocations (possibly " |
| 67 | "cloned) during whole program analysis" ); |
| 68 | STATISTIC(AllocTypeCold, "Number of cold static allocations (possibly cloned) " |
| 69 | "during whole program analysis" ); |
| 70 | STATISTIC(AllocTypeAmbiguous, |
| 71 | "Number of ambiguous static allocations (possibly cloned) during " |
| 72 | "whole program analysis" ); |
| 73 | STATISTIC(AllocTypeNotColdThinBackend, |
| 74 | "Number of not cold static allocations (possibly cloned) during " |
| 75 | "ThinLTO backend" ); |
| 76 | STATISTIC(AllocTypeColdThinBackend, "Number of cold static allocations " |
| 77 | "(possibly cloned) during ThinLTO backend" ); |
| 78 | STATISTIC(AllocTypeAmbiguousThinBackend, |
| 79 | "Number of ambiguous static allocations (possibly cloned) during " |
| 80 | "ThinLTO backend" ); |
| 81 | STATISTIC(OrigAllocsThinBackend, |
| 82 | "Number of original (not cloned) allocations with memprof profiles " |
| 83 | "during ThinLTO backend" ); |
| 84 | STATISTIC( |
| 85 | AllocVersionsThinBackend, |
| 86 | "Number of allocation versions (including clones) during ThinLTO backend" ); |
| 87 | STATISTIC(MaxAllocVersionsThinBackend, |
| 88 | "Maximum number of allocation versions created for an original " |
| 89 | "allocation during ThinLTO backend" ); |
| 90 | STATISTIC(UnclonableAllocsThinBackend, |
| 91 | "Number of unclonable ambigous allocations during ThinLTO backend" ); |
| 92 | STATISTIC(RemovedEdgesWithMismatchedCallees, |
| 93 | "Number of edges removed due to mismatched callees (profiled vs IR)" ); |
| 94 | STATISTIC(FoundProfiledCalleeCount, |
| 95 | "Number of profiled callees found via tail calls" ); |
| 96 | STATISTIC(FoundProfiledCalleeDepth, |
| 97 | "Aggregate depth of profiled callees found via tail calls" ); |
| 98 | STATISTIC(FoundProfiledCalleeMaxDepth, |
| 99 | "Maximum depth of profiled callees found via tail calls" ); |
| 100 | STATISTIC(FoundProfiledCalleeNonUniquelyCount, |
| 101 | "Number of profiled callees found via multiple tail call chains" ); |
| 102 | STATISTIC(FoundProfiledCalleeIndirectTailCallCount, |
| 103 | "Number of tail calls in chains to profiled callees that are " |
| 104 | "indirect (thin link)" ); |
| 105 | STATISTIC(IndirectTailCallsRecordedForICPThinBackend, |
| 106 | "Number of indirect tail calls recorded for ICP for cloning during " |
| 107 | "ThinLTO backend" ); |
| 108 | STATISTIC(DeferredBackedges, "Number of backedges with deferred cloning" ); |
| 109 | STATISTIC(NewMergedNodes, "Number of new nodes created during merging" ); |
| 110 | STATISTIC(NonNewMergedNodes, "Number of non new nodes used during merging" ); |
| 111 | STATISTIC(MissingAllocForContextId, |
| 112 | "Number of missing alloc nodes for context ids" ); |
| 113 | STATISTIC(SkippedCallsCloning, |
| 114 | "Number of calls skipped during cloning due to unexpected operand" ); |
| 115 | STATISTIC(MismatchedCloneAssignments, |
| 116 | "Number of callsites assigned to call multiple non-matching clones" ); |
| 117 | STATISTIC(TotalMergeInvokes, "Number of merge invocations for nodes" ); |
| 118 | STATISTIC(TotalMergeIters, "Number of merge iterations for nodes" ); |
| 119 | STATISTIC(MaxMergeIters, "Max merge iterations for nodes" ); |
| 120 | STATISTIC(NumImportantContextIds, "Number of important context ids" ); |
| 121 | STATISTIC(NumFixupEdgeIdsInserted, "Number of fixup edge ids inserted" ); |
| 122 | STATISTIC(NumFixupEdgesAdded, "Number of fixup edges added" ); |
| 123 | STATISTIC(NumFixedContexts, "Number of contexts with fixed edges" ); |
| 124 | STATISTIC(NumImplicitCallersAssigned, |
| 125 | "Number of callsite clones without an edge to a callee's clones " |
| 126 | "assigned the callee function clone of their original node" ); |
| 127 | STATISTIC(, |
| 128 | "Number of calls in function clones whose callee was taken from a " |
| 129 | "callsite clone an implicit caller requires" ); |
| 130 | STATISTIC(NumTrimmedContextIds, |
| 131 | "Number of context ids with call stacks trimmed at a callsite" ); |
| 132 | STATISTIC(NumImplicitCallersUnresolved, |
| 133 | "Number of callsite clones without an edge to a callee's clones " |
| 134 | "left in a function clone calling a conflicting callee clone" ); |
| 135 | STATISTIC(AliaseesPrevailingInDiffModuleFromAlias, |
| 136 | "Number of aliasees prevailing in a different module than its alias" ); |
| 137 | STATISTIC(NumClonedCalleeCallsites, |
| 138 | "Number of summary callsites assigned to call a cloned callee (thin " |
| 139 | "link)" ); |
| 140 | STATISTIC(NumClonedCalleeIndirectCallsites, |
| 141 | "Number of summary callsites assigned to call a cloned callee with " |
| 142 | "an indirect call edge (thin link)" ); |
| 143 | STATISTIC(NumClonedCalleeEdgesMarkedRequiresDeclImport, |
| 144 | "Number of summary indirect call edges to cloned callees marked as " |
| 145 | "requiring a declaration import (thin link)" ); |
| 146 | STATISTIC(NumCallsAssignedTrimmedInlinedContexts, |
| 147 | "Number of calls assigned contexts trimmed within their inlined " |
| 148 | "stack id sequence" ); |
| 149 | STATISTIC(NumTrimmedInlinedContextIds, |
| 150 | "Number of (duplicated) context ids for contexts trimmed within a " |
| 151 | "call's inlined stack id sequence" ); |
| 152 | |
| 153 | static cl::opt<std::string> DotFilePathPrefix( |
| 154 | "memprof-dot-file-path-prefix" , cl::init(Val: "" ), cl::Hidden, |
| 155 | cl::value_desc("filename" ), |
| 156 | cl::desc("Specify the path prefix of the MemProf dot files." )); |
| 157 | |
| 158 | static cl::opt<bool> ExportToDot("memprof-export-to-dot" , cl::init(Val: false), |
| 159 | cl::Hidden, |
| 160 | cl::desc("Export graph to dot files." )); |
| 161 | |
| 162 | // TODO: Remove this option once new handling is validated more widely. |
| 163 | static cl::opt<bool> DoMergeIteration( |
| 164 | "memprof-merge-iteration" , cl::init(Val: true), cl::Hidden, |
| 165 | cl::desc("Iteratively apply merging on a node to catch new callers" )); |
| 166 | |
| 167 | // How much of the graph to export to dot. |
| 168 | enum DotScope { |
| 169 | All, // The full CCG graph. |
| 170 | Alloc, // Only contexts for the specified allocation. |
| 171 | Context, // Only the specified context. |
| 172 | }; |
| 173 | |
| 174 | static cl::opt<DotScope> DotGraphScope( |
| 175 | "memprof-dot-scope" , cl::desc("Scope of graph to export to dot" ), |
| 176 | cl::Hidden, cl::init(Val: DotScope::All), |
| 177 | cl::values( |
| 178 | clEnumValN(DotScope::All, "all" , "Export full callsite graph" ), |
| 179 | clEnumValN(DotScope::Alloc, "alloc" , |
| 180 | "Export only nodes with contexts feeding given " |
| 181 | "-memprof-dot-alloc-id" ), |
| 182 | clEnumValN(DotScope::Context, "context" , |
| 183 | "Export only nodes with given -memprof-dot-context-id" ))); |
| 184 | |
| 185 | static cl::opt<unsigned> |
| 186 | AllocIdForDot("memprof-dot-alloc-id" , cl::init(Val: 0), cl::Hidden, |
| 187 | cl::desc("Id of alloc to export if -memprof-dot-scope=alloc " |
| 188 | "or to highlight if -memprof-dot-scope=all" )); |
| 189 | |
| 190 | static cl::opt<unsigned> ContextIdForDot( |
| 191 | "memprof-dot-context-id" , cl::init(Val: 0), cl::Hidden, |
| 192 | cl::desc("Id of context to export if -memprof-dot-scope=context or to " |
| 193 | "highlight otherwise" )); |
| 194 | |
| 195 | static cl::opt<bool> |
| 196 | DumpCCG("memprof-dump-ccg" , cl::init(Val: false), cl::Hidden, |
| 197 | cl::desc("Dump CallingContextGraph to stdout after each stage." )); |
| 198 | |
| 199 | static cl::opt<bool> |
| 200 | VerifyCCG("memprof-verify-ccg" , cl::init(Val: false), cl::Hidden, |
| 201 | cl::desc("Perform verification checks on CallingContextGraph." )); |
| 202 | |
| 203 | static cl::opt<bool> |
| 204 | VerifyNodes("memprof-verify-nodes" , cl::init(Val: false), cl::Hidden, |
| 205 | cl::desc("Perform frequent verification checks on nodes." )); |
| 206 | |
| 207 | static cl::opt<std::string> MemProfImportSummary( |
| 208 | "memprof-import-summary" , |
| 209 | cl::desc("Import summary to use for testing the ThinLTO backend via opt" ), |
| 210 | cl::Hidden); |
| 211 | |
| 212 | static cl::opt<unsigned> |
| 213 | TailCallSearchDepth("memprof-tail-call-search-depth" , cl::init(Val: 5), |
| 214 | cl::Hidden, |
| 215 | cl::desc("Max depth to recursively search for missing " |
| 216 | "frames through tail calls." )); |
| 217 | |
| 218 | // Whether the ThinLTO search for frames missing due to tail calls should look |
| 219 | // through indirect tail calls, via the summary call edges to their profiled |
| 220 | // targets. |
| 221 | static cl::opt<bool> IndirectTailCallSearch( |
| 222 | "memprof-indirect-tail-call-search" , cl::init(Val: true), cl::Hidden, |
| 223 | cl::desc("Search for missing frames through indirect tail calls to their " |
| 224 | "profiled targets (ThinLTO)." )); |
| 225 | |
| 226 | // Optionally enable cloning of callsites involved with recursive cycles |
| 227 | static cl::opt<bool> AllowRecursiveCallsites( |
| 228 | "memprof-allow-recursive-callsites" , cl::init(Val: true), cl::Hidden, |
| 229 | cl::desc("Allow cloning of callsites involved in recursive cycles" )); |
| 230 | |
| 231 | static cl::opt<bool> CloneRecursiveContexts( |
| 232 | "memprof-clone-recursive-contexts" , cl::init(Val: true), cl::Hidden, |
| 233 | cl::desc("Allow cloning of contexts through recursive cycles" )); |
| 234 | |
| 235 | // Generally this is needed for correct assignment of allocation clones to |
| 236 | // function clones, however, allow it to be disabled for debugging while the |
| 237 | // functionality is new and being tested more widely. |
| 238 | static cl::opt<bool> |
| 239 | MergeClones("memprof-merge-clones" , cl::init(Val: true), cl::Hidden, |
| 240 | cl::desc("Merge clones before assigning functions" )); |
| 241 | |
| 242 | // When disabled, try to detect and prevent cloning of recursive contexts. |
| 243 | // This is only necessary until we support cloning through recursive cycles. |
| 244 | // Leave on by default for now, as disabling requires a little bit of compile |
| 245 | // time overhead and doesn't affect correctness, it will just inflate the cold |
| 246 | // hinted bytes reporting a bit when -memprof-report-hinted-sizes is enabled. |
| 247 | static cl::opt<bool> AllowRecursiveContexts( |
| 248 | "memprof-allow-recursive-contexts" , cl::init(Val: true), cl::Hidden, |
| 249 | cl::desc("Allow cloning of contexts having recursive cycles" )); |
| 250 | |
| 251 | // When updating the stack nodes for callsites, a call with inlined frames is |
| 252 | // assigned the contexts that include all of its stack ids that have context |
| 253 | // nodes. The profile may have trimmed some contexts at an earlier stack id in |
| 254 | // that sequence (because they had a single allocation type from that point |
| 255 | // on); when a later stack id has a node due to other contexts, those trimmed |
| 256 | // contexts are not assigned to the call, only to calls whose last stack id |
| 257 | // with a node is the earlier id (e.g. an out-of-line copy of the inlined |
| 258 | // callee). When enabled, also assign them to the call with the inlined frames, |
| 259 | // via duplicates of their context ids (the other calls keep the originals). |
| 260 | // TODO: Remove this option once the new handling is validated more widely. |
| 261 | static cl::opt<bool> AssignTrimmedInlinedContexts( |
| 262 | "memprof-assign-trimmed-inlined-contexts" , cl::init(Val: true), cl::Hidden, |
| 263 | cl::desc("Assign contexts trimmed within a call's inlined stack id " |
| 264 | "sequence to that call" )); |
| 265 | |
| 266 | // Set the minimum absolute count threshold for allowing inlining of indirect |
| 267 | // calls promoted during cloning. |
| 268 | static cl::opt<unsigned> MemProfICPNoInlineThreshold( |
| 269 | "memprof-icp-noinline-threshold" , cl::init(Val: 0), cl::Hidden, |
| 270 | cl::desc("Minimum absolute count for promoted target to be inlinable" )); |
| 271 | |
| 272 | namespace llvm { |
| 273 | cl::opt<bool> EnableMemProfContextDisambiguation( |
| 274 | "enable-memprof-context-disambiguation" , cl::Hidden, |
| 275 | cl::desc("Enable MemProf context disambiguation" )); |
| 276 | |
| 277 | // Indicate we are linking with an allocator that supports hot/cold operator |
| 278 | // new interfaces. |
| 279 | cl::opt<bool> SupportsHotColdNew( |
| 280 | "supports-hot-cold-new" , cl::init(Val: false), cl::Hidden, |
| 281 | cl::desc("Linking with hot/cold operator new interfaces" )); |
| 282 | |
| 283 | static cl::opt<bool> MemProfRequireDefinitionForPromotion( |
| 284 | "memprof-require-definition-for-promotion" , cl::init(Val: false), cl::Hidden, |
| 285 | cl::desc( |
| 286 | "Require target function definition when promoting indirect calls" )); |
| 287 | |
| 288 | // Applying the thin link cloning decisions to indirect calls requires the |
| 289 | // ThinLTO backend to perform ICP to the profiled target, which requires that |
| 290 | // the target be available in the caller's module (or in any module the caller |
| 291 | // is imported into). When the target is defined in a different module it must |
| 292 | // be imported, but the profiled targets are often synthesized from the MemProf |
| 293 | // profile with a small count, so their summary edges are not hot and large |
| 294 | // targets are often not imported. Since the thin link cloning runs before the |
| 295 | // function importing analysis, we can mark the indirect call edges to callees |
| 296 | // of callsites that were assigned to call a cloned callee as requiring a |
| 297 | // declaration import, in which case the function importer imports at least a |
| 298 | // declaration of the callee into any module containing the caller that |
| 299 | // doesn't define the callee. Direct calls don't require importing the callee, |
| 300 | // since the ThinLTO backend can simply call a declaration of the callee clone, |
| 301 | // so we only do this for edges marked as being from indirect calls in the |
| 302 | // summary. Note that unlike regular ICP, which by default requires the |
| 303 | // target's definition (see -icp-allow-decls), MemProf ICP can promote to a |
| 304 | // declaration of the target (see -memprof-require-definition-for-promotion), |
| 305 | // since the purpose is to call the assigned clone rather than to inline. This |
| 306 | // is therefore not done under -memprof-require-definition-for-promotion. |
| 307 | static cl::opt<bool> MemProfImportClonedIndirectTargets( |
| 308 | "memprof-import-cloned-indirect-targets" , cl::init(Val: true), cl::Hidden, |
| 309 | cl::desc("Import at least a declaration of the indirect call targets of " |
| 310 | "callsites assigned to call a cloned callee (needed for MemProf " |
| 311 | "ICP in the ThinLTO backend)" )); |
| 312 | |
| 313 | extern cl::opt<bool> MemProfReportHintedSizes; |
| 314 | extern cl::opt<unsigned> MinClonedColdBytePercent; |
| 315 | extern cl::opt<bool> MemProfUseAmbiguousAttributes; |
| 316 | |
| 317 | // Appends a per-context diagnostic suffix to each -memprof-report-hinted-sizes |
| 318 | // line: the allocation function clone, how far the context matched, and the |
| 319 | // post-cloning call paths from the outermost matched caller to the reached |
| 320 | // allocation clone(s) and their final hints. Requires |
| 321 | // -memprof-report-hinted-sizes, and adds some memory and time overhead. |
| 322 | static cl::opt<bool> MemProfReportHintedSizesDetail( |
| 323 | "memprof-report-hinted-sizes-detail" , cl::init(Val: false), cl::Hidden, |
| 324 | cl::desc("Append post-cloning call path and matching details to each " |
| 325 | "-memprof-report-hinted-sizes line" )); |
| 326 | |
| 327 | // Post-cloning call paths in the detail suffix longer than this (in functions) |
| 328 | // have runs of uncloned (or cloned) functions in their middle elided to |
| 329 | // "...", keeping the ends and every cloned/uncloned transition. |
| 330 | static cl::opt<unsigned> MemProfReportHintedSizesDetailMaxPathLen( |
| 331 | "memprof-report-hinted-sizes-detail-max-path-len" , cl::init(Val: 7), cl::Hidden, |
| 332 | cl::desc("Maximum number of functions printed in full for a post-cloning " |
| 333 | "call path in the -memprof-report-hinted-sizes-detail suffix; " |
| 334 | "longer paths are elided in the middle" )); |
| 335 | |
| 336 | cl::opt<unsigned> MemProfTopNImportant( |
| 337 | "memprof-top-n-important" , cl::init(Val: 10), cl::Hidden, |
| 338 | cl::desc("Number of largest cold contexts to consider important" )); |
| 339 | |
| 340 | cl::opt<bool> MemProfFixupImportant( |
| 341 | "memprof-fixup-important" , cl::init(Val: true), cl::Hidden, |
| 342 | cl::desc("Enables edge fixup for important contexts" )); |
| 343 | |
| 344 | } // namespace llvm |
| 345 | |
| 346 | namespace { |
| 347 | |
| 348 | /// CRTP base for graphs built from either IR or ThinLTO summary index. |
| 349 | /// |
| 350 | /// The graph represents the call contexts in all memprof metadata on allocation |
| 351 | /// calls, with nodes for the allocations themselves, as well as for the calls |
| 352 | /// in each context. The graph is initially built from the allocation memprof |
| 353 | /// metadata (or summary) MIBs. It is then updated to match calls with callsite |
| 354 | /// metadata onto the nodes, updating it to reflect any inlining performed on |
| 355 | /// those calls. |
| 356 | /// |
| 357 | /// Each MIB (representing an allocation's call context with allocation |
| 358 | /// behavior) is assigned a unique context id during the graph build. The edges |
| 359 | /// and nodes in the graph are decorated with the context ids they carry. This |
| 360 | /// is used to correctly update the graph when cloning is performed so that we |
| 361 | /// can uniquify the context for a single (possibly cloned) allocation. |
| 362 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 363 | class CallsiteContextGraph { |
| 364 | public: |
| 365 | CallsiteContextGraph() = default; |
| 366 | CallsiteContextGraph(const CallsiteContextGraph &) = default; |
| 367 | CallsiteContextGraph(CallsiteContextGraph &&) = default; |
| 368 | |
| 369 | /// Main entry point to perform analysis and transformations on graph. |
| 370 | bool process(function_ref<void(StringRef, StringRef, const Twine &)> |
| 371 | = nullptr, |
| 372 | bool = false); |
| 373 | |
| 374 | /// Perform cloning on the graph necessary to uniquely identify the allocation |
| 375 | /// behavior of an allocation based on its context. |
| 376 | void identifyClones(); |
| 377 | |
| 378 | /// Assign callsite clones to functions, cloning functions as needed to |
| 379 | /// accommodate the combinations of their callsite clones reached by callers. |
| 380 | /// For regular LTO this clones functions and callsites in the IR, but for |
| 381 | /// ThinLTO the cloning decisions are noted in the summaries and later applied |
| 382 | /// in applyImport. |
| 383 | bool assignFunctions(); |
| 384 | |
| 385 | void dump() const; |
| 386 | void print(raw_ostream &OS) const; |
| 387 | void printTotalSizes(raw_ostream &OS, |
| 388 | function_ref<void(StringRef, StringRef, const Twine &)> |
| 389 | = nullptr) const; |
| 390 | |
| 391 | friend raw_ostream &operator<<(raw_ostream &OS, |
| 392 | const CallsiteContextGraph &CCG) { |
| 393 | CCG.print(OS); |
| 394 | return OS; |
| 395 | } |
| 396 | |
| 397 | friend struct GraphTraits< |
| 398 | const CallsiteContextGraph<DerivedCCG, FuncTy, CallTy> *>; |
| 399 | friend struct DOTGraphTraits< |
| 400 | const CallsiteContextGraph<DerivedCCG, FuncTy, CallTy> *>; |
| 401 | |
| 402 | void exportToDot(std::string Label) const; |
| 403 | |
| 404 | /// Represents a function clone via FuncTy pointer and clone number pair. |
| 405 | struct FuncInfo final |
| 406 | : public std::pair<FuncTy *, unsigned /*Clone number*/> { |
| 407 | using Base = std::pair<FuncTy *, unsigned>; |
| 408 | FuncInfo(const Base &B) : Base(B) {} |
| 409 | FuncInfo(FuncTy *F = nullptr, unsigned CloneNo = 0) : Base(F, CloneNo) {} |
| 410 | explicit operator bool() const { return this->first != nullptr; } |
| 411 | FuncTy *func() const { return this->first; } |
| 412 | unsigned cloneNo() const { return this->second; } |
| 413 | }; |
| 414 | |
| 415 | /// Represents a callsite clone via CallTy and clone number pair. |
| 416 | struct CallInfo final : public std::pair<CallTy, unsigned /*Clone number*/> { |
| 417 | using Base = std::pair<CallTy, unsigned>; |
| 418 | CallInfo(const Base &B) : Base(B) {} |
| 419 | CallInfo(CallTy Call = nullptr, unsigned CloneNo = 0) |
| 420 | : Base(Call, CloneNo) {} |
| 421 | explicit operator bool() const { return (bool)this->first; } |
| 422 | CallTy call() const { return this->first; } |
| 423 | unsigned cloneNo() const { return this->second; } |
| 424 | void setCloneNo(unsigned N) { this->second = N; } |
| 425 | void print(raw_ostream &OS) const { |
| 426 | if (!operator bool()) { |
| 427 | assert(!cloneNo()); |
| 428 | OS << "null Call" ; |
| 429 | return; |
| 430 | } |
| 431 | call()->print(OS); |
| 432 | OS << "\t(clone " << cloneNo() << ")" ; |
| 433 | } |
| 434 | void dump() const { |
| 435 | print(OS&: dbgs()); |
| 436 | dbgs() << "\n" ; |
| 437 | } |
| 438 | friend raw_ostream &operator<<(raw_ostream &OS, const CallInfo &Call) { |
| 439 | Call.print(OS); |
| 440 | return OS; |
| 441 | } |
| 442 | }; |
| 443 | |
| 444 | struct ContextEdge; |
| 445 | |
| 446 | /// Node in the Callsite Context Graph |
| 447 | struct ContextNode { |
| 448 | // Assigned to nodes as they are created, useful for debugging. |
| 449 | unsigned NodeId = 0; |
| 450 | |
| 451 | // Keep this for now since in the IR case where we have an Instruction* it |
| 452 | // is not as immediately discoverable. Used for printing richer information |
| 453 | // when dumping graph. |
| 454 | bool IsAllocation; |
| 455 | |
| 456 | // Keeps track of when the Call was reset to null because there was |
| 457 | // recursion. |
| 458 | bool Recursive = false; |
| 459 | |
| 460 | // This will be formed by ORing together the AllocationType enum values |
| 461 | // for contexts including this node. |
| 462 | uint8_t AllocTypes = 0; |
| 463 | |
| 464 | // The corresponding allocation or interior call. This is the primary call |
| 465 | // for which we have created this node. |
| 466 | CallInfo Call; |
| 467 | |
| 468 | // List of other calls that can be treated the same as the primary call |
| 469 | // through cloning. I.e. located in the same function and have the same |
| 470 | // (possibly trimmed) stack ids. They will be updated the same way as the |
| 471 | // primary call when assigning to function clones. |
| 472 | SmallVector<CallInfo, 0> MatchingCalls; |
| 473 | |
| 474 | // For alloc nodes this is a unique id assigned when constructed, and for |
| 475 | // callsite stack nodes it is the original stack id when the node is |
| 476 | // constructed from the memprof MIB metadata on the alloc nodes. Note that |
| 477 | // this is only used when matching callsite metadata onto the stack nodes |
| 478 | // created when processing the allocation memprof MIBs, and for labeling |
| 479 | // nodes in the dot graph. Therefore we don't bother to assign a value for |
| 480 | // clones. |
| 481 | uint64_t OrigStackOrAllocId = 0; |
| 482 | |
| 483 | // Edges to all callees in the profiled call stacks. |
| 484 | // TODO: Should this be a map (from Callee node) for more efficient lookup? |
| 485 | std::vector<std::shared_ptr<ContextEdge>> CalleeEdges; |
| 486 | |
| 487 | // Edges to all callers in the profiled call stacks. |
| 488 | // TODO: Should this be a map (from Caller node) for more efficient lookup? |
| 489 | std::vector<std::shared_ptr<ContextEdge>> CallerEdges; |
| 490 | |
| 491 | // Returns true if we need to look at the callee edges for determining the |
| 492 | // node context ids and allocation type. |
| 493 | bool useCallerEdgesForContextInfo() const { |
| 494 | // Typically if the callee edges are empty either the caller edges are |
| 495 | // also empty, or this is an allocation (leaf node). However, if we are |
| 496 | // allowing recursive callsites and contexts this will be violated for |
| 497 | // incompletely cloned recursive cycles. |
| 498 | assert(!CalleeEdges.empty() || CallerEdges.empty() || IsAllocation || |
| 499 | (AllowRecursiveCallsites && AllowRecursiveContexts)); |
| 500 | // When cloning for a recursive context, during cloning we might be in the |
| 501 | // midst of cloning for a recurrence and have moved context ids off of a |
| 502 | // caller edge onto the clone but not yet off of the incoming caller |
| 503 | // (back) edge. If we don't look at those we miss the fact that this node |
| 504 | // still has context ids of interest. |
| 505 | return IsAllocation || CloneRecursiveContexts; |
| 506 | } |
| 507 | |
| 508 | // Compute the context ids for this node from the union of its edge context |
| 509 | // ids. |
| 510 | DenseSet<uint32_t> getContextIds() const { |
| 511 | unsigned Count = 0; |
| 512 | // Compute the number of ids for reserve below. In general we only need to |
| 513 | // look at one set of edges, typically the callee edges, since other than |
| 514 | // allocations and in some cases during recursion cloning, all the context |
| 515 | // ids on the callers should also flow out via callee edges. |
| 516 | for (auto &Edge : CalleeEdges.empty() ? CallerEdges : CalleeEdges) |
| 517 | Count += Edge->getContextIds().size(); |
| 518 | DenseSet<uint32_t> ContextIds; |
| 519 | ContextIds.reserve(Size: Count); |
| 520 | auto Edges = llvm::concat<const std::shared_ptr<ContextEdge>>( |
| 521 | CalleeEdges, useCallerEdgesForContextInfo() |
| 522 | ? CallerEdges |
| 523 | : std::vector<std::shared_ptr<ContextEdge>>()); |
| 524 | for (const auto &Edge : Edges) |
| 525 | ContextIds.insert_range(Edge->getContextIds()); |
| 526 | return ContextIds; |
| 527 | } |
| 528 | |
| 529 | // Compute the allocation type for this node from the OR of its edge |
| 530 | // allocation types. |
| 531 | uint8_t computeAllocType() const { |
| 532 | uint8_t BothTypes = |
| 533 | (uint8_t)AllocationType::Cold | (uint8_t)AllocationType::NotCold; |
| 534 | uint8_t AllocType = (uint8_t)AllocationType::None; |
| 535 | auto Edges = llvm::concat<const std::shared_ptr<ContextEdge>>( |
| 536 | CalleeEdges, useCallerEdgesForContextInfo() |
| 537 | ? CallerEdges |
| 538 | : std::vector<std::shared_ptr<ContextEdge>>()); |
| 539 | for (const auto &Edge : Edges) { |
| 540 | AllocType |= Edge->AllocTypes; |
| 541 | // Bail early if alloc type reached both, no further refinement. |
| 542 | if (AllocType == BothTypes) |
| 543 | return AllocType; |
| 544 | } |
| 545 | return AllocType; |
| 546 | } |
| 547 | |
| 548 | // The context ids set for this node is empty if its edge context ids are |
| 549 | // also all empty. |
| 550 | bool emptyContextIds() const { |
| 551 | auto Edges = llvm::concat<const std::shared_ptr<ContextEdge>>( |
| 552 | CalleeEdges, useCallerEdgesForContextInfo() |
| 553 | ? CallerEdges |
| 554 | : std::vector<std::shared_ptr<ContextEdge>>()); |
| 555 | for (const auto &Edge : Edges) { |
| 556 | if (!Edge->getContextIds().empty()) |
| 557 | return false; |
| 558 | } |
| 559 | return true; |
| 560 | } |
| 561 | |
| 562 | // Whether the given context id is among this node's context ids. |
| 563 | bool hasContextId(uint32_t Id) const { |
| 564 | auto Edges = llvm::concat<const std::shared_ptr<ContextEdge>>( |
| 565 | CalleeEdges, useCallerEdgesForContextInfo() |
| 566 | ? CallerEdges |
| 567 | : std::vector<std::shared_ptr<ContextEdge>>()); |
| 568 | for (const auto &Edge : Edges) { |
| 569 | if (Edge->getContextIds().contains(Id)) |
| 570 | return true; |
| 571 | } |
| 572 | return false; |
| 573 | } |
| 574 | |
| 575 | // Removes from Ids any that are not on this node (i.e. on none of the |
| 576 | // edges its context ids are computed from). Walks the edges once for all |
| 577 | // ids, stopping early once every id has been found. |
| 578 | void removeContextIdsNotPresent(DenseSet<uint32_t> &Ids) const { |
| 579 | DenseSet<uint32_t> NotFound = Ids; |
| 580 | auto Edges = llvm::concat<const std::shared_ptr<ContextEdge>>( |
| 581 | CalleeEdges, useCallerEdgesForContextInfo() |
| 582 | ? CallerEdges |
| 583 | : std::vector<std::shared_ptr<ContextEdge>>()); |
| 584 | for (const auto &Edge : Edges) { |
| 585 | const auto &EdgeIds = Edge->getContextIds(); |
| 586 | // Iterate over the smaller of the two sets. |
| 587 | if (EdgeIds.size() < NotFound.size()) { |
| 588 | for (auto Id : EdgeIds) |
| 589 | NotFound.erase(Id); |
| 590 | } else { |
| 591 | NotFound.remove_if( |
| 592 | [&EdgeIds](uint32_t Id) { return EdgeIds.contains(Id); }); |
| 593 | } |
| 594 | if (NotFound.empty()) |
| 595 | return; |
| 596 | } |
| 597 | set_subtract(S1&: Ids, S2: NotFound); |
| 598 | } |
| 599 | |
| 600 | // List of clones of this ContextNode, initially empty. |
| 601 | std::vector<ContextNode *> Clones; |
| 602 | |
| 603 | // If a clone, points to the original uncloned node. |
| 604 | ContextNode *CloneOf = nullptr; |
| 605 | |
| 606 | ContextNode(bool IsAllocation) : IsAllocation(IsAllocation), Call() {} |
| 607 | |
| 608 | ContextNode(bool IsAllocation, CallInfo C) |
| 609 | : IsAllocation(IsAllocation), Call(C) {} |
| 610 | |
| 611 | void addClone(ContextNode *Clone) { |
| 612 | if (CloneOf) { |
| 613 | CloneOf->Clones.push_back(Clone); |
| 614 | Clone->CloneOf = CloneOf; |
| 615 | } else { |
| 616 | Clones.push_back(Clone); |
| 617 | assert(!Clone->CloneOf); |
| 618 | Clone->CloneOf = this; |
| 619 | } |
| 620 | } |
| 621 | |
| 622 | ContextNode *getOrigNode() { |
| 623 | if (!CloneOf) |
| 624 | return this; |
| 625 | return CloneOf; |
| 626 | } |
| 627 | |
| 628 | const ContextNode *getOrigNode() const { |
| 629 | if (!CloneOf) |
| 630 | return this; |
| 631 | return CloneOf; |
| 632 | } |
| 633 | |
| 634 | void addOrUpdateCallerEdge(ContextNode *Caller, AllocationType AllocType, |
| 635 | unsigned int ContextId); |
| 636 | |
| 637 | ContextEdge *findEdgeFromCallee(const ContextNode *Callee); |
| 638 | ContextEdge *findEdgeFromCaller(const ContextNode *Caller); |
| 639 | void eraseCalleeEdge(const ContextEdge *Edge); |
| 640 | void eraseCallerEdge(const ContextEdge *Edge); |
| 641 | |
| 642 | void setCall(CallInfo C) { Call = std::move(C); } |
| 643 | |
| 644 | bool hasCall() const { return (bool)Call.call(); } |
| 645 | |
| 646 | void printCall(raw_ostream &OS) const { Call.print(OS); } |
| 647 | |
| 648 | // True if this node was effectively removed from the graph, in which case |
| 649 | // it should have an allocation type of None and empty context ids. |
| 650 | bool isRemoved() const { |
| 651 | // Typically if the callee edges are empty either the caller edges are |
| 652 | // also empty, or this is an allocation (leaf node). However, if we are |
| 653 | // allowing recursive callsites and contexts this will be violated for |
| 654 | // incompletely cloned recursive cycles. |
| 655 | assert((AllowRecursiveCallsites && AllowRecursiveContexts) || |
| 656 | (AllocTypes == (uint8_t)AllocationType::None) == |
| 657 | emptyContextIds()); |
| 658 | return AllocTypes == (uint8_t)AllocationType::None; |
| 659 | } |
| 660 | |
| 661 | void dump() const; |
| 662 | void print(raw_ostream &OS) const; |
| 663 | |
| 664 | friend raw_ostream &operator<<(raw_ostream &OS, const ContextNode &Node) { |
| 665 | Node.print(OS); |
| 666 | return OS; |
| 667 | } |
| 668 | }; |
| 669 | |
| 670 | /// Edge in the Callsite Context Graph from a ContextNode N to a caller or |
| 671 | /// callee. |
| 672 | struct ContextEdge { |
| 673 | ContextNode *Callee; |
| 674 | ContextNode *Caller; |
| 675 | |
| 676 | // This will be formed by ORing together the AllocationType enum values |
| 677 | // for contexts including this edge. |
| 678 | uint8_t AllocTypes = 0; |
| 679 | |
| 680 | // Set just before initiating cloning when cloning of recursive contexts is |
| 681 | // enabled. Used to defer cloning of backedges until we have done cloning of |
| 682 | // the callee node for non-backedge caller edges. This exposes cloning |
| 683 | // opportunities through the backedge of the cycle. |
| 684 | // TODO: Note that this is not updated during cloning, and it is unclear |
| 685 | // whether that would be needed. |
| 686 | bool IsBackedge = false; |
| 687 | |
| 688 | // The set of IDs for contexts including this edge. |
| 689 | DenseSet<uint32_t> ContextIds; |
| 690 | |
| 691 | ContextEdge(ContextNode *Callee, ContextNode *Caller, uint8_t AllocType, |
| 692 | DenseSet<uint32_t> ContextIds) |
| 693 | : Callee(Callee), Caller(Caller), AllocTypes(AllocType), |
| 694 | ContextIds(std::move(ContextIds)) {} |
| 695 | |
| 696 | DenseSet<uint32_t> &getContextIds() { return ContextIds; } |
| 697 | |
| 698 | // Helper to clear the fields of this edge when we are removing it from the |
| 699 | // graph. |
| 700 | inline void clear() { |
| 701 | ContextIds.clear(); |
| 702 | AllocTypes = (uint8_t)AllocationType::None; |
| 703 | Caller = nullptr; |
| 704 | Callee = nullptr; |
| 705 | } |
| 706 | |
| 707 | // Check if edge was removed from the graph. This is useful while iterating |
| 708 | // over a copy of edge lists when performing operations that mutate the |
| 709 | // graph in ways that might remove one of the edges. |
| 710 | inline bool isRemoved() const { |
| 711 | if (Callee || Caller) |
| 712 | return false; |
| 713 | // Any edges that have been removed from the graph but are still in a |
| 714 | // shared_ptr somewhere should have all fields null'ed out by clear() |
| 715 | // above. |
| 716 | assert(AllocTypes == (uint8_t)AllocationType::None); |
| 717 | assert(ContextIds.empty()); |
| 718 | return true; |
| 719 | } |
| 720 | |
| 721 | void dump() const; |
| 722 | void print(raw_ostream &OS) const; |
| 723 | |
| 724 | friend raw_ostream &operator<<(raw_ostream &OS, const ContextEdge &Edge) { |
| 725 | Edge.print(OS); |
| 726 | return OS; |
| 727 | } |
| 728 | }; |
| 729 | |
| 730 | /// Helpers to remove edges that have allocation type None (due to not |
| 731 | /// carrying any context ids) after transformations. |
| 732 | void removeNoneTypeCalleeEdges(ContextNode *Node); |
| 733 | void removeNoneTypeCallerEdges(ContextNode *Node); |
| 734 | void |
| 735 | recursivelyRemoveNoneTypeCalleeEdges(ContextNode *Node, |
| 736 | DenseSet<const ContextNode *> &Visited); |
| 737 | |
| 738 | protected: |
| 739 | /// Get the list of context nodes for the stack ids in the given callsite |
| 740 | /// context, stopping at the first stack id that does not have a node. |
| 741 | template <class NodeT, class IteratorT> |
| 742 | std::vector<ContextNode *> |
| 743 | getStackNodesForCallsite(CallStack<NodeT, IteratorT> &CallsiteContext); |
| 744 | |
| 745 | /// Adds nodes for the given allocation and any stack ids on its memprof MIB |
| 746 | /// metadata (or summary). |
| 747 | ContextNode *addAllocNode(CallInfo Call, const FuncTy *F); |
| 748 | |
| 749 | /// Adds nodes for the given MIB stack ids. |
| 750 | template <class NodeT, class IteratorT> |
| 751 | void addStackNodesForMIB( |
| 752 | ContextNode *AllocNode, CallStack<NodeT, IteratorT> &StackContext, |
| 753 | CallStack<NodeT, IteratorT> &CallsiteContext, AllocationType AllocType, |
| 754 | ArrayRef<ContextTotalSize> ContextSizeInfo, |
| 755 | std::map<uint64_t, uint32_t> &TotalSizeToContextIdTopNCold); |
| 756 | |
| 757 | /// Matches all callsite metadata (or summary) to the nodes created for |
| 758 | /// allocation memprof MIB metadata, synthesizing new nodes to reflect any |
| 759 | /// inlining performed on those callsite instructions. |
| 760 | void updateStackNodes(); |
| 761 | |
| 762 | /// Optionally fixup edges for the N largest cold contexts to better enable |
| 763 | /// cloning. This is particularly helpful if the context includes recursion |
| 764 | /// as well as inlining, resulting in a single stack node for multiple stack |
| 765 | /// ids in the context. With recursion it is particularly difficult to get the |
| 766 | /// edge updates correct as in the general case we have lost the original |
| 767 | /// stack id ordering for the context. Do more expensive fixup for the largest |
| 768 | /// contexts, controlled by MemProfTopNImportant and MemProfFixupImportant. |
| 769 | void fixupImportantContexts(); |
| 770 | |
| 771 | /// Update graph to conservatively handle any callsite stack nodes that target |
| 772 | /// multiple different callee target functions. |
| 773 | void handleCallsitesWithMultipleTargets(); |
| 774 | |
| 775 | /// Mark backedges via the standard DFS based backedge algorithm. |
| 776 | void markBackedges(); |
| 777 | |
| 778 | /// Merge clones generated during cloning for different allocations but that |
| 779 | /// are called by the same caller node, to ensure proper function assignment. |
| 780 | void mergeClones(); |
| 781 | |
| 782 | // Try to partition calls on the given node (already placed into the AllCalls |
| 783 | // array) by callee function, creating new copies of Node as needed to hold |
| 784 | // calls with different callees, and moving the callee edges appropriately. |
| 785 | // Returns true if partitioning was successful. |
| 786 | bool partitionCallsByCallee( |
| 787 | ContextNode *Node, ArrayRef<CallInfo> AllCalls, |
| 788 | std::vector<std::pair<CallInfo, ContextNode *>> &NewCallToNode); |
| 789 | |
| 790 | /// Save lists of calls with MemProf metadata in each function, for faster |
| 791 | /// iteration. |
| 792 | MapVector<FuncTy *, std::vector<CallInfo>> FuncToCallsWithMetadata; |
| 793 | |
| 794 | /// Map from callsite node to the enclosing caller function. |
| 795 | std::map<const ContextNode *, const FuncTy *> NodeToCallingFunc; |
| 796 | |
| 797 | const FuncTy *getCallingFunc(const ContextNode *N) const { |
| 798 | auto It = NodeToCallingFunc.find(N); |
| 799 | return It != NodeToCallingFunc.end() ? It->second : nullptr; |
| 800 | } |
| 801 | |
| 802 | /// Functions that are referenced (called directly, or address-taken) in the |
| 803 | /// module or summary index. Only populated, at the start of reporting, when |
| 804 | /// MemProfReportHintedSizesDetail is enabled; used to distinguish allocation |
| 805 | /// functions with no references at all from those whose callers simply did |
| 806 | /// not match the profile. |
| 807 | DenseSet<const FuncTy *> FunctionsWithCallers; |
| 808 | |
| 809 | // When exporting to dot, and an allocation id is specified, contains the |
| 810 | // context ids on that allocation. |
| 811 | DenseSet<uint32_t> DotAllocContextIds; |
| 812 | |
| 813 | private: |
| 814 | using EdgeIter = typename std::vector<std::shared_ptr<ContextEdge>>::iterator; |
| 815 | |
| 816 | // Structure to keep track of information for each call as we are matching |
| 817 | // non-allocation callsites onto context nodes created from the allocation |
| 818 | // call metadata / summary contexts. |
| 819 | struct CallContextInfo { |
| 820 | // The callsite we're trying to match. |
| 821 | CallTy Call; |
| 822 | // The context nodes for the callsite's stack ids that have one, saved to |
| 823 | // avoid repeatedly looking them up. The stack ids are available via the |
| 824 | // nodes' OrigStackOrAllocId. Since each stack id has its own node, and the |
| 825 | // stack id to node map is not modified while matching calls, two calls |
| 826 | // have the same stack ids iff they have the same nodes. |
| 827 | std::vector<ContextNode *> StackNodes; |
| 828 | // The function containing this callsite. |
| 829 | const FuncTy *Func; |
| 830 | // Initially empty, if needed this will be updated to contain the context |
| 831 | // ids for use in a new context node created for this callsite. |
| 832 | DenseSet<uint32_t> ContextIds; |
| 833 | // Duplicates, created for this callsite, of the ids of contexts trimmed at |
| 834 | // a stack id before the last one in StackNodes (see |
| 835 | // -memprof-assign-trimmed-inlined-contexts). Any new node created for this |
| 836 | // callsite gets these in addition to ContextIds. |
| 837 | DenseSet<uint32_t> TrimmedInlinedDupIds; |
| 838 | }; |
| 839 | |
| 840 | /// Helper to remove edge from graph, updating edge iterator if it is provided |
| 841 | /// (in which case CalleeIter indicates which edge list is being iterated). |
| 842 | /// This will also perform the necessary clearing of the ContextEdge members |
| 843 | /// to enable later checking if the edge has been removed (since we may have |
| 844 | /// other copies of the shared_ptr in existence, and in fact rely on this to |
| 845 | /// enable removal while iterating over a copy of a node's edge list). |
| 846 | void removeEdgeFromGraph(ContextEdge *Edge, EdgeIter *EI = nullptr, |
| 847 | bool CalleeIter = true); |
| 848 | |
| 849 | /// Assigns the given Node to calls at or inlined into the location with |
| 850 | /// the Node's stack id, after post order traversing and processing its |
| 851 | /// caller nodes. Uses the call information recorded in the given |
| 852 | /// StackIdToMatchingCalls map, and creates new nodes for inlined sequences |
| 853 | /// as needed. Called by updateStackNodes which sets up the given |
| 854 | /// StackIdToMatchingCalls map. |
| 855 | void assignStackNodesPostOrder( |
| 856 | ContextNode *Node, DenseSet<const ContextNode *> &Visited, |
| 857 | DenseMap<uint64_t, std::vector<CallContextInfo>> &StackIdToMatchingCalls, |
| 858 | DenseMap<CallInfo, CallInfo> &CallToMatchingCall, |
| 859 | const DenseSet<uint32_t> &ImportantContextIds); |
| 860 | |
| 861 | /// Duplicates the given set of context ids, updating the provided |
| 862 | /// map from each original id with the newly generated context ids, |
| 863 | /// and returning the new duplicated id set. If provided, OnDuplicate is |
| 864 | /// invoked with each original id and its new duplicate. |
| 865 | DenseSet<uint32_t> duplicateContextIds( |
| 866 | const DenseSet<uint32_t> &StackSequenceContextIds, |
| 867 | DenseMap<uint32_t, DenseSet<uint32_t>> &OldToNewContextIds, |
| 868 | function_ref<void(uint32_t OldId, uint32_t NewId)> OnDuplicate = nullptr); |
| 869 | |
| 870 | /// Propagates all duplicated context ids across the graph. |
| 871 | void propagateDuplicateContextIds( |
| 872 | const DenseMap<uint32_t, DenseSet<uint32_t>> &OldToNewContextIds); |
| 873 | |
| 874 | /// Cache of the ids of contexts trimmed at each node (i.e. that are on the |
| 875 | /// node but not on any of its caller edges). The sets are stored in a vector |
| 876 | /// and looked up by index, so that references to them are not invalidated |
| 877 | /// by adding more entries. |
| 878 | struct TrimmedAtNodeCache { |
| 879 | DenseMap<const ContextNode *, unsigned> NodeToIndex; |
| 880 | std::vector<DenseSet<uint32_t>> TrimmedAtNodeIds; |
| 881 | }; |
| 882 | |
| 883 | /// Returns the ids of contexts that were trimmed at a stack node before the |
| 884 | /// last one in the given call's inlined stack node sequence (see the |
| 885 | /// definition for details). |
| 886 | DenseSet<uint32_t> |
| 887 | computeTrimmedInlinedContextIds(ArrayRef<ContextNode *> StackNodes, |
| 888 | TrimmedAtNodeCache &Cache); |
| 889 | |
| 890 | /// Connect the NewNode to OrigNode's callees if TowardsCallee is true, |
| 891 | /// else to its callers. Also updates OrigNode's edges to remove any context |
| 892 | /// ids moved to the newly created edge. |
| 893 | void connectNewNode(ContextNode *NewNode, ContextNode *OrigNode, |
| 894 | bool TowardsCallee, const DenseSet<uint32_t> &ContextIds); |
| 895 | |
| 896 | /// Get the stack id corresponding to the given Id or Index (for IR this will |
| 897 | /// return itself, for a summary index this will return the id recorded in the |
| 898 | /// index for that stack id index value). |
| 899 | uint64_t getStackId(uint64_t IdOrIndex) const { |
| 900 | return static_cast<const DerivedCCG *>(this)->getStackId(IdOrIndex); |
| 901 | } |
| 902 | |
| 903 | /// Returns true if the given call targets the callee of the given edge, or if |
| 904 | /// we were able to identify the call chain through intermediate tail calls. |
| 905 | /// In the latter case new context nodes are added to the graph for the |
| 906 | /// identified tail calls, and their synthesized nodes are added to |
| 907 | /// TailCallToContextNodeMap. The EdgeIter is updated in the latter case for |
| 908 | /// the updated edges and to prepare it for an increment in the caller. |
| 909 | bool |
| 910 | calleesMatch(CallTy Call, EdgeIter &EI, |
| 911 | MapVector<CallInfo, ContextNode *> &TailCallToContextNodeMap); |
| 912 | |
| 913 | // Return the callee function of the given call, or nullptr if it can't be |
| 914 | // determined |
| 915 | const FuncTy *getCalleeFunc(CallTy Call) { |
| 916 | return static_cast<DerivedCCG *>(this)->getCalleeFunc(Call); |
| 917 | } |
| 918 | |
| 919 | /// Returns true if the given call targets the given function, or if we were |
| 920 | /// able to identify the call chain through intermediate tail calls (in which |
| 921 | /// case FoundCalleeChain will be populated). |
| 922 | bool calleeMatchesFunc( |
| 923 | CallTy Call, const FuncTy *Func, const FuncTy *CallerFunc, |
| 924 | std::vector<std::pair<CallTy, FuncTy *>> &FoundCalleeChain) { |
| 925 | return static_cast<DerivedCCG *>(this)->calleeMatchesFunc( |
| 926 | Call, Func, CallerFunc, FoundCalleeChain); |
| 927 | } |
| 928 | |
| 929 | /// Returns true if both call instructions have the same callee. |
| 930 | bool sameCallee(CallTy Call1, CallTy Call2) { |
| 931 | return static_cast<DerivedCCG *>(this)->sameCallee(Call1, Call2); |
| 932 | } |
| 933 | |
| 934 | /// Get the list of context nodes for the stack ids in the given callsite's |
| 935 | /// context, stopping at the first stack id that does not have a node. |
| 936 | std::vector<ContextNode *> getStackNodesForCall(CallTy Call) { |
| 937 | return static_cast<DerivedCCG *>(this)->getStackNodesForCall(Call); |
| 938 | } |
| 939 | |
| 940 | /// Get the last stack id in the context for callsite. |
| 941 | uint64_t getLastStackId(CallTy Call) { |
| 942 | return static_cast<DerivedCCG *>(this)->getLastStackId(Call); |
| 943 | } |
| 944 | |
| 945 | /// Update the allocation call to record type of allocated memory. |
| 946 | void updateAllocationCall(CallInfo &Call, AllocationType AllocType) { |
| 947 | if (AllocType == AllocationType::Cold) |
| 948 | AllocTypeCold++; |
| 949 | else if (isAmbiguousAllocType(AllocTypes: (uint8_t)AllocType)) |
| 950 | AllocTypeAmbiguous++; |
| 951 | else |
| 952 | AllocTypeNotCold++; |
| 953 | static_cast<DerivedCCG *>(this)->updateAllocationCall(Call, AllocType); |
| 954 | } |
| 955 | |
| 956 | /// Get the AllocationType assigned to the given allocation instruction clone. |
| 957 | AllocationType getAllocationCallType(const CallInfo &Call) const { |
| 958 | return static_cast<const DerivedCCG *>(this)->getAllocationCallType(Call); |
| 959 | } |
| 960 | |
| 961 | /// Update non-allocation call to invoke (possibly cloned) function |
| 962 | /// CalleeFunc. |
| 963 | void updateCall(CallInfo &CallerCall, FuncInfo CalleeFunc) { |
| 964 | static_cast<DerivedCCG *>(this)->updateCall(CallerCall, CalleeFunc); |
| 965 | } |
| 966 | |
| 967 | /// Clone the given function for the given callsite, recording mapping of all |
| 968 | /// of the functions tracked calls to their new versions in the CallMap. |
| 969 | /// Assigns new clones to clone number CloneNo. |
| 970 | FuncInfo cloneFunctionForCallsite( |
| 971 | FuncInfo &Func, CallInfo &Call, DenseMap<CallInfo, CallInfo> &CallMap, |
| 972 | std::vector<CallInfo> &CallsWithMetadataInFunc, unsigned CloneNo) { |
| 973 | return static_cast<DerivedCCG *>(this)->cloneFunctionForCallsite( |
| 974 | Func, Call, CallMap, CallsWithMetadataInFunc, CloneNo); |
| 975 | } |
| 976 | |
| 977 | /// Gets a label to use in the dot graph for the given call clone in the given |
| 978 | /// function. |
| 979 | std::string getLabel(const FuncTy *Func, const CallTy Call, |
| 980 | unsigned CloneNo) const { |
| 981 | return static_cast<const DerivedCCG *>(this)->getLabel(Func, Call, CloneNo); |
| 982 | } |
| 983 | |
| 984 | /// Gets the function name (with optional clone suffix) for reporting. |
| 985 | std::string getFuncName(const FuncTy *Func, unsigned CloneNo = 0) const { |
| 986 | return static_cast<const DerivedCCG *>(this)->getFuncName(Func, CloneNo); |
| 987 | } |
| 988 | |
| 989 | /// Returns true if the given function was marked dead by ThinLTO dead symbol |
| 990 | /// analysis. |
| 991 | bool isFuncDead(const FuncTy *Func) const { |
| 992 | return static_cast<const DerivedCCG *>(this)->isFuncDead(Func); |
| 993 | } |
| 994 | |
| 995 | /// Populates FunctionsWithCallers (MemProfReportHintedSizesDetail only). |
| 996 | void populateFunctionsWithCallers() { |
| 997 | static_cast<DerivedCCG *>(this)->populateFunctionsWithCallers(); |
| 998 | } |
| 999 | |
| 1000 | /// Records, before identifyClones(), the chain of (uncloned) nodes of every |
| 1001 | /// context and how far it matched (MemProfReportHintedSizesDetail only). |
| 1002 | void capturePreCloningContextInfo(); |
| 1003 | |
| 1004 | /// Formats the detail suffix for the report line of context Id, whose |
| 1005 | /// (possibly cloned) allocation node is Node (MemProfReportHintedSizesDetail |
| 1006 | /// only). See the option description for the format. |
| 1007 | std::string formatContextReportSuffix(const ContextNode *Node, |
| 1008 | uint32_t Id) const; |
| 1009 | |
| 1010 | // Create and return a new ContextNode. |
| 1011 | ContextNode *createNewNode(bool IsAllocation, const FuncTy *F = nullptr, |
| 1012 | CallInfo C = CallInfo()) { |
| 1013 | NodeOwner.push_back(std::make_unique<ContextNode>(IsAllocation, C)); |
| 1014 | auto *NewNode = NodeOwner.back().get(); |
| 1015 | if (F) |
| 1016 | NodeToCallingFunc[NewNode] = F; |
| 1017 | NewNode->NodeId = NodeOwner.size(); |
| 1018 | return NewNode; |
| 1019 | } |
| 1020 | |
| 1021 | /// Helpers to find the node corresponding to the given call or stackid. |
| 1022 | ContextNode *getNodeForInst(const CallInfo &C); |
| 1023 | ContextNode *getNodeForAlloc(const CallInfo &C); |
| 1024 | ContextNode *getNodeForStackId(uint64_t StackId); |
| 1025 | |
| 1026 | /// Computes the alloc type corresponding to the given context ids, by |
| 1027 | /// unioning their recorded alloc types. |
| 1028 | uint8_t computeAllocType(DenseSet<uint32_t> &ContextIds) const; |
| 1029 | |
| 1030 | /// Returns the allocation type of the intersection of the contexts of two |
| 1031 | /// nodes (based on their provided context id sets), optimized for the case |
| 1032 | /// when Node1Ids is smaller than Node2Ids. |
| 1033 | uint8_t intersectAllocTypesImpl(const DenseSet<uint32_t> &Node1Ids, |
| 1034 | const DenseSet<uint32_t> &Node2Ids) const; |
| 1035 | |
| 1036 | /// Returns the allocation type of the intersection of the contexts of two |
| 1037 | /// nodes (based on their provided context id sets). |
| 1038 | uint8_t intersectAllocTypes(const DenseSet<uint32_t> &Node1Ids, |
| 1039 | const DenseSet<uint32_t> &Node2Ids) const; |
| 1040 | |
| 1041 | /// Create a clone of Edge's callee and move Edge to that new callee node, |
| 1042 | /// performing the necessary context id and allocation type updates. |
| 1043 | /// If ContextIdsToMove is non-empty, only that subset of Edge's ids are |
| 1044 | /// moved to an edge to the new callee. |
| 1045 | ContextNode * |
| 1046 | moveEdgeToNewCalleeClone(const std::shared_ptr<ContextEdge> &Edge, |
| 1047 | DenseSet<uint32_t> ContextIdsToMove = {}); |
| 1048 | |
| 1049 | /// Change the callee of Edge to existing callee clone NewCallee, performing |
| 1050 | /// the necessary context id and allocation type updates. |
| 1051 | /// If ContextIdsToMove is non-empty, only that subset of Edge's ids are |
| 1052 | /// moved to an edge to the new callee. |
| 1053 | void moveEdgeToExistingCalleeClone(const std::shared_ptr<ContextEdge> &Edge, |
| 1054 | ContextNode *NewCallee, |
| 1055 | bool NewClone = false, |
| 1056 | DenseSet<uint32_t> ContextIdsToMove = {}); |
| 1057 | |
| 1058 | /// Change the caller of the edge at the given callee edge iterator to be |
| 1059 | /// NewCaller, performing the necessary context id and allocation type |
| 1060 | /// updates. This is similar to the above moveEdgeToExistingCalleeClone, but |
| 1061 | /// a simplified version of it as we always move the given edge and all of its |
| 1062 | /// context ids. |
| 1063 | void moveCalleeEdgeToNewCaller(const std::shared_ptr<ContextEdge> &Edge, |
| 1064 | ContextNode *NewCaller); |
| 1065 | |
| 1066 | /// Recursive helper for marking backedges via DFS. |
| 1067 | void markBackedges(ContextNode *Node, DenseSet<const ContextNode *> &Visited, |
| 1068 | DenseSet<const ContextNode *> &CurrentStack); |
| 1069 | |
| 1070 | /// Recursive helper for merging clones. |
| 1071 | void |
| 1072 | mergeClones(ContextNode *Node, DenseSet<const ContextNode *> &Visited, |
| 1073 | DenseMap<uint32_t, ContextNode *> &ContextIdToAllocationNode); |
| 1074 | /// Main worker for merging callee clones for a given node. |
| 1075 | void mergeNodeCalleeClones( |
| 1076 | ContextNode *Node, DenseSet<const ContextNode *> &Visited, |
| 1077 | DenseMap<uint32_t, ContextNode *> &ContextIdToAllocationNode); |
| 1078 | /// Helper to find other callers of the given set of callee edges that can |
| 1079 | /// share the same callee merge node. |
| 1080 | void findOtherCallersToShareMerge( |
| 1081 | ContextNode *Node, std::vector<std::shared_ptr<ContextEdge>> &CalleeEdges, |
| 1082 | DenseMap<uint32_t, ContextNode *> &ContextIdToAllocationNode, |
| 1083 | DenseSet<ContextNode *> &OtherCallersToShareMerge); |
| 1084 | |
| 1085 | /// Recursively perform cloning on the graph for the given Node and its |
| 1086 | /// callers, in order to uniquely identify the allocation behavior of an |
| 1087 | /// allocation given its context. The context ids of the allocation being |
| 1088 | /// processed are given in AllocContextIds. |
| 1089 | void identifyClones(ContextNode *Node, DenseSet<const ContextNode *> &Visited, |
| 1090 | const DenseSet<uint32_t> &AllocContextIds); |
| 1091 | |
| 1092 | /// Records which clone of a callsite node an implicit caller requires (see |
| 1093 | /// assignImplicitCallers). The call to that callsite in a given function |
| 1094 | /// clone (the map key below) must call the same callee version as Source. |
| 1095 | struct ImplicitSource { |
| 1096 | /// The required clone of the callsite node. |
| 1097 | const ContextNode *Source; |
| 1098 | /// One of the trimmed context ids that made Source required. If Source is |
| 1099 | /// later emptied by further cloning, this id is used to find the clone |
| 1100 | /// that now carries the trimmed contexts (see resolveImplicitSource). |
| 1101 | uint32_t ContextId; |
| 1102 | }; |
| 1103 | /// Map from (original callsite node, function clone number) to the clone of |
| 1104 | /// that callsite node that the call in that function clone must follow. |
| 1105 | using ImplicitSourceMap = |
| 1106 | DenseMap<std::pair<const ContextNode *, unsigned>, ImplicitSource>; |
| 1107 | |
| 1108 | /// Assign the implicit callers of the clones of Node (callsite clones without |
| 1109 | /// an edge to any clone of Node which contexts trimmed at Node's frame would |
| 1110 | /// have reached) the function clone called by the caller carrying those |
| 1111 | /// contexts, and record the clone of Node their calls must follow. See |
| 1112 | /// assignFunctions for the maps. |
| 1113 | void assignImplicitCallers( |
| 1114 | ContextNode *Node, const DenseSet<uint32_t> &TrimmedContextIds, |
| 1115 | DenseMap<ContextNode *, FuncInfo> &CallsiteToCalleeFuncCloneMap, |
| 1116 | const std::map<FuncInfo, ContextNode *> &FuncCloneToCurNodeCloneMap, |
| 1117 | ImplicitSourceMap &ImplicitSources); |
| 1118 | |
| 1119 | /// Returns the clone of Node (the original callsite node) that the implicit |
| 1120 | /// source Src refers to: Src.Source unless it has since been emptied, in |
| 1121 | /// which case the clone that now carries Src.ContextId (null if none does, |
| 1122 | /// which is not expected). |
| 1123 | static const ContextNode *resolveImplicitSource(const ContextNode *Node, |
| 1124 | const ImplicitSource &Src); |
| 1125 | |
| 1126 | /// Map from each context ID to the AllocationType assigned to that context. |
| 1127 | DenseMap<uint32_t, AllocationType> ContextIdToAllocationType; |
| 1128 | |
| 1129 | /// Map from each contextID to the profiled full contexts and their total |
| 1130 | /// sizes (there may be more than one due to context trimming), |
| 1131 | /// optionally populated when requested (via MemProfReportHintedSizes or |
| 1132 | /// MinClonedColdBytePercent). |
| 1133 | DenseMap<uint32_t, std::vector<ContextTotalSize>> ContextIdToContextSizeInfos; |
| 1134 | |
| 1135 | /// Identifies the context node created for a stack id when adding the MIB |
| 1136 | /// contexts to the graph. This is used to locate the context nodes when |
| 1137 | /// trying to assign the corresponding callsites with those stack ids to these |
| 1138 | /// nodes. |
| 1139 | DenseMap<uint64_t, ContextNode *> StackEntryIdToContextNodeMap; |
| 1140 | |
| 1141 | /// Saves information for the contexts identified as important (the largest |
| 1142 | /// cold contexts up to MemProfTopNImportant). |
| 1143 | struct ImportantContextInfo { |
| 1144 | // The original list of leaf first stack ids corresponding to this context. |
| 1145 | std::vector<uint64_t> StackIds; |
| 1146 | // Max length of stack ids corresponding to a single stack ContextNode for |
| 1147 | // this context (i.e. the max length of a key in StackIdsToNode below). |
| 1148 | unsigned MaxLength = 0; |
| 1149 | // Mapping of slices of the stack ids to the corresponding ContextNode |
| 1150 | // (there can be multiple stack ids due to inlining). Populated when |
| 1151 | // updating stack nodes while matching them to the IR or summary. |
| 1152 | std::map<std::vector<uint64_t>, ContextNode *> StackIdsToNode; |
| 1153 | }; |
| 1154 | |
| 1155 | // Map of important full context ids to information about each. |
| 1156 | DenseMap<uint32_t, ImportantContextInfo> ImportantContextIdInfo; |
| 1157 | |
| 1158 | // For each important context id found in Node (if any), records the list of |
| 1159 | // stack ids (those of the given StackNodes) that corresponded to the given |
| 1160 | // callsite Node. There can be more than one in the case of inlining. |
| 1161 | void recordStackNode(ArrayRef<ContextNode *> StackNodes, ContextNode *Node, |
| 1162 | // We pass in the Node's context ids to avoid the |
| 1163 | // overhead of computing them as the caller already has |
| 1164 | // them in some cases. |
| 1165 | const DenseSet<uint32_t> &NodeContextIds, |
| 1166 | const DenseSet<uint32_t> &ImportantContextIds) { |
| 1167 | if (!MemProfTopNImportant) { |
| 1168 | assert(ImportantContextIds.empty()); |
| 1169 | return; |
| 1170 | } |
| 1171 | DenseSet<uint32_t> Ids = |
| 1172 | set_intersection(S1: NodeContextIds, S2: ImportantContextIds); |
| 1173 | if (Ids.empty()) |
| 1174 | return; |
| 1175 | std::vector<uint64_t> StackIds; |
| 1176 | StackIds.reserve(n: StackNodes.size()); |
| 1177 | for (const ContextNode *N : StackNodes) |
| 1178 | StackIds.push_back(N->OrigStackOrAllocId); |
| 1179 | auto Size = StackIds.size(); |
| 1180 | for (auto Id : Ids) { |
| 1181 | auto &Entry = ImportantContextIdInfo[Id]; |
| 1182 | Entry.StackIdsToNode[StackIds] = Node; |
| 1183 | // Keep track of the max to simplify later analysis. |
| 1184 | if (Size > Entry.MaxLength) |
| 1185 | Entry.MaxLength = Size; |
| 1186 | } |
| 1187 | } |
| 1188 | |
| 1189 | /// Maps to track the calls to their corresponding nodes in the graph. |
| 1190 | MapVector<CallInfo, ContextNode *> AllocationCallToContextNodeMap; |
| 1191 | MapVector<CallInfo, ContextNode *> NonAllocationCallToContextNodeMap; |
| 1192 | |
| 1193 | /// Owner of all ContextNode unique_ptrs. |
| 1194 | std::vector<std::unique_ptr<ContextNode>> NodeOwner; |
| 1195 | |
| 1196 | /// Perform sanity checks on graph when requested. |
| 1197 | void check() const; |
| 1198 | |
| 1199 | /// Keeps track of the last unique context id assigned. |
| 1200 | unsigned int LastContextId = 0; |
| 1201 | |
| 1202 | /// Per-context information captured before identifyClones() when |
| 1203 | /// MemProfReportHintedSizesDetail is enabled, so that the report can relate |
| 1204 | /// the final (cloned) state back to the original context. "Frames" below are |
| 1205 | /// CCG nodes, not profile stack frames: a node matched to a call covers all |
| 1206 | /// of that call's inlined stack frames, so a chain can be shorter than the |
| 1207 | /// profiled stack. |
| 1208 | struct PreCloningContextInfo { |
| 1209 | // Ordered chain of uncloned ContextNodes for this context ID from the leaf |
| 1210 | // allocation node (Chain[0]) up to the outermost node in the CCG for this |
| 1211 | // context. |
| 1212 | std::vector<const ContextNode *> Chain; |
| 1213 | // Index in Chain of the outermost matched callsite node, i.e. the |
| 1214 | // outermost node such that it and all nodes below it |
| 1215 | // (Chain[1..MaxMatchedFrame]) were matched to a call in the IR/summary |
| 1216 | // (hasCall() == true). If the whole context matched, this is the |
| 1217 | // context's outermost node (possibly where the context was trimmed); |
| 1218 | // otherwise it is the last matched node below the first unmatched one. The |
| 1219 | // function containing it (the "outermost matched caller") is where the |
| 1220 | // reported post-cloning paths start. 0 if no caller of the allocation |
| 1221 | // matched. |
| 1222 | unsigned MaxMatchedFrame = 0; |
| 1223 | // True if the outermost matched node had caller edges for other contexts, |
| 1224 | // i.e. the context was trimmed below callers that other contexts reach. |
| 1225 | bool OutermostMatchedNodeHasOtherCallers = false; |
| 1226 | // If MaxMatchedFrame + 1 < Chain.size(), why did Chain[MaxMatchedFrame + 1] |
| 1227 | // not have a call: it was a recursive cycle node, or its calling function |
| 1228 | // was found but the call there did not match the callee (mismatched |
| 1229 | // callee); otherwise the caller had no !callsite for this stack id. |
| 1230 | bool StoppedAtRecursive = false; |
| 1231 | bool StoppedAtMismatchedCallee = false; |
| 1232 | // Stack id and (when known) calling function of the first unmatched node. |
| 1233 | uint64_t UnmatchedStackId = 0; |
| 1234 | const FuncTy *UnmatchedCallingFunc = nullptr; |
| 1235 | }; |
| 1236 | /// Map from context id to its pre-cloning information. |
| 1237 | DenseMap<uint32_t, PreCloningContextInfo> PreCloningContextInfoMap; |
| 1238 | /// Allocation nodes (uncloned) that had at least one matched caller before |
| 1239 | /// cloning. |
| 1240 | DenseSet<const ContextNode *> PreCloningAllocHasMatchedCaller; |
| 1241 | /// Map from function to the number of function clones assigned to it |
| 1242 | /// (including the original), recorded during assignFunctions(). |
| 1243 | DenseMap<const FuncTy *, unsigned> FuncToNumClones; |
| 1244 | /// Map from (original callsite node, clone number of the function clone |
| 1245 | /// containing the call) to the clone number of the callee function the call |
| 1246 | /// was updated to in assignFunctions(). Used to follow the post-cloning call |
| 1247 | /// path from a caller clone down to the allocation clone it reaches. |
| 1248 | DenseMap<std::pair<const ContextNode *, unsigned>, unsigned> |
| 1249 | OrigNodeAndCloneToCalleeClone; |
| 1250 | /// Map from (original allocation node, clone number of the function clone |
| 1251 | /// containing the allocation) to the allocation node in that clone, whose |
| 1252 | /// call carries the final hint. |
| 1253 | DenseMap<std::pair<const ContextNode *, unsigned>, const ContextNode *> |
| 1254 | OrigAllocAndCloneToNode; |
| 1255 | }; |
| 1256 | |
| 1257 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1258 | using ContextNode = |
| 1259 | typename CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextNode; |
| 1260 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1261 | using ContextEdge = |
| 1262 | typename CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextEdge; |
| 1263 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1264 | using FuncInfo = |
| 1265 | typename CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::FuncInfo; |
| 1266 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1267 | using CallInfo = |
| 1268 | typename CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::CallInfo; |
| 1269 | |
| 1270 | /// CRTP derived class for graphs built from IR (regular LTO). |
| 1271 | class ModuleCallsiteContextGraph |
| 1272 | : public CallsiteContextGraph<ModuleCallsiteContextGraph, Function, |
| 1273 | Instruction *> { |
| 1274 | public: |
| 1275 | ModuleCallsiteContextGraph( |
| 1276 | Module &M, |
| 1277 | llvm::function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter); |
| 1278 | |
| 1279 | private: |
| 1280 | friend CallsiteContextGraph<ModuleCallsiteContextGraph, Function, |
| 1281 | Instruction *>; |
| 1282 | |
| 1283 | uint64_t getStackId(uint64_t IdOrIndex) const; |
| 1284 | const Function *getCalleeFunc(Instruction *Call); |
| 1285 | bool calleeMatchesFunc( |
| 1286 | Instruction *Call, const Function *Func, const Function *CallerFunc, |
| 1287 | std::vector<std::pair<Instruction *, Function *>> &FoundCalleeChain); |
| 1288 | bool sameCallee(Instruction *Call1, Instruction *Call2); |
| 1289 | bool findProfiledCalleeThroughTailCalls( |
| 1290 | const Function *ProfiledCallee, Value *CurCallee, unsigned Depth, |
| 1291 | std::vector<std::pair<Instruction *, Function *>> &FoundCalleeChain, |
| 1292 | bool &FoundMultipleCalleeChains); |
| 1293 | uint64_t getLastStackId(Instruction *Call); |
| 1294 | std::vector<ContextNode *> getStackNodesForCall(Instruction *Call); |
| 1295 | void updateAllocationCall(CallInfo &Call, AllocationType AllocType); |
| 1296 | AllocationType getAllocationCallType(const CallInfo &Call) const; |
| 1297 | void updateCall(CallInfo &CallerCall, FuncInfo CalleeFunc); |
| 1298 | CallsiteContextGraph<ModuleCallsiteContextGraph, Function, |
| 1299 | Instruction *>::FuncInfo |
| 1300 | cloneFunctionForCallsite(FuncInfo &Func, CallInfo &Call, |
| 1301 | DenseMap<CallInfo, CallInfo> &CallMap, |
| 1302 | std::vector<CallInfo> &CallsWithMetadataInFunc, |
| 1303 | unsigned CloneNo); |
| 1304 | std::string getLabel(const Function *Func, const Instruction *Call, |
| 1305 | unsigned CloneNo) const; |
| 1306 | std::string getFuncName(const Function *Func, unsigned CloneNo = 0) const; |
| 1307 | bool isFuncDead(const Function *Func) const { return false; } |
| 1308 | void populateFunctionsWithCallers(); |
| 1309 | |
| 1310 | const Module &Mod; |
| 1311 | llvm::function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter; |
| 1312 | }; |
| 1313 | |
| 1314 | /// Represents a call in the summary index graph, which can either be an |
| 1315 | /// allocation or an interior callsite node in an allocation's context. |
| 1316 | /// Holds a pointer to the corresponding data structure in the index. |
| 1317 | struct IndexCall : public PointerUnion<CallsiteInfo *, AllocInfo *> { |
| 1318 | IndexCall() : PointerUnion() {} |
| 1319 | IndexCall(std::nullptr_t) : IndexCall() {} |
| 1320 | IndexCall(CallsiteInfo *StackNode) : PointerUnion(StackNode) {} |
| 1321 | IndexCall(AllocInfo *AllocNode) : PointerUnion(AllocNode) {} |
| 1322 | IndexCall(PointerUnion PT) : PointerUnion(PT) {} |
| 1323 | |
| 1324 | IndexCall *operator->() { return this; } |
| 1325 | |
| 1326 | void print(raw_ostream &OS) const { |
| 1327 | PointerUnion<CallsiteInfo *, AllocInfo *> Base = *this; |
| 1328 | if (auto *AI = llvm::dyn_cast_if_present<AllocInfo *>(Val&: Base)) { |
| 1329 | OS << *AI; |
| 1330 | } else { |
| 1331 | auto *CI = llvm::dyn_cast_if_present<CallsiteInfo *>(Val&: Base); |
| 1332 | assert(CI); |
| 1333 | OS << *CI; |
| 1334 | } |
| 1335 | } |
| 1336 | }; |
| 1337 | } // namespace |
| 1338 | |
| 1339 | namespace llvm { |
| 1340 | template <> struct simplify_type<IndexCall> { |
| 1341 | using SimpleType = PointerUnion<CallsiteInfo *, AllocInfo *>; |
| 1342 | static SimpleType getSimplifiedValue(IndexCall &Val) { return Val; } |
| 1343 | }; |
| 1344 | template <> struct simplify_type<const IndexCall> { |
| 1345 | using SimpleType = const PointerUnion<CallsiteInfo *, AllocInfo *>; |
| 1346 | static SimpleType getSimplifiedValue(const IndexCall &Val) { return Val; } |
| 1347 | }; |
| 1348 | } // namespace llvm |
| 1349 | |
| 1350 | namespace { |
| 1351 | /// CRTP derived class for graphs built from summary index (ThinLTO). |
| 1352 | class IndexCallsiteContextGraph |
| 1353 | : public CallsiteContextGraph<IndexCallsiteContextGraph, FunctionSummary, |
| 1354 | IndexCall> { |
| 1355 | public: |
| 1356 | IndexCallsiteContextGraph( |
| 1357 | ModuleSummaryIndex &Index, |
| 1358 | llvm::function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> |
| 1359 | isPrevailing); |
| 1360 | |
| 1361 | ~IndexCallsiteContextGraph() { |
| 1362 | // Now that we are done with the graph it is safe to add the new |
| 1363 | // CallsiteInfo structs to the function summary vectors. The graph nodes |
| 1364 | // point into locations within these vectors, so we don't want to add them |
| 1365 | // any earlier. |
| 1366 | for (auto &I : FunctionCalleesToSynthesizedCallsiteInfos) { |
| 1367 | auto *FS = I.first; |
| 1368 | for (auto &Callsite : I.second) |
| 1369 | FS->addCallsite(Callsite: std::move(*Callsite.second)); |
| 1370 | } |
| 1371 | } |
| 1372 | |
| 1373 | private: |
| 1374 | friend CallsiteContextGraph<IndexCallsiteContextGraph, FunctionSummary, |
| 1375 | IndexCall>; |
| 1376 | |
| 1377 | uint64_t getStackId(uint64_t IdOrIndex) const; |
| 1378 | const FunctionSummary *getCalleeFunc(IndexCall &Call); |
| 1379 | bool calleeMatchesFunc( |
| 1380 | IndexCall &Call, const FunctionSummary *Func, |
| 1381 | const FunctionSummary *CallerFunc, |
| 1382 | std::vector<std::pair<IndexCall, FunctionSummary *>> &FoundCalleeChain); |
| 1383 | bool sameCallee(IndexCall &Call1, IndexCall &Call2); |
| 1384 | bool findProfiledCalleeThroughTailCalls( |
| 1385 | ValueInfo ProfiledCallee, ValueInfo CurCallee, unsigned Depth, |
| 1386 | std::vector<std::pair<ValueInfo, FunctionSummary *>> &FoundCalleeChain, |
| 1387 | bool &FoundMultipleCalleeChains); |
| 1388 | uint64_t getLastStackId(IndexCall &Call); |
| 1389 | std::vector<ContextNode *> getStackNodesForCall(IndexCall &Call); |
| 1390 | void updateAllocationCall(CallInfo &Call, AllocationType AllocType); |
| 1391 | AllocationType getAllocationCallType(const CallInfo &Call) const; |
| 1392 | void updateCall(CallInfo &CallerCall, FuncInfo CalleeFunc); |
| 1393 | CallsiteContextGraph<IndexCallsiteContextGraph, FunctionSummary, |
| 1394 | IndexCall>::FuncInfo |
| 1395 | cloneFunctionForCallsite(FuncInfo &Func, CallInfo &Call, |
| 1396 | DenseMap<CallInfo, CallInfo> &CallMap, |
| 1397 | std::vector<CallInfo> &CallsWithMetadataInFunc, |
| 1398 | unsigned CloneNo); |
| 1399 | std::string getLabel(const FunctionSummary *Func, const IndexCall &Call, |
| 1400 | unsigned CloneNo) const; |
| 1401 | std::string getFuncName(const FunctionSummary *Func, |
| 1402 | unsigned CloneNo = 0) const; |
| 1403 | bool isFuncDead(const FunctionSummary *Func) const; |
| 1404 | void populateFunctionsWithCallers(); |
| 1405 | DenseSet<GlobalValue::GUID> findAliaseeGUIDsPrevailingInDifferentModule(); |
| 1406 | |
| 1407 | // Saves mapping from function summaries containing memprof records back to |
| 1408 | // its VI, for use in checking and debugging. |
| 1409 | std::map<const FunctionSummary *, ValueInfo> FSToVIMap; |
| 1410 | |
| 1411 | const ModuleSummaryIndex &Index; |
| 1412 | llvm::function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> |
| 1413 | isPrevailing; |
| 1414 | |
| 1415 | // Saves/owns the callsite info structures synthesized for missing tail call |
| 1416 | // frames that we discover while building the graph. |
| 1417 | // It maps from the summary of the function making the tail call, to a map |
| 1418 | // of callee ValueInfo to corresponding synthesized callsite info. |
| 1419 | DenseMap<FunctionSummary *, |
| 1420 | std::map<ValueInfo, std::unique_ptr<CallsiteInfo>>> |
| 1421 | FunctionCalleesToSynthesizedCallsiteInfos; |
| 1422 | }; |
| 1423 | } // namespace |
| 1424 | |
| 1425 | template <> |
| 1426 | struct llvm::DenseMapInfo<CallsiteContextGraph< |
| 1427 | ModuleCallsiteContextGraph, Function, Instruction *>::CallInfo> |
| 1428 | : public DenseMapInfo<std::pair<Instruction *, unsigned>> {}; |
| 1429 | template <> |
| 1430 | struct llvm::DenseMapInfo<CallsiteContextGraph< |
| 1431 | IndexCallsiteContextGraph, FunctionSummary, IndexCall>::CallInfo> |
| 1432 | : public DenseMapInfo<std::pair<IndexCall, unsigned>> {}; |
| 1433 | template <> |
| 1434 | struct llvm::DenseMapInfo<IndexCall> |
| 1435 | : public DenseMapInfo<PointerUnion<CallsiteInfo *, AllocInfo *>> {}; |
| 1436 | |
| 1437 | namespace { |
| 1438 | |
| 1439 | // Map the uint8_t alloc types (which may contain NotCold|Cold) to the alloc |
| 1440 | // type we should actually use on the corresponding allocation. |
| 1441 | // If we can't clone a node that has NotCold+Cold alloc type, we will fall |
| 1442 | // back to using NotCold. So don't bother cloning to distinguish NotCold+Cold |
| 1443 | // from NotCold. |
| 1444 | AllocationType allocTypeToUse(uint8_t AllocTypes) { |
| 1445 | assert(AllocTypes != (uint8_t)AllocationType::None); |
| 1446 | if (AllocTypes == |
| 1447 | ((uint8_t)AllocationType::NotCold | (uint8_t)AllocationType::Cold)) |
| 1448 | return AllocationType::NotCold; |
| 1449 | else |
| 1450 | return (AllocationType)AllocTypes; |
| 1451 | } |
| 1452 | |
| 1453 | // Helper to check if the alloc types for all edges recorded in the |
| 1454 | // InAllocTypes vector match the alloc types for all edges in the Edges |
| 1455 | // vector. |
| 1456 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1457 | bool allocTypesMatch( |
| 1458 | const std::vector<uint8_t> &InAllocTypes, |
| 1459 | const std::vector<std::shared_ptr<ContextEdge<DerivedCCG, FuncTy, CallTy>>> |
| 1460 | &Edges) { |
| 1461 | // This should be called only when the InAllocTypes vector was computed for |
| 1462 | // this set of Edges. Make sure the sizes are the same. |
| 1463 | assert(InAllocTypes.size() == Edges.size()); |
| 1464 | return std::equal( |
| 1465 | InAllocTypes.begin(), InAllocTypes.end(), Edges.begin(), Edges.end(), |
| 1466 | [](const uint8_t &l, |
| 1467 | const std::shared_ptr<ContextEdge<DerivedCCG, FuncTy, CallTy>> &r) { |
| 1468 | // Can share if one of the edges is None type - don't |
| 1469 | // care about the type along that edge as it doesn't |
| 1470 | // exist for those context ids. |
| 1471 | if (l == (uint8_t)AllocationType::None || |
| 1472 | r->AllocTypes == (uint8_t)AllocationType::None) |
| 1473 | return true; |
| 1474 | return allocTypeToUse(AllocTypes: l) == allocTypeToUse(r->AllocTypes); |
| 1475 | }); |
| 1476 | } |
| 1477 | |
| 1478 | // Helper to check if the alloc types for all edges recorded in the |
| 1479 | // InAllocTypes vector match the alloc types for callee edges in the given |
| 1480 | // clone. Because the InAllocTypes were computed from the original node's callee |
| 1481 | // edges, and other cloning could have happened after this clone was created, we |
| 1482 | // need to find the matching clone callee edge, which may or may not exist. |
| 1483 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1484 | bool allocTypesMatchClone( |
| 1485 | const std::vector<uint8_t> &InAllocTypes, |
| 1486 | const ContextNode<DerivedCCG, FuncTy, CallTy> *Clone) { |
| 1487 | const ContextNode<DerivedCCG, FuncTy, CallTy> *Node = Clone->CloneOf; |
| 1488 | assert(Node); |
| 1489 | // InAllocTypes should have been computed for the original node's callee |
| 1490 | // edges. |
| 1491 | assert(InAllocTypes.size() == Node->CalleeEdges.size()); |
| 1492 | // First create a map of the clone callee edge callees to the edge alloc type. |
| 1493 | DenseMap<const ContextNode<DerivedCCG, FuncTy, CallTy> *, uint8_t> |
| 1494 | EdgeCalleeMap; |
| 1495 | for (const auto &E : Clone->CalleeEdges) { |
| 1496 | assert(!EdgeCalleeMap.contains(E->Callee)); |
| 1497 | EdgeCalleeMap[E->Callee] = E->AllocTypes; |
| 1498 | } |
| 1499 | // Next, walk the original node's callees, and look for the corresponding |
| 1500 | // clone edge to that callee. |
| 1501 | for (unsigned I = 0; I < Node->CalleeEdges.size(); I++) { |
| 1502 | auto Iter = EdgeCalleeMap.find(Node->CalleeEdges[I]->Callee); |
| 1503 | // Not found is ok, we will simply add an edge if we use this clone. |
| 1504 | if (Iter == EdgeCalleeMap.end()) |
| 1505 | continue; |
| 1506 | // Can share if one of the edges is None type - don't |
| 1507 | // care about the type along that edge as it doesn't |
| 1508 | // exist for those context ids. |
| 1509 | if (InAllocTypes[I] == (uint8_t)AllocationType::None || |
| 1510 | Iter->second == (uint8_t)AllocationType::None) |
| 1511 | continue; |
| 1512 | if (allocTypeToUse(Iter->second) != allocTypeToUse(AllocTypes: InAllocTypes[I])) |
| 1513 | return false; |
| 1514 | } |
| 1515 | return true; |
| 1516 | } |
| 1517 | |
| 1518 | } // end anonymous namespace |
| 1519 | |
| 1520 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1521 | typename CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextNode * |
| 1522 | CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::getNodeForInst( |
| 1523 | const CallInfo &C) { |
| 1524 | ContextNode *Node = getNodeForAlloc(C); |
| 1525 | if (Node) |
| 1526 | return Node; |
| 1527 | |
| 1528 | return NonAllocationCallToContextNodeMap.lookup(C); |
| 1529 | } |
| 1530 | |
| 1531 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1532 | typename CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextNode * |
| 1533 | CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::getNodeForAlloc( |
| 1534 | const CallInfo &C) { |
| 1535 | return AllocationCallToContextNodeMap.lookup(C); |
| 1536 | } |
| 1537 | |
| 1538 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1539 | typename CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextNode * |
| 1540 | CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::getNodeForStackId( |
| 1541 | uint64_t StackId) { |
| 1542 | auto StackEntryNode = StackEntryIdToContextNodeMap.find(StackId); |
| 1543 | if (StackEntryNode != StackEntryIdToContextNodeMap.end()) |
| 1544 | return StackEntryNode->second; |
| 1545 | return nullptr; |
| 1546 | } |
| 1547 | |
| 1548 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1549 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextNode:: |
| 1550 | addOrUpdateCallerEdge(ContextNode *Caller, AllocationType AllocType, |
| 1551 | unsigned int ContextId) { |
| 1552 | for (auto &Edge : CallerEdges) { |
| 1553 | if (Edge->Caller == Caller) { |
| 1554 | Edge->AllocTypes |= (uint8_t)AllocType; |
| 1555 | Edge->getContextIds().insert(ContextId); |
| 1556 | return; |
| 1557 | } |
| 1558 | } |
| 1559 | std::shared_ptr<ContextEdge> Edge = std::make_shared<ContextEdge>( |
| 1560 | this, Caller, (uint8_t)AllocType, DenseSet<uint32_t>({ContextId})); |
| 1561 | CallerEdges.push_back(Edge); |
| 1562 | Caller->CalleeEdges.push_back(Edge); |
| 1563 | } |
| 1564 | |
| 1565 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1566 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::removeEdgeFromGraph( |
| 1567 | ContextEdge *Edge, EdgeIter *EI, bool CalleeIter) { |
| 1568 | assert(!EI || (*EI)->get() == Edge); |
| 1569 | assert(!Edge->isRemoved()); |
| 1570 | // Save the Caller and Callee pointers so we can erase Edge from their edge |
| 1571 | // lists after clearing Edge below. We do the clearing first in case it is |
| 1572 | // destructed after removing from the edge lists (if those were the last |
| 1573 | // shared_ptr references to Edge). |
| 1574 | auto *Callee = Edge->Callee; |
| 1575 | auto *Caller = Edge->Caller; |
| 1576 | |
| 1577 | // Make sure the edge fields are cleared out so we can properly detect |
| 1578 | // removed edges if Edge is not destructed because there is still a shared_ptr |
| 1579 | // reference. |
| 1580 | Edge->clear(); |
| 1581 | |
| 1582 | #ifndef NDEBUG |
| 1583 | auto CalleeCallerCount = Callee->CallerEdges.size(); |
| 1584 | auto CallerCalleeCount = Caller->CalleeEdges.size(); |
| 1585 | #endif |
| 1586 | if (!EI) { |
| 1587 | Callee->eraseCallerEdge(Edge); |
| 1588 | Caller->eraseCalleeEdge(Edge); |
| 1589 | } else if (CalleeIter) { |
| 1590 | Callee->eraseCallerEdge(Edge); |
| 1591 | *EI = Caller->CalleeEdges.erase(*EI); |
| 1592 | } else { |
| 1593 | Caller->eraseCalleeEdge(Edge); |
| 1594 | *EI = Callee->CallerEdges.erase(*EI); |
| 1595 | } |
| 1596 | assert(Callee->CallerEdges.size() < CalleeCallerCount); |
| 1597 | assert(Caller->CalleeEdges.size() < CallerCalleeCount); |
| 1598 | } |
| 1599 | |
| 1600 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1601 | void CallsiteContextGraph< |
| 1602 | DerivedCCG, FuncTy, CallTy>::removeNoneTypeCalleeEdges(ContextNode *Node) { |
| 1603 | for (auto EI = Node->CalleeEdges.begin(); EI != Node->CalleeEdges.end();) { |
| 1604 | auto Edge = *EI; |
| 1605 | if (Edge->AllocTypes == (uint8_t)AllocationType::None) { |
| 1606 | assert(Edge->ContextIds.empty()); |
| 1607 | removeEdgeFromGraph(Edge: Edge.get(), EI: &EI, /*CalleeIter=*/true); |
| 1608 | } else |
| 1609 | ++EI; |
| 1610 | } |
| 1611 | } |
| 1612 | |
| 1613 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1614 | void CallsiteContextGraph< |
| 1615 | DerivedCCG, FuncTy, CallTy>::removeNoneTypeCallerEdges(ContextNode *Node) { |
| 1616 | for (auto EI = Node->CallerEdges.begin(); EI != Node->CallerEdges.end();) { |
| 1617 | auto Edge = *EI; |
| 1618 | if (Edge->AllocTypes == (uint8_t)AllocationType::None) { |
| 1619 | assert(Edge->ContextIds.empty()); |
| 1620 | Edge->Caller->eraseCalleeEdge(Edge.get()); |
| 1621 | EI = Node->CallerEdges.erase(EI); |
| 1622 | } else |
| 1623 | ++EI; |
| 1624 | } |
| 1625 | } |
| 1626 | |
| 1627 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1628 | typename CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextEdge * |
| 1629 | CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextNode:: |
| 1630 | findEdgeFromCallee(const ContextNode *Callee) { |
| 1631 | for (const auto &Edge : CalleeEdges) |
| 1632 | if (Edge->Callee == Callee) |
| 1633 | return Edge.get(); |
| 1634 | return nullptr; |
| 1635 | } |
| 1636 | |
| 1637 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1638 | typename CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextEdge * |
| 1639 | CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextNode:: |
| 1640 | findEdgeFromCaller(const ContextNode *Caller) { |
| 1641 | for (const auto &Edge : CallerEdges) |
| 1642 | if (Edge->Caller == Caller) |
| 1643 | return Edge.get(); |
| 1644 | return nullptr; |
| 1645 | } |
| 1646 | |
| 1647 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1648 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextNode:: |
| 1649 | eraseCalleeEdge(const ContextEdge *Edge) { |
| 1650 | auto EI = llvm::find_if( |
| 1651 | CalleeEdges, [Edge](const std::shared_ptr<ContextEdge> &CalleeEdge) { |
| 1652 | return CalleeEdge.get() == Edge; |
| 1653 | }); |
| 1654 | assert(EI != CalleeEdges.end()); |
| 1655 | CalleeEdges.erase(EI); |
| 1656 | } |
| 1657 | |
| 1658 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1659 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextNode:: |
| 1660 | eraseCallerEdge(const ContextEdge *Edge) { |
| 1661 | auto EI = llvm::find_if( |
| 1662 | CallerEdges, [Edge](const std::shared_ptr<ContextEdge> &CallerEdge) { |
| 1663 | return CallerEdge.get() == Edge; |
| 1664 | }); |
| 1665 | assert(EI != CallerEdges.end()); |
| 1666 | CallerEdges.erase(EI); |
| 1667 | } |
| 1668 | |
| 1669 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1670 | uint8_t CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::computeAllocType( |
| 1671 | DenseSet<uint32_t> &ContextIds) const { |
| 1672 | uint8_t BothTypes = |
| 1673 | (uint8_t)AllocationType::Cold | (uint8_t)AllocationType::NotCold; |
| 1674 | uint8_t AllocType = (uint8_t)AllocationType::None; |
| 1675 | for (auto Id : ContextIds) { |
| 1676 | AllocType |= (uint8_t)ContextIdToAllocationType.at(Val: Id); |
| 1677 | // Bail early if alloc type reached both, no further refinement. |
| 1678 | if (AllocType == BothTypes) |
| 1679 | return AllocType; |
| 1680 | } |
| 1681 | return AllocType; |
| 1682 | } |
| 1683 | |
| 1684 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1685 | uint8_t |
| 1686 | CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::intersectAllocTypesImpl( |
| 1687 | const DenseSet<uint32_t> &Node1Ids, |
| 1688 | const DenseSet<uint32_t> &Node2Ids) const { |
| 1689 | uint8_t BothTypes = |
| 1690 | (uint8_t)AllocationType::Cold | (uint8_t)AllocationType::NotCold; |
| 1691 | uint8_t AllocType = (uint8_t)AllocationType::None; |
| 1692 | for (auto Id : Node1Ids) { |
| 1693 | if (!Node2Ids.count(V: Id)) |
| 1694 | continue; |
| 1695 | AllocType |= (uint8_t)ContextIdToAllocationType.at(Val: Id); |
| 1696 | // Bail early if alloc type reached both, no further refinement. |
| 1697 | if (AllocType == BothTypes) |
| 1698 | return AllocType; |
| 1699 | } |
| 1700 | return AllocType; |
| 1701 | } |
| 1702 | |
| 1703 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1704 | uint8_t CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::intersectAllocTypes( |
| 1705 | const DenseSet<uint32_t> &Node1Ids, |
| 1706 | const DenseSet<uint32_t> &Node2Ids) const { |
| 1707 | if (Node1Ids.size() < Node2Ids.size()) |
| 1708 | return intersectAllocTypesImpl(Node1Ids, Node2Ids); |
| 1709 | else |
| 1710 | return intersectAllocTypesImpl(Node1Ids: Node2Ids, Node2Ids: Node1Ids); |
| 1711 | } |
| 1712 | |
| 1713 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1714 | typename CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextNode * |
| 1715 | CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::addAllocNode( |
| 1716 | CallInfo Call, const FuncTy *F) { |
| 1717 | assert(!getNodeForAlloc(Call)); |
| 1718 | ContextNode *AllocNode = createNewNode(/*IsAllocation=*/true, F, C: Call); |
| 1719 | AllocationCallToContextNodeMap[Call] = AllocNode; |
| 1720 | // Use LastContextId as a uniq id for MIB allocation nodes. |
| 1721 | AllocNode->OrigStackOrAllocId = LastContextId; |
| 1722 | // Alloc type should be updated as we add in the MIBs. We should assert |
| 1723 | // afterwards that it is not still None. |
| 1724 | AllocNode->AllocTypes = (uint8_t)AllocationType::None; |
| 1725 | |
| 1726 | return AllocNode; |
| 1727 | } |
| 1728 | |
| 1729 | static std::string getAllocTypeString(uint8_t AllocTypes) { |
| 1730 | if (!AllocTypes) |
| 1731 | return "None" ; |
| 1732 | std::string Str; |
| 1733 | if (AllocTypes & (uint8_t)AllocationType::NotCold) |
| 1734 | Str += "NotCold" ; |
| 1735 | if (AllocTypes & (uint8_t)AllocationType::Cold) |
| 1736 | Str += "Cold" ; |
| 1737 | return Str; |
| 1738 | } |
| 1739 | |
| 1740 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1741 | template <class NodeT, class IteratorT> |
| 1742 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::addStackNodesForMIB( |
| 1743 | ContextNode *AllocNode, CallStack<NodeT, IteratorT> &StackContext, |
| 1744 | CallStack<NodeT, IteratorT> &CallsiteContext, AllocationType AllocType, |
| 1745 | ArrayRef<ContextTotalSize> ContextSizeInfo, |
| 1746 | std::map<uint64_t, uint32_t> &TotalSizeToContextIdTopNCold) { |
| 1747 | // Treating the hot alloc type as NotCold before the disambiguation for "hot" |
| 1748 | // is done. |
| 1749 | if (AllocType == AllocationType::Hot) |
| 1750 | AllocType = AllocationType::NotCold; |
| 1751 | |
| 1752 | ContextIdToAllocationType[++LastContextId] = AllocType; |
| 1753 | |
| 1754 | bool IsImportant = false; |
| 1755 | if (!ContextSizeInfo.empty()) { |
| 1756 | auto &Entry = ContextIdToContextSizeInfos[LastContextId]; |
| 1757 | // If this is a cold allocation, and we are collecting non-zero largest |
| 1758 | // contexts, see if this is a candidate. |
| 1759 | if (AllocType == AllocationType::Cold && MemProfTopNImportant > 0) { |
| 1760 | uint64_t TotalCold = 0; |
| 1761 | for (auto &CSI : ContextSizeInfo) |
| 1762 | TotalCold += CSI.TotalSize; |
| 1763 | // Record this context if either we haven't found the first top-n largest |
| 1764 | // yet, or if it is larger than the smallest already recorded. |
| 1765 | if (TotalSizeToContextIdTopNCold.size() < MemProfTopNImportant || |
| 1766 | // Since TotalSizeToContextIdTopNCold is a std::map, it is implicitly |
| 1767 | // sorted in ascending size of its key which is the size. |
| 1768 | TotalCold > TotalSizeToContextIdTopNCold.begin()->first) { |
| 1769 | if (TotalSizeToContextIdTopNCold.size() == MemProfTopNImportant) { |
| 1770 | // Remove old one and its associated entries. |
| 1771 | auto IdToRemove = TotalSizeToContextIdTopNCold.begin()->second; |
| 1772 | TotalSizeToContextIdTopNCold.erase( |
| 1773 | position: TotalSizeToContextIdTopNCold.begin()); |
| 1774 | assert(ImportantContextIdInfo.count(IdToRemove)); |
| 1775 | ImportantContextIdInfo.erase(IdToRemove); |
| 1776 | } |
| 1777 | TotalSizeToContextIdTopNCold[TotalCold] = LastContextId; |
| 1778 | IsImportant = true; |
| 1779 | } |
| 1780 | } |
| 1781 | Entry.insert(position: Entry.begin(), first: ContextSizeInfo.begin(), last: ContextSizeInfo.end()); |
| 1782 | } |
| 1783 | |
| 1784 | // Update alloc type and context ids for this MIB. |
| 1785 | AllocNode->AllocTypes |= (uint8_t)AllocType; |
| 1786 | |
| 1787 | // Now add or update nodes for each stack id in alloc's context. |
| 1788 | // Later when processing the stack ids on non-alloc callsites we will adjust |
| 1789 | // for any inlining in the context. |
| 1790 | ContextNode *PrevNode = AllocNode; |
| 1791 | // Look for recursion (direct recursion should have been collapsed by |
| 1792 | // module summary analysis, here we should just be detecting mutual |
| 1793 | // recursion). Mark these nodes so we don't try to clone. |
| 1794 | SmallSet<uint64_t, 8> StackIdSet; |
| 1795 | // Skip any on the allocation call (inlining). |
| 1796 | for (auto ContextIter = StackContext.beginAfterSharedPrefix(CallsiteContext); |
| 1797 | ContextIter != StackContext.end(); ++ContextIter) { |
| 1798 | auto StackId = getStackId(IdOrIndex: *ContextIter); |
| 1799 | if (IsImportant) |
| 1800 | ImportantContextIdInfo[LastContextId].StackIds.push_back(StackId); |
| 1801 | ContextNode *StackNode = getNodeForStackId(StackId); |
| 1802 | if (!StackNode) { |
| 1803 | StackNode = createNewNode(/*IsAllocation=*/false); |
| 1804 | StackEntryIdToContextNodeMap[StackId] = StackNode; |
| 1805 | StackNode->OrigStackOrAllocId = StackId; |
| 1806 | } |
| 1807 | // Marking a node recursive will prevent its cloning completely, even for |
| 1808 | // non-recursive contexts flowing through it. |
| 1809 | if (!AllowRecursiveCallsites) { |
| 1810 | auto Ins = StackIdSet.insert(StackId); |
| 1811 | if (!Ins.second) |
| 1812 | StackNode->Recursive = true; |
| 1813 | } |
| 1814 | StackNode->AllocTypes |= (uint8_t)AllocType; |
| 1815 | PrevNode->addOrUpdateCallerEdge(StackNode, AllocType, LastContextId); |
| 1816 | PrevNode = StackNode; |
| 1817 | } |
| 1818 | } |
| 1819 | |
| 1820 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1821 | DenseSet<uint32_t> |
| 1822 | CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::duplicateContextIds( |
| 1823 | const DenseSet<uint32_t> &StackSequenceContextIds, |
| 1824 | DenseMap<uint32_t, DenseSet<uint32_t>> &OldToNewContextIds, |
| 1825 | function_ref<void(uint32_t OldId, uint32_t NewId)> OnDuplicate) { |
| 1826 | DenseSet<uint32_t> NewContextIds; |
| 1827 | NewContextIds.reserve(Size: StackSequenceContextIds.size()); |
| 1828 | for (auto OldId : StackSequenceContextIds) { |
| 1829 | NewContextIds.insert(V: ++LastContextId); |
| 1830 | OldToNewContextIds[OldId].insert(V: LastContextId); |
| 1831 | if (OnDuplicate) |
| 1832 | OnDuplicate(OldId, LastContextId); |
| 1833 | assert(ContextIdToAllocationType.count(OldId)); |
| 1834 | // The new context has the same allocation type and size info as original. |
| 1835 | ContextIdToAllocationType[LastContextId] = ContextIdToAllocationType[OldId]; |
| 1836 | auto CSI = ContextIdToContextSizeInfos.find(Val: OldId); |
| 1837 | if (CSI != ContextIdToContextSizeInfos.end()) |
| 1838 | ContextIdToContextSizeInfos[LastContextId] = CSI->second; |
| 1839 | if (DotAllocContextIds.contains(V: OldId)) |
| 1840 | DotAllocContextIds.insert(V: LastContextId); |
| 1841 | } |
| 1842 | return NewContextIds; |
| 1843 | } |
| 1844 | |
| 1845 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1846 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>:: |
| 1847 | propagateDuplicateContextIds( |
| 1848 | const DenseMap<uint32_t, DenseSet<uint32_t>> &OldToNewContextIds) { |
| 1849 | // Build a set of duplicated context ids corresponding to the input id set. |
| 1850 | auto GetNewIds = [&OldToNewContextIds](const DenseSet<uint32_t> &ContextIds) { |
| 1851 | DenseSet<uint32_t> NewIds; |
| 1852 | for (auto Id : ContextIds) |
| 1853 | if (auto NewId = OldToNewContextIds.find(Val: Id); |
| 1854 | NewId != OldToNewContextIds.end()) |
| 1855 | NewIds.insert_range(R: NewId->second); |
| 1856 | return NewIds; |
| 1857 | }; |
| 1858 | |
| 1859 | // Recursively update context ids sets along caller edges. |
| 1860 | auto UpdateCallers = [&](ContextNode *Node, |
| 1861 | DenseSet<const ContextEdge *> &Visited, |
| 1862 | auto &&UpdateCallers) -> void { |
| 1863 | for (const auto &Edge : Node->CallerEdges) { |
| 1864 | auto Inserted = Visited.insert(Edge.get()); |
| 1865 | if (!Inserted.second) |
| 1866 | continue; |
| 1867 | ContextNode *NextNode = Edge->Caller; |
| 1868 | DenseSet<uint32_t> NewIdsToAdd = GetNewIds(Edge->getContextIds()); |
| 1869 | // Only need to recursively iterate to NextNode via this caller edge if |
| 1870 | // it resulted in any added ids to NextNode. |
| 1871 | if (!NewIdsToAdd.empty()) { |
| 1872 | Edge->getContextIds().insert_range(NewIdsToAdd); |
| 1873 | UpdateCallers(NextNode, Visited, UpdateCallers); |
| 1874 | } |
| 1875 | } |
| 1876 | }; |
| 1877 | |
| 1878 | DenseSet<const ContextEdge *> Visited; |
| 1879 | for (auto &Entry : AllocationCallToContextNodeMap) { |
| 1880 | auto *Node = Entry.second; |
| 1881 | UpdateCallers(Node, Visited, UpdateCallers); |
| 1882 | } |
| 1883 | } |
| 1884 | |
| 1885 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1886 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::connectNewNode( |
| 1887 | ContextNode *NewNode, ContextNode *OrigNode, bool TowardsCallee, |
| 1888 | const DenseSet<uint32_t> &ContextIds) { |
| 1889 | auto &OrigEdges = |
| 1890 | TowardsCallee ? OrigNode->CalleeEdges : OrigNode->CallerEdges; |
| 1891 | // Increment iterator in loop so that we can remove edges as needed. |
| 1892 | for (auto EI = OrigEdges.begin(); EI != OrigEdges.end();) { |
| 1893 | auto Edge = *EI; |
| 1894 | DenseSet<uint32_t> NewEdgeContextIds; |
| 1895 | // Remove any matching context ids from Edge, return set that were found and |
| 1896 | // removed, these are the new edge's context ids. |
| 1897 | set_subtract(Edge->getContextIds(), ContextIds, NewEdgeContextIds); |
| 1898 | // If no matching context ids for this edge, skip it. |
| 1899 | if (NewEdgeContextIds.empty()) { |
| 1900 | ++EI; |
| 1901 | continue; |
| 1902 | } |
| 1903 | if (TowardsCallee) { |
| 1904 | uint8_t NewAllocType = computeAllocType(ContextIds&: NewEdgeContextIds); |
| 1905 | auto NewEdge = std::make_shared<ContextEdge>( |
| 1906 | Edge->Callee, NewNode, NewAllocType, std::move(NewEdgeContextIds)); |
| 1907 | NewNode->CalleeEdges.push_back(NewEdge); |
| 1908 | NewEdge->Callee->CallerEdges.push_back(NewEdge); |
| 1909 | } else { |
| 1910 | uint8_t NewAllocType = computeAllocType(ContextIds&: NewEdgeContextIds); |
| 1911 | auto NewEdge = std::make_shared<ContextEdge>( |
| 1912 | NewNode, Edge->Caller, NewAllocType, std::move(NewEdgeContextIds)); |
| 1913 | NewNode->CallerEdges.push_back(NewEdge); |
| 1914 | NewEdge->Caller->CalleeEdges.push_back(NewEdge); |
| 1915 | } |
| 1916 | // Remove old edge if context ids empty, otherwise update its alloc type to |
| 1917 | // reflect the removed context ids. |
| 1918 | if (Edge->getContextIds().empty()) { |
| 1919 | removeEdgeFromGraph(Edge: Edge.get(), EI: &EI, CalleeIter: TowardsCallee); |
| 1920 | continue; |
| 1921 | } |
| 1922 | Edge->AllocTypes = computeAllocType(ContextIds&: Edge->getContextIds()); |
| 1923 | ++EI; |
| 1924 | } |
| 1925 | } |
| 1926 | |
| 1927 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1928 | static void checkEdge( |
| 1929 | const std::shared_ptr<ContextEdge<DerivedCCG, FuncTy, CallTy>> &Edge) { |
| 1930 | // Confirm that alloc type is not None and that we have at least one context |
| 1931 | // id. |
| 1932 | assert(Edge->AllocTypes != (uint8_t)AllocationType::None); |
| 1933 | assert(!Edge->ContextIds.empty()); |
| 1934 | } |
| 1935 | |
| 1936 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1937 | static void checkNode(const ContextNode<DerivedCCG, FuncTy, CallTy> *Node, |
| 1938 | bool CheckEdges = true) { |
| 1939 | if (Node->isRemoved()) |
| 1940 | return; |
| 1941 | #ifndef NDEBUG |
| 1942 | // Compute node's context ids once for use in asserts. |
| 1943 | auto NodeContextIds = Node->getContextIds(); |
| 1944 | #endif |
| 1945 | // Node's context ids should be the union of both its callee and caller edge |
| 1946 | // context ids. |
| 1947 | if (Node->CallerEdges.size()) { |
| 1948 | DenseSet<uint32_t> CallerEdgeContextIds( |
| 1949 | Node->CallerEdges.front()->ContextIds); |
| 1950 | for (const auto &Edge : llvm::drop_begin(Node->CallerEdges)) { |
| 1951 | if (CheckEdges) |
| 1952 | checkEdge<DerivedCCG, FuncTy, CallTy>(Edge); |
| 1953 | set_union(CallerEdgeContextIds, Edge->ContextIds); |
| 1954 | } |
| 1955 | // Node can have more context ids than callers if some contexts terminate at |
| 1956 | // node and some are longer. If we are allowing recursive callsites and |
| 1957 | // contexts this will be violated for incompletely cloned recursive cycles, |
| 1958 | // so skip the checking in that case. |
| 1959 | assert((AllowRecursiveCallsites && AllowRecursiveContexts) || |
| 1960 | NodeContextIds == CallerEdgeContextIds || |
| 1961 | set_is_subset(CallerEdgeContextIds, NodeContextIds)); |
| 1962 | } |
| 1963 | if (Node->CalleeEdges.size()) { |
| 1964 | DenseSet<uint32_t> CalleeEdgeContextIds( |
| 1965 | Node->CalleeEdges.front()->ContextIds); |
| 1966 | for (const auto &Edge : llvm::drop_begin(Node->CalleeEdges)) { |
| 1967 | if (CheckEdges) |
| 1968 | checkEdge<DerivedCCG, FuncTy, CallTy>(Edge); |
| 1969 | set_union(CalleeEdgeContextIds, Edge->getContextIds()); |
| 1970 | } |
| 1971 | // If we are allowing recursive callsites and contexts this will be violated |
| 1972 | // for incompletely cloned recursive cycles, so skip the checking in that |
| 1973 | // case. |
| 1974 | assert((AllowRecursiveCallsites && AllowRecursiveContexts) || |
| 1975 | NodeContextIds == CalleeEdgeContextIds); |
| 1976 | } |
| 1977 | // FIXME: Since this checking is only invoked under an option, we should |
| 1978 | // change the error checking from using assert to something that will trigger |
| 1979 | // an error on a release build. |
| 1980 | #ifndef NDEBUG |
| 1981 | // Make sure we don't end up with duplicate edges between the same caller and |
| 1982 | // callee. |
| 1983 | DenseSet<ContextNode<DerivedCCG, FuncTy, CallTy> *> NodeSet; |
| 1984 | for (const auto &E : Node->CalleeEdges) |
| 1985 | NodeSet.insert(E->Callee); |
| 1986 | assert(NodeSet.size() == Node->CalleeEdges.size()); |
| 1987 | #endif |
| 1988 | } |
| 1989 | |
| 1990 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 1991 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>:: |
| 1992 | assignStackNodesPostOrder(ContextNode *Node, |
| 1993 | DenseSet<const ContextNode *> &Visited, |
| 1994 | DenseMap<uint64_t, std::vector<CallContextInfo>> |
| 1995 | &StackIdToMatchingCalls, |
| 1996 | DenseMap<CallInfo, CallInfo> &CallToMatchingCall, |
| 1997 | const DenseSet<uint32_t> &ImportantContextIds) { |
| 1998 | auto Inserted = Visited.insert(Node); |
| 1999 | if (!Inserted.second) |
| 2000 | return; |
| 2001 | // Post order traversal. Iterate over a copy since we may add nodes and |
| 2002 | // therefore new callers during the recursive call, invalidating any |
| 2003 | // iterator over the original edge vector. We don't need to process these |
| 2004 | // new nodes as they were already processed on creation. |
| 2005 | auto CallerEdges = Node->CallerEdges; |
| 2006 | for (auto &Edge : CallerEdges) { |
| 2007 | // Skip any that have been removed during the recursion. |
| 2008 | if (Edge->isRemoved()) { |
| 2009 | assert(!is_contained(Node->CallerEdges, Edge)); |
| 2010 | continue; |
| 2011 | } |
| 2012 | assignStackNodesPostOrder(Node: Edge->Caller, Visited, StackIdToMatchingCalls, |
| 2013 | CallToMatchingCall, ImportantContextIds); |
| 2014 | } |
| 2015 | |
| 2016 | // If this node's stack id is in the map, update the graph to contain new |
| 2017 | // nodes representing any inlining at interior callsites. Note we move the |
| 2018 | // associated context ids over to the new nodes. |
| 2019 | |
| 2020 | // Ignore this node if it is for an allocation or we didn't record any |
| 2021 | // stack id lists ending at it. |
| 2022 | if (Node->IsAllocation || |
| 2023 | !StackIdToMatchingCalls.count(Node->OrigStackOrAllocId)) |
| 2024 | return; |
| 2025 | |
| 2026 | auto &Calls = StackIdToMatchingCalls[Node->OrigStackOrAllocId]; |
| 2027 | // Handle the simple case first. A single call with a single stack id. |
| 2028 | // In this case there is no need to create any new context nodes, simply |
| 2029 | // assign the context node for stack id to this Call. |
| 2030 | if (Calls.size() == 1) { |
| 2031 | auto &[Call, StackNodes, Func, SavedContextIds, TrimmedInlinedDupIds] = |
| 2032 | Calls[0]; |
| 2033 | if (StackNodes.size() == 1) { |
| 2034 | assert(SavedContextIds.empty()); |
| 2035 | // Only calls with inlined frames get these. |
| 2036 | assert(TrimmedInlinedDupIds.empty()); |
| 2037 | // It should be this Node |
| 2038 | assert(Node == StackNodes[0]); |
| 2039 | if (Node->Recursive) |
| 2040 | return; |
| 2041 | Node->setCall(Call); |
| 2042 | NonAllocationCallToContextNodeMap[Call] = Node; |
| 2043 | NodeToCallingFunc[Node] = Func; |
| 2044 | recordStackNode(StackNodes, Node, NodeContextIds: Node->getContextIds(), |
| 2045 | ImportantContextIds); |
| 2046 | return; |
| 2047 | } |
| 2048 | } |
| 2049 | |
| 2050 | #ifndef NDEBUG |
| 2051 | // Find the node for the last stack id, which should be the same |
| 2052 | // across all calls recorded for this id, and is this node's id. |
| 2053 | uint64_t LastId = Node->OrigStackOrAllocId; |
| 2054 | ContextNode *LastNode = getNodeForStackId(LastId); |
| 2055 | // We should only have kept stack ids that had nodes. |
| 2056 | assert(LastNode); |
| 2057 | assert(LastNode == Node); |
| 2058 | #else |
| 2059 | ContextNode *LastNode = Node; |
| 2060 | #endif |
| 2061 | |
| 2062 | // Compute the last node's context ids once, as it is shared by all calls in |
| 2063 | // this entry. |
| 2064 | DenseSet<uint32_t> LastNodeContextIds = LastNode->getContextIds(); |
| 2065 | |
| 2066 | [[maybe_unused]] bool PrevIterCreatedNode = false; |
| 2067 | bool CreatedNode = false; |
| 2068 | for (unsigned I = 0; I < Calls.size(); |
| 2069 | I++, PrevIterCreatedNode = CreatedNode) { |
| 2070 | CreatedNode = false; |
| 2071 | auto &[Call, StackNodes, Func, SavedContextIds, TrimmedInlinedDupIds] = |
| 2072 | Calls[I]; |
| 2073 | // Skip any for which we didn't assign any ids, these don't get a node in |
| 2074 | // the graph. |
| 2075 | if (SavedContextIds.empty() && TrimmedInlinedDupIds.empty()) { |
| 2076 | // If this call has a matching call (located in the same function and |
| 2077 | // having the same stack ids), simply add it to the context node created |
| 2078 | // for its matching call earlier. These can be treated the same through |
| 2079 | // cloning and get updated at the same time. |
| 2080 | if (!CallToMatchingCall.contains(Call)) |
| 2081 | continue; |
| 2082 | auto MatchingCall = CallToMatchingCall[Call]; |
| 2083 | if (!NonAllocationCallToContextNodeMap.contains(MatchingCall)) { |
| 2084 | // This should only happen if we had a prior iteration, and it didn't |
| 2085 | // create a node because of the below recomputation of context ids |
| 2086 | // finding none remaining and continuing early. |
| 2087 | assert(I > 0 && !PrevIterCreatedNode); |
| 2088 | continue; |
| 2089 | } |
| 2090 | NonAllocationCallToContextNodeMap[MatchingCall]->MatchingCalls.push_back( |
| 2091 | Call); |
| 2092 | continue; |
| 2093 | } |
| 2094 | |
| 2095 | assert(LastNode == StackNodes.back()); |
| 2096 | |
| 2097 | // Recompute the context ids for this stack id sequence (the |
| 2098 | // intersection of the context ids of the corresponding nodes). |
| 2099 | // Start with the ids we saved in the map for this call, which could be |
| 2100 | // duplicated context ids. We have to recompute as we might have overlap |
| 2101 | // overlap between the saved context ids for different last nodes, and |
| 2102 | // removed them already during the post order traversal. |
| 2103 | set_intersect(SavedContextIds, LastNodeContextIds); |
| 2104 | ContextNode *PrevNode = LastNode; |
| 2105 | bool Skip = SavedContextIds.empty(); |
| 2106 | // Iterate backwards through the stack nodes, starting after the last one |
| 2107 | // in the list, which was handled once outside for all Calls. |
| 2108 | for (auto NodeIter = StackNodes.rbegin() + 1; |
| 2109 | !Skip && NodeIter != StackNodes.rend(); NodeIter++) { |
| 2110 | ContextNode *CurNode = *NodeIter; |
| 2111 | // We should only have kept stack ids that had nodes and weren't |
| 2112 | // recursive. |
| 2113 | assert(CurNode); |
| 2114 | assert(!CurNode->Recursive); |
| 2115 | |
| 2116 | auto *Edge = CurNode->findEdgeFromCaller(PrevNode); |
| 2117 | if (!Edge) { |
| 2118 | Skip = true; |
| 2119 | break; |
| 2120 | } |
| 2121 | PrevNode = CurNode; |
| 2122 | |
| 2123 | // Update the context ids, which is the intersection of the ids along |
| 2124 | // all edges in the sequence. |
| 2125 | set_intersect(SavedContextIds, Edge->getContextIds()); |
| 2126 | |
| 2127 | // If we now have no context ids for clone, skip this call. |
| 2128 | if (SavedContextIds.empty()) { |
| 2129 | Skip = true; |
| 2130 | break; |
| 2131 | } |
| 2132 | } |
| 2133 | if (Skip) |
| 2134 | SavedContextIds.clear(); |
| 2135 | // Add the duplicates created for this call (see updateStackNodes) that are |
| 2136 | // still on its first node. Other calls given them end below that node and |
| 2137 | // leave them on it, but a recursive context may traverse this call's |
| 2138 | // initial nodes more than once with different continuations, in which |
| 2139 | // case a call assigned one of the other traversals may have been processed |
| 2140 | // first and claimed them (as happens to the original ids too). |
| 2141 | if (!TrimmedInlinedDupIds.empty()) { |
| 2142 | StackNodes[0]->removeContextIdsNotPresent(TrimmedInlinedDupIds); |
| 2143 | SavedContextIds.insert_range(TrimmedInlinedDupIds); |
| 2144 | } |
| 2145 | if (SavedContextIds.empty()) |
| 2146 | continue; |
| 2147 | |
| 2148 | // Create new context node. |
| 2149 | ContextNode *NewNode = createNewNode(/*IsAllocation=*/false, F: Func, C: Call); |
| 2150 | NonAllocationCallToContextNodeMap[Call] = NewNode; |
| 2151 | CreatedNode = true; |
| 2152 | NewNode->AllocTypes = computeAllocType(ContextIds&: SavedContextIds); |
| 2153 | |
| 2154 | ContextNode *FirstNode = StackNodes[0]; |
| 2155 | assert(FirstNode); |
| 2156 | |
| 2157 | // Connect to callees of innermost stack frame in inlined call chain. |
| 2158 | // This updates context ids for FirstNode's callee's to reflect those |
| 2159 | // moved to NewNode. |
| 2160 | connectNewNode(NewNode, OrigNode: FirstNode, /*TowardsCallee=*/true, ContextIds: SavedContextIds); |
| 2161 | |
| 2162 | // Connect to callers of outermost stack frame in inlined call chain. |
| 2163 | // This updates context ids for FirstNode's caller's to reflect those |
| 2164 | // moved to NewNode. |
| 2165 | connectNewNode(NewNode, OrigNode: LastNode, /*TowardsCallee=*/false, ContextIds: SavedContextIds); |
| 2166 | |
| 2167 | // Now we need to remove context ids from edges/nodes between First and |
| 2168 | // Last Node. |
| 2169 | PrevNode = nullptr; |
| 2170 | for (ContextNode *CurNode : StackNodes) { |
| 2171 | // We should only have kept stack ids that had nodes. |
| 2172 | assert(CurNode); |
| 2173 | |
| 2174 | // Remove the context ids moved to NewNode from CurNode, and the |
| 2175 | // edge from the prior node. |
| 2176 | if (PrevNode) { |
| 2177 | auto *PrevEdge = CurNode->findEdgeFromCallee(PrevNode); |
| 2178 | // If the sequence contained recursion, we might have already removed |
| 2179 | // some edges during the connectNewNode calls above. |
| 2180 | if (!PrevEdge) { |
| 2181 | PrevNode = CurNode; |
| 2182 | continue; |
| 2183 | } |
| 2184 | set_subtract(PrevEdge->getContextIds(), SavedContextIds); |
| 2185 | if (PrevEdge->getContextIds().empty()) |
| 2186 | removeEdgeFromGraph(Edge: PrevEdge); |
| 2187 | else |
| 2188 | PrevEdge->AllocTypes = computeAllocType(ContextIds&: PrevEdge->getContextIds()); |
| 2189 | } |
| 2190 | // Since we update the edges from leaf to tail, only look at the callee |
| 2191 | // edges. This isn't an alloc node, so if there are no callee edges, the |
| 2192 | // alloc type is None. |
| 2193 | CurNode->AllocTypes = CurNode->CalleeEdges.empty() |
| 2194 | ? (uint8_t)AllocationType::None |
| 2195 | : CurNode->computeAllocType(); |
| 2196 | PrevNode = CurNode; |
| 2197 | } |
| 2198 | |
| 2199 | recordStackNode(StackNodes, Node: NewNode, NodeContextIds: SavedContextIds, ImportantContextIds); |
| 2200 | |
| 2201 | if (VerifyNodes) { |
| 2202 | checkNode<DerivedCCG, FuncTy, CallTy>(NewNode, /*CheckEdges=*/true); |
| 2203 | for (ContextNode *CurNode : StackNodes) { |
| 2204 | // We should only have kept stack ids that had nodes. |
| 2205 | assert(CurNode); |
| 2206 | checkNode<DerivedCCG, FuncTy, CallTy>(CurNode, /*CheckEdges=*/true); |
| 2207 | } |
| 2208 | } |
| 2209 | } |
| 2210 | } |
| 2211 | |
| 2212 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 2213 | void CallsiteContextGraph<DerivedCCG, FuncTy, |
| 2214 | CallTy>::fixupImportantContexts() { |
| 2215 | if (ImportantContextIdInfo.empty()) |
| 2216 | return; |
| 2217 | |
| 2218 | // Update statistics as we are done building this map at this point. |
| 2219 | NumImportantContextIds = ImportantContextIdInfo.size(); |
| 2220 | |
| 2221 | if (!MemProfFixupImportant) |
| 2222 | return; |
| 2223 | |
| 2224 | if (ExportToDot) |
| 2225 | exportToDot(Label: "beforestackfixup" ); |
| 2226 | |
| 2227 | // For each context we identified as important, walk through the saved context |
| 2228 | // stack ids in order from leaf upwards, and make sure all edges are correct. |
| 2229 | // These can be difficult to get right when updating the graph while mapping |
| 2230 | // nodes onto summary or IR, especially when there is recursion. In |
| 2231 | // particular, when we have created new nodes to reflect inlining, it is |
| 2232 | // sometimes impossible to know exactly how to update the edges in the face of |
| 2233 | // recursion, as we have lost the original ordering of the stack ids in the |
| 2234 | // contexts. |
| 2235 | // TODO: Consider only doing this if we detect the context has recursive |
| 2236 | // cycles. |
| 2237 | // |
| 2238 | // I.e. assume we have a context with stack ids like: {A B A C A D E} |
| 2239 | // and let's say A was inlined into B, C, and D. The original graph will have |
| 2240 | // multiple recursive cycles through A. When we match the original context |
| 2241 | // nodes onto the IR or summary, we will merge {A B} into one context node, |
| 2242 | // {A C} onto another, and {A D} onto another. Looking at the stack sequence |
| 2243 | // above, we should end up with a non-cyclic set of edges like: |
| 2244 | // {AB} <- {AC} <- {AD} <- E. However, because we normally have lost the |
| 2245 | // original ordering, we won't get the edges correct initially (it's |
| 2246 | // impossible without the original ordering). Here we do the fixup (add and |
| 2247 | // removing edges where necessary) for this context. In the |
| 2248 | // ImportantContextInfo struct in this case we should have a MaxLength = 2, |
| 2249 | // and map entries for {A B}, {A C}, {A D}, and {E}. |
| 2250 | for (auto &[CurContextId, Info] : ImportantContextIdInfo) { |
| 2251 | if (Info.StackIdsToNode.empty()) |
| 2252 | continue; |
| 2253 | bool Changed = false; |
| 2254 | ContextNode *PrevNode = nullptr; |
| 2255 | ContextNode *CurNode = nullptr; |
| 2256 | DenseSet<const ContextEdge *> VisitedEdges; |
| 2257 | ArrayRef<uint64_t> AllStackIds(Info.StackIds); |
| 2258 | // Try to identify what callsite ContextNode maps to which slice of the |
| 2259 | // context's ordered stack ids. |
| 2260 | for (unsigned I = 0; I < AllStackIds.size(); I++, PrevNode = CurNode) { |
| 2261 | // We will do this greedily, trying up to MaxLength stack ids in a row, to |
| 2262 | // see if we recorded a context node for that sequence. |
| 2263 | auto Len = Info.MaxLength; |
| 2264 | auto LenToEnd = AllStackIds.size() - I; |
| 2265 | if (Len > LenToEnd) |
| 2266 | Len = LenToEnd; |
| 2267 | CurNode = nullptr; |
| 2268 | // Try to find a recorded context node starting with the longest length |
| 2269 | // recorded, and on down until we check for just a single stack node. |
| 2270 | for (; Len > 0; Len--) { |
| 2271 | // Get the slice of the original stack id sequence to check. |
| 2272 | auto CheckStackIds = AllStackIds.slice(I, Len); |
| 2273 | auto EntryIt = Info.StackIdsToNode.find(CheckStackIds); |
| 2274 | if (EntryIt == Info.StackIdsToNode.end()) |
| 2275 | continue; |
| 2276 | CurNode = EntryIt->second; |
| 2277 | // Skip forward so we don't try to look for the ones we just matched. |
| 2278 | // We increment by Len - 1, because the outer for loop will increment I. |
| 2279 | I += Len - 1; |
| 2280 | break; |
| 2281 | } |
| 2282 | // Give up if we couldn't find a node. Since we need to clone from the |
| 2283 | // leaf allocation upwards, no sense in doing anymore fixup further up |
| 2284 | // the context if we couldn't match part of the original stack context |
| 2285 | // onto a callsite node. |
| 2286 | if (!CurNode) |
| 2287 | break; |
| 2288 | // No edges to fix up until we have a pair of nodes that should be |
| 2289 | // adjacent in the graph. |
| 2290 | if (!PrevNode) |
| 2291 | continue; |
| 2292 | // See if we already have a call edge from CurNode to PrevNode. |
| 2293 | auto *CurEdge = PrevNode->findEdgeFromCaller(CurNode); |
| 2294 | if (CurEdge) { |
| 2295 | // We already have an edge. Make sure it contains this context id. |
| 2296 | if (CurEdge->getContextIds().insert(CurContextId).second) { |
| 2297 | NumFixupEdgeIdsInserted++; |
| 2298 | Changed = true; |
| 2299 | } |
| 2300 | } else { |
| 2301 | // No edge exists - add one. |
| 2302 | NumFixupEdgesAdded++; |
| 2303 | DenseSet<uint32_t> ContextIds({CurContextId}); |
| 2304 | auto AllocType = computeAllocType(ContextIds); |
| 2305 | auto NewEdge = std::make_shared<ContextEdge>( |
| 2306 | PrevNode, CurNode, AllocType, std::move(ContextIds)); |
| 2307 | PrevNode->CallerEdges.push_back(NewEdge); |
| 2308 | CurNode->CalleeEdges.push_back(NewEdge); |
| 2309 | // Save the new edge for the below handling. |
| 2310 | CurEdge = NewEdge.get(); |
| 2311 | Changed = true; |
| 2312 | } |
| 2313 | VisitedEdges.insert(CurEdge); |
| 2314 | // Now remove this context id from any other caller edges calling |
| 2315 | // PrevNode. |
| 2316 | for (auto &Edge : PrevNode->CallerEdges) { |
| 2317 | // Skip the edge updating/created above and edges we have already |
| 2318 | // visited (due to recursion). |
| 2319 | if (Edge.get() != CurEdge && !VisitedEdges.contains(Edge.get())) |
| 2320 | Edge->getContextIds().erase(CurContextId); |
| 2321 | } |
| 2322 | } |
| 2323 | if (Changed) |
| 2324 | NumFixedContexts++; |
| 2325 | } |
| 2326 | } |
| 2327 | |
| 2328 | // Returns the ids of contexts whose stacks end partway through the given |
| 2329 | // call's inlined sequence of stack nodes: those that follow the sequence from |
| 2330 | // StackNodes[0] (the innermost inlined callee frame) through some node before |
| 2331 | // the last, and were trimmed at that node, i.e. have no caller edge out of it. |
| 2332 | // For example, given nodes [A, B, C], a context whose stack ends at A, or that |
| 2333 | // follows A->B and ends at B, is returned, but one that follows A->B->C (which |
| 2334 | // is assigned to the call normally) or A->D is not. |
| 2335 | // |
| 2336 | // Nothing is returned if the sequence contains a recursive node before the |
| 2337 | // last, as updateStackNodes does not assign such calls a node at all. |
| 2338 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 2339 | DenseSet<uint32_t> CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>:: |
| 2340 | computeTrimmedInlinedContextIds(ArrayRef<ContextNode *> StackNodes, |
| 2341 | TrimmedAtNodeCache &Cache) { |
| 2342 | assert(StackNodes.size() >= 2); |
| 2343 | DenseSet<uint32_t> Result; |
| 2344 | |
| 2345 | // Returns the index in the cache of the ids of contexts trimmed at the given |
| 2346 | // node. This only depends on the node, which is often shared by many calls' |
| 2347 | // inlined sequences, so it is computed once per node. |
| 2348 | auto GetTrimmedAtNodeIndex = [&Cache](const ContextNode *Node) { |
| 2349 | auto [It, Inserted] = |
| 2350 | Cache.NodeToIndex.try_emplace(Node, Cache.TrimmedAtNodeIds.size()); |
| 2351 | unsigned Index = It->second; |
| 2352 | if (!Inserted) |
| 2353 | return Index; |
| 2354 | // All of the node's context ids flow in through its callee edges. Those |
| 2355 | // not flowing out through a caller edge were trimmed here. |
| 2356 | DenseSet<uint32_t> &Trimmed = Cache.TrimmedAtNodeIds.emplace_back(); |
| 2357 | for (const auto &CalleeEdge : Node->CalleeEdges) |
| 2358 | Trimmed.insert_range(CalleeEdge->getContextIds()); |
| 2359 | for (const auto &CallerEdge : Node->CallerEdges) { |
| 2360 | set_subtract(Trimmed, CallerEdge->getContextIds()); |
| 2361 | if (Trimmed.empty()) |
| 2362 | break; |
| 2363 | } |
| 2364 | return Index; |
| 2365 | }; |
| 2366 | |
| 2367 | // Collect the cache index for each node but the last. Skip the walk |
| 2368 | // entirely if no contexts were trimmed at any of them (the common case). |
| 2369 | SmallVector<unsigned, 8> TrimmedAtNodeIdsIndices; |
| 2370 | bool AnyTrimmed = false; |
| 2371 | for (ContextNode *Node : StackNodes.drop_back()) { |
| 2372 | assert(Node); |
| 2373 | if (Node->Recursive) |
| 2374 | return Result; |
| 2375 | unsigned Index = GetTrimmedAtNodeIndex(Node); |
| 2376 | AnyTrimmed |= !Cache.TrimmedAtNodeIds[Index].empty(); |
| 2377 | TrimmedAtNodeIdsIndices.push_back(Elt: Index); |
| 2378 | } |
| 2379 | if (!AnyTrimmed) |
| 2380 | return Result; |
| 2381 | |
| 2382 | // Walk the sequence from the innermost node outward, tracking in |
| 2383 | // PrefixContextIds the contexts that follow the call's stack nodes from the |
| 2384 | // first through the current one. Any of those contexts that were trimmed at |
| 2385 | // the current node end partway through the call's sequence, so add them to |
| 2386 | // Result. All contexts on the first node trivially follow the sequence up to |
| 2387 | // it, so start with all those trimmed there, and start PrefixContextIds with |
| 2388 | // the ids on the edge out of the first node (rather than the possibly very |
| 2389 | // large set of ids on the first node itself). |
| 2390 | // |
| 2391 | // If an edge is missing between two of the nodes, the contexts trimmed at |
| 2392 | // the nodes before it (in particular all those trimmed at the first node) |
| 2393 | // are deliberately kept, even though updateStackNodes does not assign such |
| 2394 | // a call any contexts otherwise: a trimmed context is unambiguous for every |
| 2395 | // continuation, including an inlined sequence never observed in the profile. |
| 2396 | Result = Cache.TrimmedAtNodeIds[TrimmedAtNodeIdsIndices[0]]; |
| 2397 | DenseSet<uint32_t> PrefixContextIds; |
| 2398 | for (unsigned I = 1; I < TrimmedAtNodeIdsIndices.size(); I++) { |
| 2399 | auto *Edge = StackNodes[I]->findEdgeFromCallee(StackNodes[I - 1]); |
| 2400 | if (!Edge) |
| 2401 | break; |
| 2402 | if (I == 1) |
| 2403 | PrefixContextIds = Edge->getContextIds(); |
| 2404 | else |
| 2405 | set_intersect(PrefixContextIds, Edge->getContextIds()); |
| 2406 | if (PrefixContextIds.empty()) |
| 2407 | break; |
| 2408 | Result.insert_range(set_intersection( |
| 2409 | PrefixContextIds, Cache.TrimmedAtNodeIds[TrimmedAtNodeIdsIndices[I]])); |
| 2410 | } |
| 2411 | return Result; |
| 2412 | } |
| 2413 | |
| 2414 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 2415 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::updateStackNodes() { |
| 2416 | // Map of stack id to all calls with that as the last (outermost caller) |
| 2417 | // callsite id that has a context node (some might not due to trimming |
| 2418 | // performed during matching of the allocation profile contexts). |
| 2419 | // The CallContextInfo contains the Call and a list of the context nodes for |
| 2420 | // its stack ids, the function containing Call, and the set of context ids |
| 2421 | // the analysis will eventually identify for use in any new node created |
| 2422 | // for that callsite. |
| 2423 | DenseMap<uint64_t, std::vector<CallContextInfo>> StackIdToMatchingCalls; |
| 2424 | for (auto &[Func, CallsWithMetadata] : FuncToCallsWithMetadata) { |
| 2425 | for (auto &Call : CallsWithMetadata) { |
| 2426 | // Ignore allocations, already handled. |
| 2427 | if (AllocationCallToContextNodeMap.count(Call)) |
| 2428 | continue; |
| 2429 | std::vector<ContextNode *> StackNodes = getStackNodesForCall(Call: Call.call()); |
| 2430 | // If there were no nodes created for MIBs on allocs (maybe this was in |
| 2431 | // the unambiguous part of the MIB stack that was trimmed), ignore. |
| 2432 | if (StackNodes.empty()) |
| 2433 | continue; |
| 2434 | // Otherwise, record this Call along with the list of nodes for the last |
| 2435 | // (outermost caller) stack id with a node. |
| 2436 | uint64_t LastId = StackNodes.back()->OrigStackOrAllocId; |
| 2437 | StackIdToMatchingCalls[LastId].push_back( |
| 2438 | {Call.call(), std::move(StackNodes), Func, {}, {}}); |
| 2439 | } |
| 2440 | } |
| 2441 | |
| 2442 | // In order to do the best and maximal matching of inlined calls to context |
| 2443 | // node sequences, sort the calls with each last stack id by the length of |
| 2444 | // their stack node sequences in descending order, and within each length, |
| 2445 | // lexicographically by stack id. The latter is so that we can specially |
| 2446 | // handle calls that have identical stack id sequences (either due to cloning |
| 2447 | // or artificially because of the MIB context trimming). Those with the same |
| 2448 | // stack ids are then sorted by function to facilitate efficiently mapping |
| 2449 | // them to the same context node. Because the functions are pointers, to |
| 2450 | // ensure a stable sort first assign each function pointer to its first index |
| 2451 | // in the Calls array, and then use that to sort by. The stack nodes are |
| 2452 | // compared by stack id rather than by pointer, for a deterministic order. |
| 2453 | auto StackIdsLess = [](ArrayRef<ContextNode *> A, ArrayRef<ContextNode *> B) { |
| 2454 | return std::lexicographical_compare( |
| 2455 | A.begin(), A.end(), B.begin(), B.end(), |
| 2456 | [](const ContextNode *X, const ContextNode *Y) { |
| 2457 | return X->OrigStackOrAllocId < Y->OrigStackOrAllocId; |
| 2458 | }); |
| 2459 | }; |
| 2460 | // Save a map from each call to any that are found to match it. I.e. located |
| 2461 | // in the same function and have the same (possibly trimmed) stack ids. We use |
| 2462 | // this to avoid creating extra graph nodes as they can be treated the same. |
| 2463 | DenseMap<CallInfo, CallInfo> CallToMatchingCall; |
| 2464 | for (auto &Entry : StackIdToMatchingCalls) { |
| 2465 | auto &Calls = Entry.second; |
| 2466 | if (Calls.size() == 1) |
| 2467 | continue; |
| 2468 | DenseMap<const FuncTy *, unsigned> FuncToIndex; |
| 2469 | for (const auto &[Idx, CallCtxInfo] : enumerate(Calls)) |
| 2470 | FuncToIndex.insert({CallCtxInfo.Func, Idx}); |
| 2471 | auto CallLess = [&FuncToIndex, &StackIdsLess](const CallContextInfo &A, |
| 2472 | const CallContextInfo &B) { |
| 2473 | return A.StackNodes.size() > B.StackNodes.size() || |
| 2474 | (A.StackNodes.size() == B.StackNodes.size() && |
| 2475 | (StackIdsLess(A.StackNodes, B.StackNodes) || |
| 2476 | (A.StackNodes == B.StackNodes && |
| 2477 | FuncToIndex[A.Func] < FuncToIndex[B.Func]))); |
| 2478 | }; |
| 2479 | llvm::stable_sort(Calls, CallLess); |
| 2480 | // Matching calls are now adjacent. Record each run's calls after the first |
| 2481 | // as matching it. Comparing the stack nodes is equivalent to comparing the |
| 2482 | // stack ids, as each stack id has its own node and the stack id to node |
| 2483 | // map is not modified here. |
| 2484 | unsigned FirstInRun = 0; |
| 2485 | for (unsigned I = 1; I < Calls.size(); I++) { |
| 2486 | if (Calls[I].Func != Calls[FirstInRun].Func || |
| 2487 | Calls[I].StackNodes != Calls[FirstInRun].StackNodes) { |
| 2488 | FirstInRun = I; |
| 2489 | continue; |
| 2490 | } |
| 2491 | CallToMatchingCall[Calls[I].Call] = Calls[FirstInRun].Call; |
| 2492 | } |
| 2493 | } |
| 2494 | |
| 2495 | // First make a pass through all stack ids that correspond to a call, |
| 2496 | // as identified in the above loop. Compute the context ids corresponding to |
| 2497 | // each of these calls when they correspond to multiple stack ids due to |
| 2498 | // due to inlining. Perform any duplication of context ids required when |
| 2499 | // there is more than one call with the same stack ids. Their (possibly newly |
| 2500 | // duplicated) context ids are saved in the StackIdToMatchingCalls map. |
| 2501 | DenseMap<uint32_t, DenseSet<uint32_t>> OldToNewContextIds; |
| 2502 | |
| 2503 | // First, give each call with inlined frames duplicates of the ids of any |
| 2504 | // contexts trimmed at a stack id before the last one in its sequence (see |
| 2505 | // computeTrimmedInlinedContextIds), saved in its TrimmedInlinedDupIds. |
| 2506 | // Those contexts are also assigned, with their original ids, to calls whose |
| 2507 | // last stack id with a node is the earlier id, so this call gets its own |
| 2508 | // copies. This pass only reads the graph, as does the main loop below (the |
| 2509 | // duplicates are only added to the graph afterwards, by |
| 2510 | // propagateDuplicateContextIds), so it can run first; that lets the main |
| 2511 | // loop hand these duplicates to the other calls that should also get them |
| 2512 | // (see AddTrimmedInlinedDups). |
| 2513 | |
| 2514 | // The stack nodes of each call that duplicates were created for. These |
| 2515 | // reference the vectors in StackIdToMatchingCalls, which is not modified |
| 2516 | // after this point. |
| 2517 | std::vector<ArrayRef<ContextNode *>> TrimmedInlinedDupCreatorNodes; |
| 2518 | // Map from each original trimmed context id to its duplicates, each paired |
| 2519 | // with the index in TrimmedInlinedDupCreatorNodes of the creating call. |
| 2520 | DenseMap<uint32_t, SmallVector<std::pair<uint32_t, unsigned>, 1>> |
| 2521 | OrigToTrimmedInlinedDups; |
| 2522 | if (AssignTrimmedInlinedContexts) { |
| 2523 | TrimmedAtNodeCache Cache; |
| 2524 | for (auto &Entry : StackIdToMatchingCalls) { |
| 2525 | auto &Calls = Entry.second; |
| 2526 | // The main loop below skips these, see the comments there. All calls in |
| 2527 | // this entry share the same last node. |
| 2528 | if (Calls[0].StackNodes.back()->Recursive) |
| 2529 | continue; |
| 2530 | for (unsigned I = 0; I < Calls.size(); I++) { |
| 2531 | auto &CallCtxInfo = Calls[I]; |
| 2532 | auto &StackNodes = CallCtxInfo.StackNodes; |
| 2533 | // Calls are sorted longest sequence first, so no later call has |
| 2534 | // inlined frames either. |
| 2535 | if (StackNodes.size() < 2) |
| 2536 | break; |
| 2537 | // Calls matching an earlier call in the same function will share its |
| 2538 | // context node, so only give duplicates to the first. |
| 2539 | if (CallToMatchingCall.contains(CallCtxInfo.Call)) |
| 2540 | continue; |
| 2541 | auto Trimmed = computeTrimmedInlinedContextIds(StackNodes, Cache); |
| 2542 | if (Trimmed.empty()) |
| 2543 | continue; |
| 2544 | unsigned CreatorIdx = TrimmedInlinedDupCreatorNodes.size(); |
| 2545 | TrimmedInlinedDupCreatorNodes.push_back(StackNodes); |
| 2546 | CallCtxInfo.TrimmedInlinedDupIds = duplicateContextIds( |
| 2547 | StackSequenceContextIds: Trimmed, OldToNewContextIds, OnDuplicate: [&](uint32_t OldId, uint32_t NewId) { |
| 2548 | OrigToTrimmedInlinedDups[OldId].push_back(Elt: {NewId, CreatorIdx}); |
| 2549 | }); |
| 2550 | NumCallsAssignedTrimmedInlinedContexts++; |
| 2551 | NumTrimmedInlinedContextIds += Trimmed.size(); |
| 2552 | } |
| 2553 | } |
| 2554 | } |
| 2555 | |
| 2556 | // Adds to the given saved context ids, for a call with the given last stack |
| 2557 | // node, the trimmed inlined duplicates of any of those ids created for other |
| 2558 | // calls. The saved context ids are computed from the graph, which does not |
| 2559 | // include the duplicates until propagateDuplicateContextIds below adds them |
| 2560 | // to all of the same nodes and edges as the originals. A call assigned an |
| 2561 | // original should also be assigned its duplicates; otherwise, when that |
| 2562 | // call's new node takes the originals, the duplicates would be left behind |
| 2563 | // on the original stack nodes, which have no call to clone. |
| 2564 | // |
| 2565 | // TODO: The saved context ids may themselves be duplicates (created for a |
| 2566 | // call with the same stack ids in a different function), in which case the |
| 2567 | // original context id is not found here and the trimmed inlined duplicates |
| 2568 | // are not added. In the future, handle that by looking up the originals of |
| 2569 | // the saved ids. |
| 2570 | auto AddTrimmedInlinedDups = [&](DenseSet<uint32_t> &SavedContextIds, |
| 2571 | const ContextNode *LastNode) { |
| 2572 | SmallVector<uint32_t> DupIdsToAdd; |
| 2573 | // Collects the duplicates of one original id that this call should get. |
| 2574 | auto AddDups = [&](ArrayRef<std::pair<uint32_t, unsigned>> Dups) { |
| 2575 | for (auto [DupId, CreatorIdx] : Dups) { |
| 2576 | ArrayRef<ContextNode *> CreatorNodes = |
| 2577 | TrimmedInlinedDupCreatorNodes[CreatorIdx]; |
| 2578 | // Skip the duplicate if this call ends at any node of the creating |
| 2579 | // call's sequence. Such a call overlaps the stretch of the context |
| 2580 | // the creating call needs the duplicate on, and could claim the |
| 2581 | // duplicate first (the post order processes the creating call first |
| 2582 | // only when all edges along its sequence exist); being assigned the |
| 2583 | // original there is enough for it. Calls ending below the creating |
| 2584 | // call's first node only cover the callee side of the context, and |
| 2585 | // leave the duplicate on that first node for the creating call. |
| 2586 | if (is_contained(CreatorNodes, LastNode)) |
| 2587 | continue; |
| 2588 | DupIdsToAdd.push_back(Elt: DupId); |
| 2589 | } |
| 2590 | }; |
| 2591 | // Iterate over the smaller of the two. |
| 2592 | if (OrigToTrimmedInlinedDups.size() < SavedContextIds.size()) { |
| 2593 | for (auto &[OrigId, Dups] : OrigToTrimmedInlinedDups) |
| 2594 | if (SavedContextIds.contains(V: OrigId)) |
| 2595 | AddDups(Dups); |
| 2596 | } else { |
| 2597 | for (auto Id : SavedContextIds) { |
| 2598 | auto It = OrigToTrimmedInlinedDups.find(Val: Id); |
| 2599 | if (It != OrigToTrimmedInlinedDups.end()) |
| 2600 | AddDups(It->second); |
| 2601 | } |
| 2602 | } |
| 2603 | SavedContextIds.insert_range(R&: DupIdsToAdd); |
| 2604 | }; |
| 2605 | for (auto &It : StackIdToMatchingCalls) { |
| 2606 | auto &Calls = It.getSecond(); |
| 2607 | // Skip single calls with a single stack id. These don't need a new node. |
| 2608 | if (Calls.size() == 1) { |
| 2609 | auto &StackNodes = Calls[0].StackNodes; |
| 2610 | if (StackNodes.size() == 1) |
| 2611 | continue; |
| 2612 | } |
| 2613 | |
| 2614 | // Find the node for the last stack id, which should be the same |
| 2615 | // across all calls recorded for this id, and is the id for this |
| 2616 | // entry in the StackIdToMatchingCalls map. |
| 2617 | uint64_t LastId = It.getFirst(); |
| 2618 | // All calls in this entry share the same last node. |
| 2619 | ContextNode *LastNode = Calls[0].StackNodes.back(); |
| 2620 | // We should only have kept stack ids that had nodes. |
| 2621 | assert(LastNode); |
| 2622 | assert(LastNode == getNodeForStackId(LastId)); |
| 2623 | |
| 2624 | if (LastNode->Recursive) |
| 2625 | continue; |
| 2626 | |
| 2627 | // Initialize the context ids with the last node's. We will subsequently |
| 2628 | // refine the context ids by computing the intersection along all edges. |
| 2629 | DenseSet<uint32_t> LastNodeContextIds = LastNode->getContextIds(); |
| 2630 | assert(!LastNodeContextIds.empty()); |
| 2631 | |
| 2632 | #ifndef NDEBUG |
| 2633 | // Save the set of functions seen for a particular set of the same stack |
| 2634 | // ids. This is used to ensure that they have been correctly sorted to be |
| 2635 | // adjacent in the Calls list, since we rely on that to efficiently place |
| 2636 | // all such matching calls onto the same context node. |
| 2637 | DenseSet<const FuncTy *> MatchingIdsFuncSet; |
| 2638 | #endif |
| 2639 | |
| 2640 | for (unsigned I = 0; I < Calls.size(); I++) { |
| 2641 | auto &[Call, StackNodes, Func, SavedContextIds, TrimmedInlinedDupIds] = |
| 2642 | Calls[I]; |
| 2643 | assert(SavedContextIds.empty()); |
| 2644 | assert(LastNode == StackNodes.back()); |
| 2645 | |
| 2646 | #ifndef NDEBUG |
| 2647 | // If this call has a different set of ids than the last one, clear the |
| 2648 | // set used to ensure they are sorted properly. |
| 2649 | if (I > 0 && StackNodes != Calls[I - 1].StackNodes) |
| 2650 | MatchingIdsFuncSet.clear(); |
| 2651 | #endif |
| 2652 | |
| 2653 | // Skip the calls matching this one (recorded in CallToMatchingCall when |
| 2654 | // sorting above), which are adjacent to it, so that they get assigned to |
| 2655 | // the same context node. Also note whether a call in a different |
| 2656 | // function has the same stack ids. |
| 2657 | bool DuplicateContextIds = false; |
| 2658 | for (unsigned J = I + 1; J < Calls.size(); J++) { |
| 2659 | auto &CallCtxInfo = Calls[J]; |
| 2660 | if (CallCtxInfo.StackNodes != StackNodes) |
| 2661 | break; |
| 2662 | auto *NextFunc = CallCtxInfo.Func; |
| 2663 | if (NextFunc != Func) { |
| 2664 | // We have another Call with the same ids but that cannot share this |
| 2665 | // node, must duplicate ids for it. |
| 2666 | DuplicateContextIds = true; |
| 2667 | break; |
| 2668 | } |
| 2669 | assert(CallToMatchingCall.lookup(CallCtxInfo.Call).call() == Call); |
| 2670 | // Update I so that it gets incremented correctly to skip this call. |
| 2671 | I = J; |
| 2672 | } |
| 2673 | |
| 2674 | // First compute the context ids for this stack id sequence (the |
| 2675 | // intersection of the context ids of the corresponding nodes). |
| 2676 | // Start with the remaining saved ids for the last node. |
| 2677 | assert(!LastNodeContextIds.empty()); |
| 2678 | DenseSet<uint32_t> StackSequenceContextIds = LastNodeContextIds; |
| 2679 | |
| 2680 | ContextNode *PrevNode = LastNode; |
| 2681 | ContextNode *CurNode = LastNode; |
| 2682 | bool Skip = false; |
| 2683 | |
| 2684 | // Iterate backwards through the stack nodes, starting after the last |
| 2685 | // one in the list, which was handled once outside for all Calls. |
| 2686 | for (auto NodeIter = StackNodes.rbegin() + 1; |
| 2687 | NodeIter != StackNodes.rend(); NodeIter++) { |
| 2688 | CurNode = *NodeIter; |
| 2689 | // We should only have kept stack ids that had nodes. |
| 2690 | assert(CurNode); |
| 2691 | |
| 2692 | if (CurNode->Recursive) { |
| 2693 | Skip = true; |
| 2694 | break; |
| 2695 | } |
| 2696 | |
| 2697 | auto *Edge = CurNode->findEdgeFromCaller(PrevNode); |
| 2698 | // If there is no edge then the nodes belong to different MIB contexts, |
| 2699 | // and we should skip this inlined context sequence. For example, this |
| 2700 | // particular inlined context may include stack ids A->B, and we may |
| 2701 | // indeed have nodes for both A and B, but it is possible that they were |
| 2702 | // never profiled in sequence in a single MIB for any allocation (i.e. |
| 2703 | // we might have profiled an allocation that involves the callsite A, |
| 2704 | // but through a different one of its callee callsites, and we might |
| 2705 | // have profiled an allocation that involves callsite B, but reached |
| 2706 | // from a different caller callsite). |
| 2707 | if (!Edge) { |
| 2708 | Skip = true; |
| 2709 | break; |
| 2710 | } |
| 2711 | PrevNode = CurNode; |
| 2712 | |
| 2713 | // Update the context ids, which is the intersection of the ids along |
| 2714 | // all edges in the sequence. |
| 2715 | set_intersect(StackSequenceContextIds, Edge->getContextIds()); |
| 2716 | |
| 2717 | // If we now have no context ids for clone, skip this call. |
| 2718 | if (StackSequenceContextIds.empty()) { |
| 2719 | Skip = true; |
| 2720 | break; |
| 2721 | } |
| 2722 | } |
| 2723 | if (Skip) |
| 2724 | continue; |
| 2725 | |
| 2726 | // If some of this call's stack ids did not have corresponding nodes (due |
| 2727 | // to trimming), don't include any context ids for contexts that extend |
| 2728 | // beyond these nodes. Otherwise we would be matching part of unrelated / |
| 2729 | // not fully matching stack contexts. To do this, subtract any context ids |
| 2730 | // found in caller nodes of the last node found above. |
| 2731 | if (LastId != getLastStackId(Call)) { |
| 2732 | for (const auto &PE : LastNode->CallerEdges) { |
| 2733 | set_subtract(StackSequenceContextIds, PE->getContextIds()); |
| 2734 | if (StackSequenceContextIds.empty()) |
| 2735 | break; |
| 2736 | } |
| 2737 | // If we now have no context ids for clone, skip this call. |
| 2738 | if (StackSequenceContextIds.empty()) |
| 2739 | continue; |
| 2740 | } |
| 2741 | |
| 2742 | #ifndef NDEBUG |
| 2743 | // If the prior call had the same stack ids this set would not be empty. |
| 2744 | // Check if we already have a call that "matches" because it is located |
| 2745 | // in the same function. If the Calls list was sorted properly we should |
| 2746 | // not encounter this situation as all such entries should be adjacent |
| 2747 | // and processed in bulk further below. |
| 2748 | assert(!MatchingIdsFuncSet.contains(Func)); |
| 2749 | |
| 2750 | MatchingIdsFuncSet.insert(Func); |
| 2751 | #endif |
| 2752 | |
| 2753 | // If we don't have duplicate context ids, then we can assign all the |
| 2754 | // context ids computed for the original node sequence to this call. |
| 2755 | // If there are duplicate calls with the same stack ids then we synthesize |
| 2756 | // new context ids that are duplicates of the originals. These are |
| 2757 | // assigned to SavedContextIds, which is a reference into the map entry |
| 2758 | // for this call, allowing us to access these ids later on. |
| 2759 | OldToNewContextIds.reserve(NumEntries: OldToNewContextIds.size() + |
| 2760 | StackSequenceContextIds.size()); |
| 2761 | SavedContextIds = |
| 2762 | DuplicateContextIds |
| 2763 | ? duplicateContextIds(StackSequenceContextIds, OldToNewContextIds) |
| 2764 | : StackSequenceContextIds; |
| 2765 | assert(!SavedContextIds.empty()); |
| 2766 | if (!OrigToTrimmedInlinedDups.empty()) |
| 2767 | AddTrimmedInlinedDups(SavedContextIds, LastNode); |
| 2768 | |
| 2769 | if (!DuplicateContextIds) { |
| 2770 | // Update saved last node's context ids to remove those that are |
| 2771 | // assigned to other calls, so that it is ready for the next call at |
| 2772 | // this stack id. |
| 2773 | set_subtract(S1&: LastNodeContextIds, S2: StackSequenceContextIds); |
| 2774 | if (LastNodeContextIds.empty()) |
| 2775 | break; |
| 2776 | } |
| 2777 | } |
| 2778 | } |
| 2779 | |
| 2780 | // Propagate the duplicate context ids over the graph. |
| 2781 | propagateDuplicateContextIds(OldToNewContextIds); |
| 2782 | |
| 2783 | if (VerifyCCG) |
| 2784 | check(); |
| 2785 | |
| 2786 | // Now perform a post-order traversal over the graph, starting with the |
| 2787 | // allocation nodes, essentially processing nodes from callers to callees. |
| 2788 | // For any that contains an id in the map, update the graph to contain new |
| 2789 | // nodes representing any inlining at interior callsites. Note we move the |
| 2790 | // associated context ids over to the new nodes. |
| 2791 | DenseSet<const ContextNode *> Visited; |
| 2792 | DenseSet<uint32_t> ImportantContextIds(llvm::from_range, |
| 2793 | ImportantContextIdInfo.keys()); |
| 2794 | for (auto &Entry : AllocationCallToContextNodeMap) |
| 2795 | assignStackNodesPostOrder(Node: Entry.second, Visited, StackIdToMatchingCalls, |
| 2796 | CallToMatchingCall, ImportantContextIds); |
| 2797 | |
| 2798 | fixupImportantContexts(); |
| 2799 | |
| 2800 | if (VerifyCCG) |
| 2801 | check(); |
| 2802 | } |
| 2803 | |
| 2804 | uint64_t ModuleCallsiteContextGraph::getLastStackId(Instruction *Call) { |
| 2805 | CallStack<MDNode, MDNode::op_iterator> CallsiteContext( |
| 2806 | Call->getMetadata(KindID: LLVMContext::MD_callsite)); |
| 2807 | return CallsiteContext.back(); |
| 2808 | } |
| 2809 | |
| 2810 | uint64_t IndexCallsiteContextGraph::getLastStackId(IndexCall &Call) { |
| 2811 | assert(isa<CallsiteInfo *>(Call)); |
| 2812 | CallStack<CallsiteInfo, SmallVector<unsigned>::const_iterator> |
| 2813 | CallsiteContext(dyn_cast_if_present<CallsiteInfo *>(Val&: Call)); |
| 2814 | // Need to convert index into stack id. |
| 2815 | return Index.getStackIdAtIndex(Index: CallsiteContext.back()); |
| 2816 | } |
| 2817 | |
| 2818 | static std::string getMemProfFuncName(Twine Base, unsigned CloneNo) { |
| 2819 | // We use CloneNo == 0 to refer to the original version, which doesn't get |
| 2820 | // renamed with a suffix. |
| 2821 | if (!CloneNo) |
| 2822 | return Base.str(); |
| 2823 | return (Base + MemProfCloneSuffix + Twine(CloneNo)).str(); |
| 2824 | } |
| 2825 | |
| 2826 | // Return the clone number of the given function by extracting it from the |
| 2827 | // memprof suffix. Assumes the caller has already confirmed it is a memprof |
| 2828 | // clone. |
| 2829 | static unsigned getMemProfCloneNum(const Function &F) { |
| 2830 | assert(isMemProfClone(F)); |
| 2831 | auto Pos = F.getName().find_last_of(C: '.'); |
| 2832 | assert(Pos > 0); |
| 2833 | unsigned CloneNo; |
| 2834 | bool Err = F.getName().drop_front(N: Pos + 1).getAsInteger(Radix: 10, Result&: CloneNo); |
| 2835 | assert(!Err); |
| 2836 | (void)Err; |
| 2837 | return CloneNo; |
| 2838 | } |
| 2839 | |
| 2840 | std::string ModuleCallsiteContextGraph::getLabel(const Function *Func, |
| 2841 | const Instruction *Call, |
| 2842 | unsigned CloneNo) const { |
| 2843 | return (Twine(Call->getFunction()->getName()) + " -> " + |
| 2844 | cast<CallBase>(Val: Call)->getCalledFunction()->getName()) |
| 2845 | .str(); |
| 2846 | } |
| 2847 | |
| 2848 | std::string IndexCallsiteContextGraph::getLabel(const FunctionSummary *Func, |
| 2849 | const IndexCall &Call, |
| 2850 | unsigned CloneNo) const { |
| 2851 | auto VI = FSToVIMap.find(x: Func); |
| 2852 | assert(VI != FSToVIMap.end()); |
| 2853 | std::string CallerName = getMemProfFuncName(Base: VI->second.name(), CloneNo); |
| 2854 | if (isa<AllocInfo *>(Val: Call)) |
| 2855 | return CallerName + " -> alloc" ; |
| 2856 | else { |
| 2857 | auto *Callsite = dyn_cast_if_present<CallsiteInfo *>(Val: Call); |
| 2858 | return CallerName + " -> " + |
| 2859 | getMemProfFuncName(Base: Callsite->Callee.name(), |
| 2860 | CloneNo: Callsite->Clones[CloneNo]); |
| 2861 | } |
| 2862 | } |
| 2863 | |
| 2864 | std::string ModuleCallsiteContextGraph::getFuncName(const Function *Func, |
| 2865 | unsigned CloneNo) const { |
| 2866 | if (!Func) |
| 2867 | return "<unknown>" ; |
| 2868 | return getMemProfFuncName(Base: Func->getName(), CloneNo); |
| 2869 | } |
| 2870 | |
| 2871 | void ModuleCallsiteContextGraph::populateFunctionsWithCallers() { |
| 2872 | // Any use (direct call, address taken, alias, etc.) means the function is |
| 2873 | // referenced and may be reached by callers above the profiled context. |
| 2874 | for (const auto &F : Mod) |
| 2875 | if (!F.use_empty()) |
| 2876 | FunctionsWithCallers.insert(V: &F); |
| 2877 | } |
| 2878 | |
| 2879 | std::string IndexCallsiteContextGraph::getFuncName(const FunctionSummary *Func, |
| 2880 | unsigned CloneNo) const { |
| 2881 | if (!Func) |
| 2882 | return "<unknown>" ; |
| 2883 | auto VI = FSToVIMap.find(x: Func); |
| 2884 | if (VI != FSToVIMap.end()) { |
| 2885 | if (!VI->second.name().empty()) |
| 2886 | return getMemProfFuncName(Base: VI->second.name(), CloneNo); |
| 2887 | return getMemProfFuncName(Base: "GUID:" + std::to_string(val: VI->second.getGUID()), |
| 2888 | CloneNo); |
| 2889 | } |
| 2890 | return getMemProfFuncName(Base: "<unknown>" , CloneNo); |
| 2891 | } |
| 2892 | |
| 2893 | bool IndexCallsiteContextGraph::isFuncDead(const FunctionSummary *Func) const { |
| 2894 | return Index.withGlobalValueDeadStripping() && Func && !Func->isLive(); |
| 2895 | } |
| 2896 | |
| 2897 | void IndexCallsiteContextGraph::populateFunctionsWithCallers() { |
| 2898 | for (const auto &I : Index) { |
| 2899 | auto VI = Index.getValueInfo(R: I); |
| 2900 | for (const auto &S : VI.getSummaryList()) { |
| 2901 | if (!GlobalValue::isLocalLinkage(Linkage: S->linkage()) && |
| 2902 | !isPrevailing(VI.getGUID(), S.get())) |
| 2903 | continue; |
| 2904 | if (Index.withGlobalValueDeadStripping() && !S->isLive()) |
| 2905 | continue; |
| 2906 | auto AddReferenced = [&](ValueInfo RefVI) { |
| 2907 | for (const auto &RefS : RefVI.getSummaryList()) { |
| 2908 | const GlobalValueSummary *Base = RefS->getBaseObject(); |
| 2909 | if (const auto *RefFS = dyn_cast<FunctionSummary>(Val: Base)) |
| 2910 | FunctionsWithCallers.insert(V: RefFS); |
| 2911 | } |
| 2912 | }; |
| 2913 | // Address-taken references (e.g. from vtables or function pointers). |
| 2914 | for (const auto &RefVI : S->refs()) |
| 2915 | AddReferenced(RefVI); |
| 2916 | // Direct calls. |
| 2917 | if (const auto *FS = dyn_cast<FunctionSummary>(Val: S.get())) |
| 2918 | for (const auto &CallEdge : FS->calls()) |
| 2919 | AddReferenced(CallEdge.first); |
| 2920 | } |
| 2921 | } |
| 2922 | } |
| 2923 | |
| 2924 | std::vector<ModuleCallsiteContextGraph::ContextNode *> |
| 2925 | ModuleCallsiteContextGraph::getStackNodesForCall(Instruction *Call) { |
| 2926 | CallStack<MDNode, MDNode::op_iterator> CallsiteContext( |
| 2927 | Call->getMetadata(KindID: LLVMContext::MD_callsite)); |
| 2928 | return getStackNodesForCallsite<MDNode, MDNode::op_iterator>(CallsiteContext); |
| 2929 | } |
| 2930 | |
| 2931 | std::vector<IndexCallsiteContextGraph::ContextNode *> |
| 2932 | IndexCallsiteContextGraph::getStackNodesForCall(IndexCall &Call) { |
| 2933 | assert(isa<CallsiteInfo *>(Call)); |
| 2934 | CallStack<CallsiteInfo, SmallVector<unsigned>::const_iterator> |
| 2935 | CallsiteContext(dyn_cast_if_present<CallsiteInfo *>(Val&: Call)); |
| 2936 | return getStackNodesForCallsite<CallsiteInfo, |
| 2937 | SmallVector<unsigned>::const_iterator>( |
| 2938 | CallsiteContext); |
| 2939 | } |
| 2940 | |
| 2941 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 2942 | template <class NodeT, class IteratorT> |
| 2943 | std::vector< |
| 2944 | typename CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextNode *> |
| 2945 | CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::getStackNodesForCallsite( |
| 2946 | CallStack<NodeT, IteratorT> &CallsiteContext) { |
| 2947 | std::vector<ContextNode *> StackNodes; |
| 2948 | for (auto IdOrIndex : CallsiteContext) { |
| 2949 | auto StackId = getStackId(IdOrIndex); |
| 2950 | ContextNode *Node = getNodeForStackId(StackId); |
| 2951 | if (!Node) |
| 2952 | break; |
| 2953 | StackNodes.push_back(Node); |
| 2954 | } |
| 2955 | return StackNodes; |
| 2956 | } |
| 2957 | |
| 2958 | ModuleCallsiteContextGraph::( |
| 2959 | Module &M, |
| 2960 | llvm::function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter) |
| 2961 | : Mod(M), OREGetter(OREGetter) { |
| 2962 | // Map for keeping track of the largest cold contexts up to the number given |
| 2963 | // by MemProfTopNImportant. Must be a std::map (not DenseMap) because keys |
| 2964 | // must be sorted. |
| 2965 | std::map<uint64_t, uint32_t> TotalSizeToContextIdTopNCold; |
| 2966 | for (auto &F : M) { |
| 2967 | std::vector<CallInfo> CallsWithMetadata; |
| 2968 | for (auto &BB : F) { |
| 2969 | for (auto &I : BB) { |
| 2970 | if (!isa<CallBase>(Val: I)) |
| 2971 | continue; |
| 2972 | if (auto *MemProfMD = I.getMetadata(KindID: LLVMContext::MD_memprof)) { |
| 2973 | CallsWithMetadata.push_back(x: &I); |
| 2974 | auto *AllocNode = addAllocNode(Call: &I, F: &F); |
| 2975 | auto *CallsiteMD = I.getMetadata(KindID: LLVMContext::MD_callsite); |
| 2976 | assert(CallsiteMD); |
| 2977 | CallStack<MDNode, MDNode::op_iterator> CallsiteContext(CallsiteMD); |
| 2978 | // Add all of the MIBs and their stack nodes. |
| 2979 | for (auto &MDOp : MemProfMD->operands()) { |
| 2980 | auto *MIBMD = cast<const MDNode>(Val: MDOp); |
| 2981 | std::vector<ContextTotalSize> ContextSizeInfo; |
| 2982 | // Collect the context size information if it exists. |
| 2983 | if (MIBMD->getNumOperands() > 2) { |
| 2984 | for (unsigned I = 2; I < MIBMD->getNumOperands(); I++) { |
| 2985 | MDNode *ContextSizePair = |
| 2986 | dyn_cast<MDNode>(Val: MIBMD->getOperand(I)); |
| 2987 | assert(ContextSizePair->getNumOperands() == 2); |
| 2988 | uint64_t FullStackId = mdconst::dyn_extract<ConstantInt>( |
| 2989 | MD: ContextSizePair->getOperand(I: 0)) |
| 2990 | ->getZExtValue(); |
| 2991 | uint64_t TotalSize = mdconst::dyn_extract<ConstantInt>( |
| 2992 | MD: ContextSizePair->getOperand(I: 1)) |
| 2993 | ->getZExtValue(); |
| 2994 | ContextSizeInfo.push_back(x: {.FullStackId: FullStackId, .TotalSize: TotalSize}); |
| 2995 | } |
| 2996 | } |
| 2997 | MDNode *StackNode = getMIBStackNode(MIB: MIBMD); |
| 2998 | assert(StackNode); |
| 2999 | CallStack<MDNode, MDNode::op_iterator> StackContext(StackNode); |
| 3000 | addStackNodesForMIB<MDNode, MDNode::op_iterator>( |
| 3001 | AllocNode, StackContext, CallsiteContext, |
| 3002 | AllocType: getMIBAllocType(MIB: MIBMD), ContextSizeInfo, |
| 3003 | TotalSizeToContextIdTopNCold); |
| 3004 | } |
| 3005 | // If exporting the graph to dot and an allocation id of interest was |
| 3006 | // specified, record all the context ids for this allocation node. |
| 3007 | if (ExportToDot && AllocNode->OrigStackOrAllocId == AllocIdForDot) |
| 3008 | DotAllocContextIds = AllocNode->getContextIds(); |
| 3009 | assert(AllocNode->AllocTypes != (uint8_t)AllocationType::None); |
| 3010 | // Memprof and callsite metadata on memory allocations no longer |
| 3011 | // needed. |
| 3012 | I.setMetadata(KindID: LLVMContext::MD_memprof, Node: nullptr); |
| 3013 | I.setMetadata(KindID: LLVMContext::MD_callsite, Node: nullptr); |
| 3014 | } |
| 3015 | // For callsite metadata, add to list for this function for later use. |
| 3016 | else if (I.getMetadata(KindID: LLVMContext::MD_callsite)) { |
| 3017 | CallsWithMetadata.push_back(x: &I); |
| 3018 | } |
| 3019 | } |
| 3020 | } |
| 3021 | if (!CallsWithMetadata.empty()) |
| 3022 | FuncToCallsWithMetadata[&F] = CallsWithMetadata; |
| 3023 | } |
| 3024 | |
| 3025 | if (DumpCCG) { |
| 3026 | dbgs() << "CCG before updating call stack chains:\n" ; |
| 3027 | dbgs() << *this; |
| 3028 | } |
| 3029 | |
| 3030 | if (ExportToDot) |
| 3031 | exportToDot(Label: "prestackupdate" ); |
| 3032 | |
| 3033 | updateStackNodes(); |
| 3034 | |
| 3035 | if (ExportToDot) |
| 3036 | exportToDot(Label: "poststackupdate" ); |
| 3037 | |
| 3038 | handleCallsitesWithMultipleTargets(); |
| 3039 | |
| 3040 | markBackedges(); |
| 3041 | |
| 3042 | // Strip off remaining callsite metadata, no longer needed. |
| 3043 | for (auto &FuncEntry : FuncToCallsWithMetadata) |
| 3044 | for (auto &Call : FuncEntry.second) |
| 3045 | Call.call()->setMetadata(KindID: LLVMContext::MD_callsite, Node: nullptr); |
| 3046 | } |
| 3047 | |
| 3048 | // Finds the set of GUIDs for weak aliasees that are prevailing in different |
| 3049 | // modules than any of their aliases. We need to handle these specially. |
| 3050 | DenseSet<GlobalValue::GUID> |
| 3051 | IndexCallsiteContextGraph::findAliaseeGUIDsPrevailingInDifferentModule() { |
| 3052 | DenseSet<GlobalValue::GUID> AliaseeGUIDs; |
| 3053 | for (auto &I : Index) { |
| 3054 | auto VI = Index.getValueInfo(R: I); |
| 3055 | for (auto &S : VI.getSummaryList()) { |
| 3056 | // We only care about aliases to functions. |
| 3057 | auto *AS = dyn_cast<AliasSummary>(Val: S.get()); |
| 3058 | if (!AS) |
| 3059 | continue; |
| 3060 | auto *AliaseeSummary = &AS->getAliasee(); |
| 3061 | auto *AliaseeFS = dyn_cast<FunctionSummary>(Val: AliaseeSummary); |
| 3062 | if (!AliaseeFS) |
| 3063 | continue; |
| 3064 | // Skip this summary if it is not for the prevailing symbol for this GUID. |
| 3065 | // The linker doesn't resolve local linkage values so don't check whether |
| 3066 | // those are prevailing. |
| 3067 | if (!GlobalValue::isLocalLinkage(Linkage: S->linkage()) && |
| 3068 | !isPrevailing(VI.getGUID(), S.get())) |
| 3069 | continue; |
| 3070 | // Prevailing aliasee could be in a different module only if it is weak. |
| 3071 | if (!GlobalValue::isWeakForLinker(Linkage: AliaseeSummary->linkage())) |
| 3072 | continue; |
| 3073 | auto AliaseeGUID = AS->getAliaseeGUID(); |
| 3074 | // If the aliasee copy in this module is not prevailing, record it. |
| 3075 | if (!isPrevailing(AliaseeGUID, AliaseeSummary)) |
| 3076 | AliaseeGUIDs.insert(V: AliaseeGUID); |
| 3077 | } |
| 3078 | } |
| 3079 | AliaseesPrevailingInDiffModuleFromAlias += AliaseeGUIDs.size(); |
| 3080 | return AliaseeGUIDs; |
| 3081 | } |
| 3082 | |
| 3083 | IndexCallsiteContextGraph::IndexCallsiteContextGraph( |
| 3084 | ModuleSummaryIndex &Index, |
| 3085 | llvm::function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> |
| 3086 | isPrevailing) |
| 3087 | : Index(Index), isPrevailing(isPrevailing) { |
| 3088 | // Since we use the aliasee summary info to create the necessary clones for |
| 3089 | // its aliases, conservatively skip recording the aliasee function's callsites |
| 3090 | // in the CCG for any that are prevailing in a different module than one of |
| 3091 | // its aliases. We could record the necessary information to do this in the |
| 3092 | // summary, but this case should not be common. |
| 3093 | DenseSet<GlobalValue::GUID> GUIDsToSkip = |
| 3094 | findAliaseeGUIDsPrevailingInDifferentModule(); |
| 3095 | // Map for keeping track of the largest cold contexts up to the number given |
| 3096 | // by MemProfTopNImportant. Must be a std::map (not DenseMap) because keys |
| 3097 | // must be sorted. |
| 3098 | std::map<uint64_t, uint32_t> TotalSizeToContextIdTopNCold; |
| 3099 | // Sort by GUID for deterministic graph construction order. |
| 3100 | // TODO: This sort has a measurable cost on the thin link when memprof is |
| 3101 | // enabled. Investigate gating it behind an option that is only enabled for |
| 3102 | // tests that check internal state. |
| 3103 | for (const auto &I : Index.sortedGlobalValueSummariesRange()) { |
| 3104 | auto VI = Index.getValueInfo(R: I); |
| 3105 | if (GUIDsToSkip.contains(V: VI.getGUID())) |
| 3106 | continue; |
| 3107 | for (auto &S : VI.getSummaryList()) { |
| 3108 | // We should only add the prevailing nodes. Otherwise we may try to clone |
| 3109 | // in a weak copy that won't be linked (and may be different than the |
| 3110 | // prevailing version). |
| 3111 | // We only keep the memprof summary on the prevailing copy now when |
| 3112 | // building the combined index, as a space optimization, however don't |
| 3113 | // rely on this optimization. The linker doesn't resolve local linkage |
| 3114 | // values so don't check whether those are prevailing. |
| 3115 | if (!GlobalValue::isLocalLinkage(Linkage: S->linkage()) && |
| 3116 | !isPrevailing(VI.getGUID(), S.get())) |
| 3117 | continue; |
| 3118 | auto *FS = dyn_cast<FunctionSummary>(Val: S.get()); |
| 3119 | if (!FS) |
| 3120 | continue; |
| 3121 | std::vector<CallInfo> CallsWithMetadata; |
| 3122 | if (!FS->allocs().empty()) { |
| 3123 | for (auto &AN : FS->mutableAllocs()) { |
| 3124 | // This can happen because of recursion elimination handling that |
| 3125 | // currently exists in ModuleSummaryAnalysis. Skip these for now. |
| 3126 | // We still added them to the summary because we need to be able to |
| 3127 | // correlate properly in applyImport in the backends. |
| 3128 | if (AN.MIBs.empty()) |
| 3129 | continue; |
| 3130 | IndexCall AllocCall(&AN); |
| 3131 | CallsWithMetadata.push_back(x: AllocCall); |
| 3132 | auto *AllocNode = addAllocNode(Call: AllocCall, F: FS); |
| 3133 | // Pass an empty CallStack to the CallsiteContext (second) |
| 3134 | // parameter, since for ThinLTO we already collapsed out the inlined |
| 3135 | // stack ids on the allocation call during ModuleSummaryAnalysis. |
| 3136 | CallStack<MIBInfo, SmallVector<unsigned>::const_iterator> |
| 3137 | EmptyContext; |
| 3138 | unsigned I = 0; |
| 3139 | assert(!metadataMayIncludeContextSizeInfo() || |
| 3140 | AN.ContextSizeInfos.size() == AN.MIBs.size()); |
| 3141 | // Now add all of the MIBs and their stack nodes. |
| 3142 | for (auto &MIB : AN.MIBs) { |
| 3143 | CallStack<MIBInfo, SmallVector<unsigned>::const_iterator> |
| 3144 | StackContext(&MIB); |
| 3145 | std::vector<ContextTotalSize> ContextSizeInfo; |
| 3146 | if (!AN.ContextSizeInfos.empty()) { |
| 3147 | for (auto [FullStackId, TotalSize] : AN.ContextSizeInfos[I]) |
| 3148 | ContextSizeInfo.push_back(x: {.FullStackId: FullStackId, .TotalSize: TotalSize}); |
| 3149 | } |
| 3150 | addStackNodesForMIB<MIBInfo, SmallVector<unsigned>::const_iterator>( |
| 3151 | AllocNode, StackContext, CallsiteContext&: EmptyContext, AllocType: MIB.AllocType, |
| 3152 | ContextSizeInfo, TotalSizeToContextIdTopNCold); |
| 3153 | I++; |
| 3154 | } |
| 3155 | // If exporting the graph to dot and an allocation id of interest was |
| 3156 | // specified, record all the context ids for this allocation node. |
| 3157 | if (ExportToDot && AllocNode->OrigStackOrAllocId == AllocIdForDot) |
| 3158 | DotAllocContextIds = AllocNode->getContextIds(); |
| 3159 | assert(AllocNode->AllocTypes != (uint8_t)AllocationType::None); |
| 3160 | // Initialize version 0 on the summary alloc node to the current alloc |
| 3161 | // type, unless it has both types in which case make it default, so |
| 3162 | // that in the case where we aren't able to clone the original version |
| 3163 | // always ends up with the default allocation behavior. |
| 3164 | AN.Versions[0] = (uint8_t)allocTypeToUse(AllocTypes: AllocNode->AllocTypes); |
| 3165 | } |
| 3166 | } |
| 3167 | // For callsite metadata, add to list for this function for later use. |
| 3168 | if (!FS->callsites().empty()) |
| 3169 | for (auto &SN : FS->mutableCallsites()) { |
| 3170 | IndexCall StackNodeCall(&SN); |
| 3171 | CallsWithMetadata.push_back(x: StackNodeCall); |
| 3172 | } |
| 3173 | |
| 3174 | if (!CallsWithMetadata.empty()) |
| 3175 | FuncToCallsWithMetadata[FS] = CallsWithMetadata; |
| 3176 | |
| 3177 | if (!FS->allocs().empty() || !FS->callsites().empty()) |
| 3178 | FSToVIMap[FS] = VI; |
| 3179 | } |
| 3180 | } |
| 3181 | |
| 3182 | if (DumpCCG) { |
| 3183 | dbgs() << "CCG before updating call stack chains:\n" ; |
| 3184 | dbgs() << *this; |
| 3185 | } |
| 3186 | |
| 3187 | if (ExportToDot) |
| 3188 | exportToDot(Label: "prestackupdate" ); |
| 3189 | |
| 3190 | updateStackNodes(); |
| 3191 | |
| 3192 | if (ExportToDot) |
| 3193 | exportToDot(Label: "poststackupdate" ); |
| 3194 | |
| 3195 | handleCallsitesWithMultipleTargets(); |
| 3196 | |
| 3197 | markBackedges(); |
| 3198 | } |
| 3199 | |
| 3200 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 3201 | void CallsiteContextGraph<DerivedCCG, FuncTy, |
| 3202 | CallTy>::handleCallsitesWithMultipleTargets() { |
| 3203 | // Look for and workaround callsites that call multiple functions. |
| 3204 | // This can happen for indirect calls, which needs better handling, and in |
| 3205 | // more rare cases (e.g. macro expansion). |
| 3206 | // TODO: To fix this for indirect calls we will want to perform speculative |
| 3207 | // devirtualization using either the normal PGO info with ICP, or using the |
| 3208 | // information in the profiled MemProf contexts. We can do this prior to |
| 3209 | // this transformation for regular LTO, and for ThinLTO we can simulate that |
| 3210 | // effect in the summary and perform the actual speculative devirtualization |
| 3211 | // while cloning in the ThinLTO backend. |
| 3212 | |
| 3213 | // Keep track of the new nodes synthesized for discovered tail calls missing |
| 3214 | // from the profiled contexts. |
| 3215 | MapVector<CallInfo, ContextNode *> TailCallToContextNodeMap; |
| 3216 | |
| 3217 | std::vector<std::pair<CallInfo, ContextNode *>> NewCallToNode; |
| 3218 | for (auto &Entry : NonAllocationCallToContextNodeMap) { |
| 3219 | auto *Node = Entry.second; |
| 3220 | assert(Node->Clones.empty()); |
| 3221 | // Check all node callees and see if in the same function. |
| 3222 | // We need to check all of the calls recorded in this Node, because in some |
| 3223 | // cases we may have had multiple calls with the same debug info calling |
| 3224 | // different callees. This can happen, for example, when an object is |
| 3225 | // constructed in the paramter list - the destructor call of the object has |
| 3226 | // the same debug info (line/col) as the call the object was passed to. |
| 3227 | // Here we will prune any that don't match all callee nodes. |
| 3228 | std::vector<CallInfo> AllCalls; |
| 3229 | AllCalls.reserve(Node->MatchingCalls.size() + 1); |
| 3230 | AllCalls.push_back(Node->Call); |
| 3231 | llvm::append_range(AllCalls, Node->MatchingCalls); |
| 3232 | |
| 3233 | // First see if we can partition the calls by callee function, creating new |
| 3234 | // nodes to host each set of calls calling the same callees. This is |
| 3235 | // necessary for support indirect calls with ThinLTO, for which we |
| 3236 | // synthesized CallsiteInfo records for each target. They will all have the |
| 3237 | // same callsite stack ids and would be sharing a context node at this |
| 3238 | // point. We need to perform separate cloning for each, which will be |
| 3239 | // applied along with speculative devirtualization in the ThinLTO backends |
| 3240 | // as needed. Note this does not currently support looking through tail |
| 3241 | // calls, it is unclear if we need that for indirect call targets. |
| 3242 | // First partition calls by callee func. Map indexed by func, value is |
| 3243 | // struct with list of matching calls, assigned node. |
| 3244 | if (partitionCallsByCallee(Node, AllCalls, NewCallToNode)) |
| 3245 | continue; |
| 3246 | |
| 3247 | auto It = AllCalls.begin(); |
| 3248 | // Iterate through the calls until we find the first that matches. |
| 3249 | for (; It != AllCalls.end(); ++It) { |
| 3250 | auto ThisCall = *It; |
| 3251 | bool Match = true; |
| 3252 | for (auto EI = Node->CalleeEdges.begin(); EI != Node->CalleeEdges.end(); |
| 3253 | ++EI) { |
| 3254 | auto Edge = *EI; |
| 3255 | if (!Edge->Callee->hasCall()) |
| 3256 | continue; |
| 3257 | assert(NodeToCallingFunc.count(Edge->Callee)); |
| 3258 | // Check if the called function matches that of the callee node. |
| 3259 | if (!calleesMatch(Call: ThisCall.call(), EI, TailCallToContextNodeMap)) { |
| 3260 | Match = false; |
| 3261 | break; |
| 3262 | } |
| 3263 | } |
| 3264 | // Found a call that matches the callee nodes, we can quit now. |
| 3265 | if (Match) { |
| 3266 | // If the first match is not the primary call on the Node, update it |
| 3267 | // now. We will update the list of matching calls further below. |
| 3268 | if (Node->Call != ThisCall) { |
| 3269 | Node->setCall(ThisCall); |
| 3270 | // We need to update the NonAllocationCallToContextNodeMap, but don't |
| 3271 | // want to do this during iteration over that map, so save the calls |
| 3272 | // that need updated entries. |
| 3273 | NewCallToNode.push_back({ThisCall, Node}); |
| 3274 | } |
| 3275 | break; |
| 3276 | } |
| 3277 | } |
| 3278 | // We will update this list below (or leave it cleared if there was no |
| 3279 | // match found above). |
| 3280 | Node->MatchingCalls.clear(); |
| 3281 | // If we hit the end of the AllCalls vector, no call matching the callee |
| 3282 | // nodes was found, clear the call information in the node. |
| 3283 | if (It == AllCalls.end()) { |
| 3284 | RemovedEdgesWithMismatchedCallees++; |
| 3285 | // Work around by setting Node to have a null call, so it gets |
| 3286 | // skipped during cloning. Otherwise assignFunctions will assert |
| 3287 | // because its data structures are not designed to handle this case. |
| 3288 | Node->setCall(CallInfo()); |
| 3289 | continue; |
| 3290 | } |
| 3291 | // Now add back any matching calls that call the same function as the |
| 3292 | // matching primary call on Node. |
| 3293 | for (++It; It != AllCalls.end(); ++It) { |
| 3294 | auto ThisCall = *It; |
| 3295 | if (!sameCallee(Call1: Node->Call.call(), Call2: ThisCall.call())) |
| 3296 | continue; |
| 3297 | Node->MatchingCalls.push_back(ThisCall); |
| 3298 | } |
| 3299 | } |
| 3300 | |
| 3301 | // Remove all mismatched nodes identified in the above loop from the node map |
| 3302 | // (checking whether they have a null call which is set above). For a |
| 3303 | // MapVector like NonAllocationCallToContextNodeMap it is much more efficient |
| 3304 | // to do the removal via remove_if than by individually erasing entries above. |
| 3305 | // Also remove any entries if we updated the node's primary call above. |
| 3306 | NonAllocationCallToContextNodeMap.remove_if([](const auto &it) { |
| 3307 | return !it.second->hasCall() || it.second->Call != it.first; |
| 3308 | }); |
| 3309 | |
| 3310 | // Add entries for any new primary calls recorded above. |
| 3311 | for (auto &[Call, Node] : NewCallToNode) |
| 3312 | NonAllocationCallToContextNodeMap[Call] = Node; |
| 3313 | |
| 3314 | // Add the new nodes after the above loop so that the iteration is not |
| 3315 | // invalidated. |
| 3316 | for (auto &[Call, Node] : TailCallToContextNodeMap) |
| 3317 | NonAllocationCallToContextNodeMap[Call] = Node; |
| 3318 | } |
| 3319 | |
| 3320 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 3321 | bool CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::partitionCallsByCallee( |
| 3322 | ContextNode *Node, ArrayRef<CallInfo> AllCalls, |
| 3323 | std::vector<std::pair<CallInfo, ContextNode *>> &NewCallToNode) { |
| 3324 | // Struct to keep track of all the calls having the same callee function, |
| 3325 | // and the node we eventually assign to them. Eventually we will record the |
| 3326 | // context node assigned to this group of calls. |
| 3327 | struct CallsWithSameCallee { |
| 3328 | std::vector<CallInfo> Calls; |
| 3329 | ContextNode *Node = nullptr; |
| 3330 | }; |
| 3331 | |
| 3332 | // First partition calls by callee function. Build map from each function |
| 3333 | // to the list of matching calls. |
| 3334 | DenseMap<const FuncTy *, CallsWithSameCallee> CalleeFuncToCallInfo; |
| 3335 | for (auto ThisCall : AllCalls) { |
| 3336 | auto *F = getCalleeFunc(Call: ThisCall.call()); |
| 3337 | if (F) |
| 3338 | CalleeFuncToCallInfo[F].Calls.push_back(ThisCall); |
| 3339 | } |
| 3340 | |
| 3341 | // Next, walk through all callee edges. For each callee node, get its |
| 3342 | // containing function and see if it was recorded in the above map (meaning we |
| 3343 | // have at least one matching call). Build another map from each callee node |
| 3344 | // with a matching call to the structure instance created above containing all |
| 3345 | // the calls. |
| 3346 | DenseMap<ContextNode *, CallsWithSameCallee *> CalleeNodeToCallInfo; |
| 3347 | for (const auto &Edge : Node->CalleeEdges) { |
| 3348 | if (!Edge->Callee->hasCall()) |
| 3349 | continue; |
| 3350 | const FuncTy *ProfiledCalleeFunc = NodeToCallingFunc[Edge->Callee]; |
| 3351 | if (CalleeFuncToCallInfo.contains(ProfiledCalleeFunc)) |
| 3352 | CalleeNodeToCallInfo[Edge->Callee] = |
| 3353 | &CalleeFuncToCallInfo[ProfiledCalleeFunc]; |
| 3354 | } |
| 3355 | |
| 3356 | // If there are entries in the second map, then there were no matching |
| 3357 | // calls/callees, nothing to do here. Return so we can go to the handling that |
| 3358 | // looks through tail calls. |
| 3359 | if (CalleeNodeToCallInfo.empty()) |
| 3360 | return false; |
| 3361 | |
| 3362 | // Walk through all callee edges again. Any and all callee edges that didn't |
| 3363 | // match any calls (callee not in the CalleeNodeToCallInfo map) are moved to a |
| 3364 | // new caller node (UnmatchedCalleesNode) which gets a null call so that it is |
| 3365 | // ignored during cloning. If it is in the map, then we use the node recorded |
| 3366 | // in that entry (creating it if needed), and move the callee edge to it. |
| 3367 | // The first callee will use the original node instead of creating a new one. |
| 3368 | // Note that any of the original calls on this node (in AllCalls) that didn't |
| 3369 | // have a callee function automatically get dropped from the node as part of |
| 3370 | // this process. |
| 3371 | ContextNode *UnmatchedCalleesNode = nullptr; |
| 3372 | // Track whether we already assigned original node to a callee. |
| 3373 | bool UsedOrigNode = false; |
| 3374 | assert(NodeToCallingFunc[Node]); |
| 3375 | // Iterate over a copy of Node's callee edges, since we may need to remove |
| 3376 | // edges in moveCalleeEdgeToNewCaller, and this simplifies the handling and |
| 3377 | // makes it less error-prone. |
| 3378 | auto CalleeEdges = Node->CalleeEdges; |
| 3379 | for (auto &Edge : CalleeEdges) { |
| 3380 | if (!Edge->Callee->hasCall()) |
| 3381 | continue; |
| 3382 | |
| 3383 | // Will be updated below to point to whatever (caller) node this callee edge |
| 3384 | // should be moved to. |
| 3385 | ContextNode *CallerNodeToUse = nullptr; |
| 3386 | |
| 3387 | // Handle the case where there were no matching calls first. Move this |
| 3388 | // callee edge to the UnmatchedCalleesNode, creating it if needed. |
| 3389 | if (!CalleeNodeToCallInfo.contains(Edge->Callee)) { |
| 3390 | if (!UnmatchedCalleesNode) |
| 3391 | UnmatchedCalleesNode = |
| 3392 | createNewNode(/*IsAllocation=*/false, F: NodeToCallingFunc[Node]); |
| 3393 | CallerNodeToUse = UnmatchedCalleesNode; |
| 3394 | } else { |
| 3395 | // Look up the information recorded for this callee node, and use the |
| 3396 | // recorded caller node (creating it if needed). |
| 3397 | auto *Info = CalleeNodeToCallInfo[Edge->Callee]; |
| 3398 | if (!Info->Node) { |
| 3399 | // If we haven't assigned any callees to the original node use it. |
| 3400 | if (!UsedOrigNode) { |
| 3401 | Info->Node = Node; |
| 3402 | // Clear the set of matching calls which will be updated below. |
| 3403 | Node->MatchingCalls.clear(); |
| 3404 | UsedOrigNode = true; |
| 3405 | } else |
| 3406 | Info->Node = |
| 3407 | createNewNode(/*IsAllocation=*/false, F: NodeToCallingFunc[Node]); |
| 3408 | assert(!Info->Calls.empty()); |
| 3409 | // The first call becomes the primary call for this caller node, and the |
| 3410 | // rest go in the matching calls list. |
| 3411 | Info->Node->setCall(Info->Calls.front()); |
| 3412 | llvm::append_range(Info->Node->MatchingCalls, |
| 3413 | llvm::drop_begin(Info->Calls)); |
| 3414 | // Save the primary call to node correspondence so that we can update |
| 3415 | // the NonAllocationCallToContextNodeMap, which is being iterated in the |
| 3416 | // caller of this function. |
| 3417 | NewCallToNode.push_back({Info->Node->Call, Info->Node}); |
| 3418 | } |
| 3419 | CallerNodeToUse = Info->Node; |
| 3420 | } |
| 3421 | |
| 3422 | // Don't need to move edge if we are using the original node; |
| 3423 | if (CallerNodeToUse == Node) |
| 3424 | continue; |
| 3425 | |
| 3426 | moveCalleeEdgeToNewCaller(Edge, NewCaller: CallerNodeToUse); |
| 3427 | } |
| 3428 | // Now that we are done moving edges, clean up any caller edges that ended |
| 3429 | // up with no type or context ids. During moveCalleeEdgeToNewCaller all |
| 3430 | // caller edges from Node are replicated onto the new callers, and it |
| 3431 | // simplifies the handling to leave them until we have moved all |
| 3432 | // edges/context ids. |
| 3433 | for (auto &I : CalleeNodeToCallInfo) |
| 3434 | removeNoneTypeCallerEdges(Node: I.second->Node); |
| 3435 | if (UnmatchedCalleesNode) |
| 3436 | removeNoneTypeCallerEdges(Node: UnmatchedCalleesNode); |
| 3437 | removeNoneTypeCallerEdges(Node); |
| 3438 | |
| 3439 | return true; |
| 3440 | } |
| 3441 | |
| 3442 | uint64_t ModuleCallsiteContextGraph::getStackId(uint64_t IdOrIndex) const { |
| 3443 | // In the Module (IR) case this is already the Id. |
| 3444 | return IdOrIndex; |
| 3445 | } |
| 3446 | |
| 3447 | uint64_t IndexCallsiteContextGraph::getStackId(uint64_t IdOrIndex) const { |
| 3448 | // In the Index case this is an index into the stack id list in the summary |
| 3449 | // index, convert it to an Id. |
| 3450 | return Index.getStackIdAtIndex(Index: IdOrIndex); |
| 3451 | } |
| 3452 | |
| 3453 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 3454 | bool CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::calleesMatch( |
| 3455 | CallTy Call, EdgeIter &EI, |
| 3456 | MapVector<CallInfo, ContextNode *> &TailCallToContextNodeMap) { |
| 3457 | auto Edge = *EI; |
| 3458 | const FuncTy *ProfiledCalleeFunc = NodeToCallingFunc[Edge->Callee]; |
| 3459 | const FuncTy *CallerFunc = NodeToCallingFunc[Edge->Caller]; |
| 3460 | // Will be populated in order of callee to caller if we find a chain of tail |
| 3461 | // calls between the profiled caller and callee. |
| 3462 | std::vector<std::pair<CallTy, FuncTy *>> FoundCalleeChain; |
| 3463 | if (!calleeMatchesFunc(Call, Func: ProfiledCalleeFunc, CallerFunc, |
| 3464 | FoundCalleeChain)) |
| 3465 | return false; |
| 3466 | |
| 3467 | // The usual case where the profiled callee matches that of the IR/summary. |
| 3468 | if (FoundCalleeChain.empty()) |
| 3469 | return true; |
| 3470 | |
| 3471 | auto AddEdge = [Edge, &EI](ContextNode *Caller, ContextNode *Callee) { |
| 3472 | auto *CurEdge = Callee->findEdgeFromCaller(Caller); |
| 3473 | // If there is already an edge between these nodes, simply update it and |
| 3474 | // return. |
| 3475 | if (CurEdge) { |
| 3476 | CurEdge->ContextIds.insert_range(Edge->ContextIds); |
| 3477 | CurEdge->AllocTypes |= Edge->AllocTypes; |
| 3478 | return; |
| 3479 | } |
| 3480 | // Otherwise, create a new edge and insert it into the caller and callee |
| 3481 | // lists. |
| 3482 | auto NewEdge = std::make_shared<ContextEdge>( |
| 3483 | Callee, Caller, Edge->AllocTypes, Edge->ContextIds); |
| 3484 | Callee->CallerEdges.push_back(NewEdge); |
| 3485 | if (Caller == Edge->Caller) { |
| 3486 | // If we are inserting the new edge into the current edge's caller, insert |
| 3487 | // the new edge before the current iterator position, and then increment |
| 3488 | // back to the current edge. |
| 3489 | EI = Caller->CalleeEdges.insert(EI, NewEdge); |
| 3490 | ++EI; |
| 3491 | assert(*EI == Edge && |
| 3492 | "Iterator position not restored after insert and increment" ); |
| 3493 | } else |
| 3494 | Caller->CalleeEdges.push_back(NewEdge); |
| 3495 | }; |
| 3496 | |
| 3497 | // Create new nodes for each found callee and connect in between the profiled |
| 3498 | // caller and callee. |
| 3499 | auto *CurCalleeNode = Edge->Callee; |
| 3500 | for (auto &[NewCall, Func] : FoundCalleeChain) { |
| 3501 | ContextNode *NewNode = nullptr; |
| 3502 | // First check if we have already synthesized a node for this tail call. |
| 3503 | if (TailCallToContextNodeMap.count(NewCall)) { |
| 3504 | NewNode = TailCallToContextNodeMap[NewCall]; |
| 3505 | NewNode->AllocTypes |= Edge->AllocTypes; |
| 3506 | } else { |
| 3507 | FuncToCallsWithMetadata[Func].push_back({NewCall}); |
| 3508 | // Create Node and record node info. |
| 3509 | NewNode = createNewNode(/*IsAllocation=*/false, F: Func, C: NewCall); |
| 3510 | TailCallToContextNodeMap[NewCall] = NewNode; |
| 3511 | NewNode->AllocTypes = Edge->AllocTypes; |
| 3512 | } |
| 3513 | |
| 3514 | // Hook up node to its callee node |
| 3515 | AddEdge(NewNode, CurCalleeNode); |
| 3516 | |
| 3517 | CurCalleeNode = NewNode; |
| 3518 | } |
| 3519 | |
| 3520 | // Hook up edge's original caller to new callee node. |
| 3521 | AddEdge(Edge->Caller, CurCalleeNode); |
| 3522 | |
| 3523 | #ifndef NDEBUG |
| 3524 | // Save this because Edge's fields get cleared below when removed. |
| 3525 | auto *Caller = Edge->Caller; |
| 3526 | #endif |
| 3527 | |
| 3528 | // Remove old edge |
| 3529 | removeEdgeFromGraph(Edge: Edge.get(), EI: &EI, /*CalleeIter=*/true); |
| 3530 | |
| 3531 | // To simplify the increment of EI in the caller, subtract one from EI. |
| 3532 | // In the final AddEdge call we would have either added a new callee edge, |
| 3533 | // to Edge->Caller, or found an existing one. Either way we are guaranteed |
| 3534 | // that there is at least one callee edge. |
| 3535 | assert(!Caller->CalleeEdges.empty()); |
| 3536 | --EI; |
| 3537 | |
| 3538 | return true; |
| 3539 | } |
| 3540 | |
| 3541 | bool ModuleCallsiteContextGraph::findProfiledCalleeThroughTailCalls( |
| 3542 | const Function *ProfiledCallee, Value *CurCallee, unsigned Depth, |
| 3543 | std::vector<std::pair<Instruction *, Function *>> &FoundCalleeChain, |
| 3544 | bool &FoundMultipleCalleeChains) { |
| 3545 | // Stop recursive search if we have already explored the maximum specified |
| 3546 | // depth. |
| 3547 | if (Depth > TailCallSearchDepth) |
| 3548 | return false; |
| 3549 | |
| 3550 | auto SaveCallsiteInfo = [&](Instruction *Callsite, Function *F) { |
| 3551 | FoundCalleeChain.push_back(x: {Callsite, F}); |
| 3552 | }; |
| 3553 | |
| 3554 | auto *CalleeFunc = dyn_cast<Function>(Val: CurCallee); |
| 3555 | if (!CalleeFunc) { |
| 3556 | auto *Alias = dyn_cast<GlobalAlias>(Val: CurCallee); |
| 3557 | assert(Alias); |
| 3558 | CalleeFunc = dyn_cast<Function>(Val: Alias->getAliasee()); |
| 3559 | assert(CalleeFunc); |
| 3560 | } |
| 3561 | |
| 3562 | // Look for tail calls in this function, and check if they either call the |
| 3563 | // profiled callee directly, or indirectly (via a recursive search). |
| 3564 | // Only succeed if there is a single unique tail call chain found between the |
| 3565 | // profiled caller and callee, otherwise we could perform incorrect cloning. |
| 3566 | bool FoundSingleCalleeChain = false; |
| 3567 | for (auto &BB : *CalleeFunc) { |
| 3568 | for (auto &I : BB) { |
| 3569 | auto *CB = dyn_cast<CallBase>(Val: &I); |
| 3570 | if (!CB || !CB->isTailCall()) |
| 3571 | continue; |
| 3572 | auto *CalledValue = CB->getCalledOperand(); |
| 3573 | auto *CalledFunction = CB->getCalledFunction(); |
| 3574 | if (CalledValue && !CalledFunction) { |
| 3575 | CalledValue = CalledValue->stripPointerCasts(); |
| 3576 | // Stripping pointer casts can reveal a called function. |
| 3577 | CalledFunction = dyn_cast<Function>(Val: CalledValue); |
| 3578 | } |
| 3579 | // Check if this is an alias to a function. If so, get the |
| 3580 | // called aliasee for the checks below. |
| 3581 | if (auto *GA = dyn_cast<GlobalAlias>(Val: CalledValue)) { |
| 3582 | assert(!CalledFunction && |
| 3583 | "Expected null called function in callsite for alias" ); |
| 3584 | CalledFunction = dyn_cast<Function>(Val: GA->getAliaseeObject()); |
| 3585 | } |
| 3586 | if (!CalledFunction) |
| 3587 | continue; |
| 3588 | if (CalledFunction == ProfiledCallee) { |
| 3589 | if (FoundSingleCalleeChain) { |
| 3590 | FoundMultipleCalleeChains = true; |
| 3591 | return false; |
| 3592 | } |
| 3593 | FoundSingleCalleeChain = true; |
| 3594 | FoundProfiledCalleeCount++; |
| 3595 | FoundProfiledCalleeDepth += Depth; |
| 3596 | if (Depth > FoundProfiledCalleeMaxDepth) |
| 3597 | FoundProfiledCalleeMaxDepth = Depth; |
| 3598 | SaveCallsiteInfo(&I, CalleeFunc); |
| 3599 | } else if (findProfiledCalleeThroughTailCalls( |
| 3600 | ProfiledCallee, CurCallee: CalledFunction, Depth: Depth + 1, |
| 3601 | FoundCalleeChain, FoundMultipleCalleeChains)) { |
| 3602 | // findProfiledCalleeThroughTailCalls should not have returned |
| 3603 | // true if FoundMultipleCalleeChains. |
| 3604 | assert(!FoundMultipleCalleeChains); |
| 3605 | if (FoundSingleCalleeChain) { |
| 3606 | FoundMultipleCalleeChains = true; |
| 3607 | return false; |
| 3608 | } |
| 3609 | FoundSingleCalleeChain = true; |
| 3610 | SaveCallsiteInfo(&I, CalleeFunc); |
| 3611 | } else if (FoundMultipleCalleeChains) |
| 3612 | return false; |
| 3613 | } |
| 3614 | } |
| 3615 | |
| 3616 | return FoundSingleCalleeChain; |
| 3617 | } |
| 3618 | |
| 3619 | const Function *ModuleCallsiteContextGraph::getCalleeFunc(Instruction *Call) { |
| 3620 | auto *CB = dyn_cast<CallBase>(Val: Call); |
| 3621 | if (!CB->getCalledOperand() || CB->isIndirectCall()) |
| 3622 | return nullptr; |
| 3623 | auto *CalleeVal = CB->getCalledOperand()->stripPointerCasts(); |
| 3624 | auto *Alias = dyn_cast<GlobalAlias>(Val: CalleeVal); |
| 3625 | if (Alias) |
| 3626 | return dyn_cast<Function>(Val: Alias->getAliasee()); |
| 3627 | return dyn_cast<Function>(Val: CalleeVal); |
| 3628 | } |
| 3629 | |
| 3630 | bool ModuleCallsiteContextGraph::calleeMatchesFunc( |
| 3631 | Instruction *Call, const Function *Func, const Function *CallerFunc, |
| 3632 | std::vector<std::pair<Instruction *, Function *>> &FoundCalleeChain) { |
| 3633 | auto *CB = dyn_cast<CallBase>(Val: Call); |
| 3634 | if (!CB->getCalledOperand() || CB->isIndirectCall()) |
| 3635 | return false; |
| 3636 | auto *CalleeVal = CB->getCalledOperand()->stripPointerCasts(); |
| 3637 | auto *CalleeFunc = dyn_cast<Function>(Val: CalleeVal); |
| 3638 | if (CalleeFunc == Func) |
| 3639 | return true; |
| 3640 | auto *Alias = dyn_cast<GlobalAlias>(Val: CalleeVal); |
| 3641 | if (Alias && Alias->getAliasee() == Func) |
| 3642 | return true; |
| 3643 | |
| 3644 | // Recursively search for the profiled callee through tail calls starting with |
| 3645 | // the actual Callee. The discovered tail call chain is saved in |
| 3646 | // FoundCalleeChain, and we will fixup the graph to include these callsites |
| 3647 | // after returning. |
| 3648 | // FIXME: We will currently redo the same recursive walk if we find the same |
| 3649 | // mismatched callee from another callsite. We can improve this with more |
| 3650 | // bookkeeping of the created chain of new nodes for each mismatch. |
| 3651 | unsigned Depth = 1; |
| 3652 | bool FoundMultipleCalleeChains = false; |
| 3653 | if (!findProfiledCalleeThroughTailCalls(ProfiledCallee: Func, CurCallee: CalleeVal, Depth, |
| 3654 | FoundCalleeChain, |
| 3655 | FoundMultipleCalleeChains)) { |
| 3656 | LLVM_DEBUG(dbgs() << "Not found through unique tail call chain: " |
| 3657 | << Func->getName() << " from " << CallerFunc->getName() |
| 3658 | << " that actually called " << CalleeVal->getName() |
| 3659 | << (FoundMultipleCalleeChains |
| 3660 | ? " (found multiple possible chains)" |
| 3661 | : "" ) |
| 3662 | << "\n" ); |
| 3663 | if (FoundMultipleCalleeChains) |
| 3664 | FoundProfiledCalleeNonUniquelyCount++; |
| 3665 | return false; |
| 3666 | } |
| 3667 | |
| 3668 | return true; |
| 3669 | } |
| 3670 | |
| 3671 | bool ModuleCallsiteContextGraph::sameCallee(Instruction *Call1, |
| 3672 | Instruction *Call2) { |
| 3673 | auto *CB1 = cast<CallBase>(Val: Call1); |
| 3674 | if (!CB1->getCalledOperand() || CB1->isIndirectCall()) |
| 3675 | return false; |
| 3676 | auto *CalleeVal1 = CB1->getCalledOperand()->stripPointerCasts(); |
| 3677 | auto *CalleeFunc1 = dyn_cast<Function>(Val: CalleeVal1); |
| 3678 | auto *CB2 = cast<CallBase>(Val: Call2); |
| 3679 | if (!CB2->getCalledOperand() || CB2->isIndirectCall()) |
| 3680 | return false; |
| 3681 | auto *CalleeVal2 = CB2->getCalledOperand()->stripPointerCasts(); |
| 3682 | auto *CalleeFunc2 = dyn_cast<Function>(Val: CalleeVal2); |
| 3683 | return CalleeFunc1 == CalleeFunc2; |
| 3684 | } |
| 3685 | |
| 3686 | bool IndexCallsiteContextGraph::findProfiledCalleeThroughTailCalls( |
| 3687 | ValueInfo ProfiledCallee, ValueInfo CurCallee, unsigned Depth, |
| 3688 | std::vector<std::pair<ValueInfo, FunctionSummary *>> &FoundCalleeChain, |
| 3689 | bool &FoundMultipleCalleeChains) { |
| 3690 | // Stop recursive search if we have already explored the maximum specified |
| 3691 | // depth. |
| 3692 | if (Depth > TailCallSearchDepth) |
| 3693 | return false; |
| 3694 | |
| 3695 | // Record each discovered tail call (callee and calling function) in the |
| 3696 | // chain. The CallsiteInfo records are only synthesized by the caller once |
| 3697 | // the search succeeds, since the chain is discarded if the profiled callee |
| 3698 | // is found through multiple chains. |
| 3699 | auto SaveTailCall = [&](ValueInfo Callee, FunctionSummary *FS) { |
| 3700 | FoundCalleeChain.push_back(x: {Callee, FS}); |
| 3701 | }; |
| 3702 | |
| 3703 | // Returns true if the given tail call callee is the profiled callee, or is |
| 3704 | // an alias of it. Only the aliasee's function summary gets a context node |
| 3705 | // (see calleeMatchesFunc), so the profiled callee is always the aliasee. The |
| 3706 | // synthesized callsite record is still keyed by the alias (the callee of the |
| 3707 | // call edge), which is what the ThinLTO backend looks up for the tail call |
| 3708 | // (see findValueInfoForGV). |
| 3709 | auto IsProfiledCallee = [&](ValueInfo Callee) { |
| 3710 | if (Callee == ProfiledCallee) |
| 3711 | return true; |
| 3712 | for (auto &S : Callee.getSummaryList()) { |
| 3713 | if (!GlobalValue::isLocalLinkage(Linkage: S->linkage()) && |
| 3714 | !isPrevailing(Callee.getGUID(), S.get())) |
| 3715 | continue; |
| 3716 | auto *AS = dyn_cast<AliasSummary>(Val: S.get()); |
| 3717 | return AS && AS->getAliaseeVI() == ProfiledCallee; |
| 3718 | } |
| 3719 | return false; |
| 3720 | }; |
| 3721 | |
| 3722 | // Look for tail calls in this function, and check if they either call the |
| 3723 | // profiled callee directly, or indirectly (via a recursive search). |
| 3724 | // Only succeed if there is a single unique tail call chain found between the |
| 3725 | // profiled caller and callee, otherwise we could perform incorrect cloning. |
| 3726 | bool FoundSingleCalleeChain = false; |
| 3727 | for (auto &S : CurCallee.getSummaryList()) { |
| 3728 | if (!GlobalValue::isLocalLinkage(Linkage: S->linkage()) && |
| 3729 | !isPrevailing(CurCallee.getGUID(), S.get())) |
| 3730 | continue; |
| 3731 | auto *FS = dyn_cast<FunctionSummary>(Val: S->getBaseObject()); |
| 3732 | if (!FS) |
| 3733 | continue; |
| 3734 | auto FSVI = CurCallee; |
| 3735 | auto *AS = dyn_cast<AliasSummary>(Val: S.get()); |
| 3736 | if (AS) |
| 3737 | FSVI = AS->getAliaseeVI(); |
| 3738 | for (auto &CallEdge : FS->calls()) { |
| 3739 | if (!CallEdge.second.hasTailCall()) |
| 3740 | continue; |
| 3741 | // Edges marked indirect are the profiled targets of indirect calls. |
| 3742 | // If the tail call is indirect, the ThinLTO backend will perform |
| 3743 | // ICP to the target in order to call a clone of it. |
| 3744 | if (CallEdge.second.hasIndirectCall() && !IndirectTailCallSearch) |
| 3745 | continue; |
| 3746 | if (IsProfiledCallee(CallEdge.first)) { |
| 3747 | if (FoundSingleCalleeChain) { |
| 3748 | FoundMultipleCalleeChains = true; |
| 3749 | return false; |
| 3750 | } |
| 3751 | FoundSingleCalleeChain = true; |
| 3752 | FoundProfiledCalleeCount++; |
| 3753 | FoundProfiledCalleeDepth += Depth; |
| 3754 | if (Depth > FoundProfiledCalleeMaxDepth) |
| 3755 | FoundProfiledCalleeMaxDepth = Depth; |
| 3756 | if (CallEdge.second.hasIndirectCall()) |
| 3757 | FoundProfiledCalleeIndirectTailCallCount++; |
| 3758 | SaveTailCall(CallEdge.first, FS); |
| 3759 | // Add FS to FSToVIMap in case it isn't already there. |
| 3760 | assert(!FSToVIMap.count(FS) || FSToVIMap[FS] == FSVI); |
| 3761 | FSToVIMap[FS] = FSVI; |
| 3762 | } else if (findProfiledCalleeThroughTailCalls( |
| 3763 | ProfiledCallee, CurCallee: CallEdge.first, Depth: Depth + 1, |
| 3764 | FoundCalleeChain, FoundMultipleCalleeChains)) { |
| 3765 | // findProfiledCalleeThroughTailCalls should not have returned |
| 3766 | // true if FoundMultipleCalleeChains. |
| 3767 | assert(!FoundMultipleCalleeChains); |
| 3768 | if (FoundSingleCalleeChain) { |
| 3769 | FoundMultipleCalleeChains = true; |
| 3770 | return false; |
| 3771 | } |
| 3772 | FoundSingleCalleeChain = true; |
| 3773 | if (CallEdge.second.hasIndirectCall()) |
| 3774 | FoundProfiledCalleeIndirectTailCallCount++; |
| 3775 | SaveTailCall(CallEdge.first, FS); |
| 3776 | // Add FS to FSToVIMap in case it isn't already there. |
| 3777 | assert(!FSToVIMap.count(FS) || FSToVIMap[FS] == FSVI); |
| 3778 | FSToVIMap[FS] = FSVI; |
| 3779 | } else if (FoundMultipleCalleeChains) |
| 3780 | return false; |
| 3781 | } |
| 3782 | } |
| 3783 | |
| 3784 | return FoundSingleCalleeChain; |
| 3785 | } |
| 3786 | |
| 3787 | const FunctionSummary * |
| 3788 | IndexCallsiteContextGraph::getCalleeFunc(IndexCall &Call) { |
| 3789 | ValueInfo Callee = dyn_cast_if_present<CallsiteInfo *>(Val&: Call)->Callee; |
| 3790 | if (Callee.getSummaryList().empty()) |
| 3791 | return nullptr; |
| 3792 | return dyn_cast<FunctionSummary>(Val: Callee.getSummaryList()[0]->getBaseObject()); |
| 3793 | } |
| 3794 | |
| 3795 | bool IndexCallsiteContextGraph::calleeMatchesFunc( |
| 3796 | IndexCall &Call, const FunctionSummary *Func, |
| 3797 | const FunctionSummary *CallerFunc, |
| 3798 | std::vector<std::pair<IndexCall, FunctionSummary *>> &FoundCalleeChain) { |
| 3799 | ValueInfo Callee = dyn_cast_if_present<CallsiteInfo *>(Val&: Call)->Callee; |
| 3800 | // If there is no summary list then this is a call to an externally defined |
| 3801 | // symbol. |
| 3802 | AliasSummary *Alias = |
| 3803 | Callee.getSummaryList().empty() |
| 3804 | ? nullptr |
| 3805 | : dyn_cast<AliasSummary>(Val: Callee.getSummaryList()[0].get()); |
| 3806 | assert(FSToVIMap.count(Func)); |
| 3807 | auto FuncVI = FSToVIMap[Func]; |
| 3808 | if (Callee == FuncVI || |
| 3809 | // If callee is an alias, check the aliasee, since only function |
| 3810 | // summary base objects will contain the stack node summaries and thus |
| 3811 | // get a context node. |
| 3812 | (Alias && Alias->getAliaseeVI() == FuncVI)) |
| 3813 | return true; |
| 3814 | |
| 3815 | // Recursively search for the profiled callee through tail calls starting with |
| 3816 | // the actual Callee. The discovered tail call chain is saved in |
| 3817 | // FoundCalleeChain, and we will fixup the graph to include these callsites |
| 3818 | // after returning. |
| 3819 | // FIXME: We will currently redo the same recursive walk if we find the same |
| 3820 | // mismatched callee from another callsite. We can improve this with more |
| 3821 | // bookkeeping of the created chain of new nodes for each mismatch. |
| 3822 | unsigned Depth = 1; |
| 3823 | bool FoundMultipleCalleeChains = false; |
| 3824 | std::vector<std::pair<ValueInfo, FunctionSummary *>> FoundCalleeVIChain; |
| 3825 | if (!findProfiledCalleeThroughTailCalls(ProfiledCallee: FuncVI, CurCallee: Callee, Depth, |
| 3826 | FoundCalleeChain&: FoundCalleeVIChain, |
| 3827 | FoundMultipleCalleeChains)) { |
| 3828 | LLVM_DEBUG(dbgs() << "Not found through unique tail call chain: " << FuncVI |
| 3829 | << " from " << FSToVIMap[CallerFunc] |
| 3830 | << " that actually called " << Callee |
| 3831 | << (FoundMultipleCalleeChains |
| 3832 | ? " (found multiple possible chains)" |
| 3833 | : "" ) |
| 3834 | << "\n" ); |
| 3835 | if (FoundMultipleCalleeChains) |
| 3836 | FoundProfiledCalleeNonUniquelyCount++; |
| 3837 | return false; |
| 3838 | } |
| 3839 | |
| 3840 | // Now that we know the chain is unique, synthesize a CallsiteInfo for each |
| 3841 | // tail call in it, if one hasn't already been synthesized (e.g. for a chain |
| 3842 | // to a different profiled callee through the same tail call). These are only |
| 3843 | // added to the function summaries once the graph is destroyed (see |
| 3844 | // ~IndexCallsiteContextGraph). |
| 3845 | for (auto &[CalleeVI, FS] : FoundCalleeVIChain) { |
| 3846 | auto &CallsiteInfoPtr = |
| 3847 | FunctionCalleesToSynthesizedCallsiteInfos[FS][CalleeVI]; |
| 3848 | if (!CallsiteInfoPtr) |
| 3849 | // StackIds is empty (we don't have debug info available in the index for |
| 3850 | // these callsites) |
| 3851 | CallsiteInfoPtr = |
| 3852 | std::make_unique<CallsiteInfo>(args&: CalleeVI, args: SmallVector<unsigned>()); |
| 3853 | FoundCalleeChain.push_back(x: {CallsiteInfoPtr.get(), FS}); |
| 3854 | } |
| 3855 | |
| 3856 | return true; |
| 3857 | } |
| 3858 | |
| 3859 | bool IndexCallsiteContextGraph::sameCallee(IndexCall &Call1, IndexCall &Call2) { |
| 3860 | ValueInfo Callee1 = dyn_cast_if_present<CallsiteInfo *>(Val&: Call1)->Callee; |
| 3861 | ValueInfo Callee2 = dyn_cast_if_present<CallsiteInfo *>(Val&: Call2)->Callee; |
| 3862 | return Callee1 == Callee2; |
| 3863 | } |
| 3864 | |
| 3865 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 3866 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextNode::dump() |
| 3867 | const { |
| 3868 | print(OS&: dbgs()); |
| 3869 | dbgs() << "\n" ; |
| 3870 | } |
| 3871 | |
| 3872 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 3873 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextNode::print( |
| 3874 | raw_ostream &OS) const { |
| 3875 | OS << "Node " << this << "\n" ; |
| 3876 | OS << "\t" ; |
| 3877 | printCall(OS); |
| 3878 | if (Recursive) |
| 3879 | OS << " (recursive)" ; |
| 3880 | OS << "\n" ; |
| 3881 | if (!MatchingCalls.empty()) { |
| 3882 | OS << "\tMatchingCalls:\n" ; |
| 3883 | for (auto &MatchingCall : MatchingCalls) { |
| 3884 | OS << "\t" ; |
| 3885 | MatchingCall.print(OS); |
| 3886 | OS << "\n" ; |
| 3887 | } |
| 3888 | } |
| 3889 | OS << "\tNodeId: " << NodeId << "\n" ; |
| 3890 | OS << "\tAllocTypes: " << getAllocTypeString(AllocTypes) << "\n" ; |
| 3891 | OS << "\tContextIds:" ; |
| 3892 | // Make a copy of the computed context ids that we can sort for stability. |
| 3893 | auto ContextIds = getContextIds(); |
| 3894 | std::vector<uint32_t> SortedIds(ContextIds.begin(), ContextIds.end()); |
| 3895 | std::sort(first: SortedIds.begin(), last: SortedIds.end()); |
| 3896 | for (auto Id : SortedIds) |
| 3897 | OS << " " << Id; |
| 3898 | OS << "\n" ; |
| 3899 | OS << "\tCalleeEdges:\n" ; |
| 3900 | for (auto &Edge : CalleeEdges) |
| 3901 | OS << "\t\t" << *Edge << " (Callee NodeId: " << Edge->Callee->NodeId |
| 3902 | << ")\n" ; |
| 3903 | OS << "\tCallerEdges:\n" ; |
| 3904 | for (auto &Edge : CallerEdges) |
| 3905 | OS << "\t\t" << *Edge << " (Caller NodeId: " << Edge->Caller->NodeId |
| 3906 | << ")\n" ; |
| 3907 | if (!Clones.empty()) { |
| 3908 | OS << "\tClones: " ; |
| 3909 | ListSeparator LS; |
| 3910 | for (auto *C : Clones) |
| 3911 | OS << LS << C << " NodeId: " << C->NodeId; |
| 3912 | OS << "\n" ; |
| 3913 | } else if (CloneOf) { |
| 3914 | OS << "\tClone of " << CloneOf << " NodeId: " << CloneOf->NodeId << "\n" ; |
| 3915 | } |
| 3916 | } |
| 3917 | |
| 3918 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 3919 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextEdge::dump() |
| 3920 | const { |
| 3921 | print(OS&: dbgs()); |
| 3922 | dbgs() << "\n" ; |
| 3923 | } |
| 3924 | |
| 3925 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 3926 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextEdge::print( |
| 3927 | raw_ostream &OS) const { |
| 3928 | OS << "Edge from Callee " << Callee << " to Caller: " << Caller |
| 3929 | << (IsBackedge ? " (BE)" : "" ) |
| 3930 | << " AllocTypes: " << getAllocTypeString(AllocTypes); |
| 3931 | OS << " ContextIds:" ; |
| 3932 | std::vector<uint32_t> SortedIds(ContextIds.begin(), ContextIds.end()); |
| 3933 | std::sort(first: SortedIds.begin(), last: SortedIds.end()); |
| 3934 | for (auto Id : SortedIds) |
| 3935 | OS << " " << Id; |
| 3936 | } |
| 3937 | |
| 3938 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 3939 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::dump() const { |
| 3940 | print(OS&: dbgs()); |
| 3941 | } |
| 3942 | |
| 3943 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 3944 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::print( |
| 3945 | raw_ostream &OS) const { |
| 3946 | OS << "Callsite Context Graph:\n" ; |
| 3947 | using GraphType = const CallsiteContextGraph<DerivedCCG, FuncTy, CallTy> *; |
| 3948 | for (const auto Node : nodes<GraphType>(this)) { |
| 3949 | if (Node->isRemoved()) |
| 3950 | continue; |
| 3951 | Node->print(OS); |
| 3952 | OS << "\n" ; |
| 3953 | } |
| 3954 | } |
| 3955 | |
| 3956 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 3957 | void CallsiteContextGraph<DerivedCCG, FuncTy, |
| 3958 | CallTy>::capturePreCloningContextInfo() { |
| 3959 | for (const auto &Entry : AllocationCallToContextNodeMap) { |
| 3960 | const ContextNode *AllocNode = Entry.second; |
| 3961 | // Remember whether any caller of this allocation matched a call, so that |
| 3962 | // the report can distinguish "no matched callers" from an allocation whose |
| 3963 | // callers are simply not part of this context. |
| 3964 | for (const auto &Edge : AllocNode->CallerEdges) { |
| 3965 | if (!Edge->isRemoved() && Edge->Caller->hasCall()) { |
| 3966 | PreCloningAllocHasMatchedCaller.insert(AllocNode); |
| 3967 | break; |
| 3968 | } |
| 3969 | } |
| 3970 | DenseSet<uint32_t> ContextIds = AllocNode->getContextIds(); |
| 3971 | for (uint32_t Id : ContextIds) { |
| 3972 | PreCloningContextInfo Info; |
| 3973 | // Walk the caller edges carrying this context id from the allocation |
| 3974 | // node outwards to build the chain of nodes for this context. Each node |
| 3975 | // is visited once; backedges (recursive cycles) are only followed when |
| 3976 | // no other caller edge carries the id, so that the chain reflects the |
| 3977 | // context's stack order rather than looping. |
| 3978 | Info.Chain.push_back(AllocNode); |
| 3979 | SmallPtrSet<const ContextNode *, 16> Visited; |
| 3980 | Visited.insert(AllocNode); |
| 3981 | const ContextNode *Cur = AllocNode; |
| 3982 | while (true) { |
| 3983 | const ContextNode *Next = nullptr; |
| 3984 | for (const auto &Edge : Cur->CallerEdges) { |
| 3985 | if (Edge->isRemoved() || Edge->IsBackedge) |
| 3986 | continue; |
| 3987 | if (Edge->getContextIds().contains(Id) && |
| 3988 | !Visited.contains(Edge->Caller)) { |
| 3989 | Next = Edge->Caller; |
| 3990 | break; |
| 3991 | } |
| 3992 | } |
| 3993 | if (!Next) { |
| 3994 | for (const auto &Edge : Cur->CallerEdges) { |
| 3995 | if (Edge->isRemoved()) |
| 3996 | continue; |
| 3997 | if (Edge->getContextIds().contains(Id) && |
| 3998 | !Visited.contains(Edge->Caller)) { |
| 3999 | Next = Edge->Caller; |
| 4000 | break; |
| 4001 | } |
| 4002 | } |
| 4003 | } |
| 4004 | if (!Next) |
| 4005 | break; |
| 4006 | Visited.insert(Next); |
| 4007 | Info.Chain.push_back(Next); |
| 4008 | Cur = Next; |
| 4009 | } |
| 4010 | |
| 4011 | // Find the outermost node (walking from the allocation) such that it |
| 4012 | // and all nodes below it matched a call, and record why the first |
| 4013 | // unmatched node (if any) did not match. |
| 4014 | Info.MaxMatchedFrame = 0; |
| 4015 | for (unsigned I = 1; I < Info.Chain.size(); ++I) { |
| 4016 | if (Info.Chain[I]->hasCall()) { |
| 4017 | Info.MaxMatchedFrame = I; |
| 4018 | } else { |
| 4019 | const ContextNode *Unmatched = Info.Chain[I]; |
| 4020 | Info.UnmatchedStackId = Unmatched->OrigStackOrAllocId; |
| 4021 | Info.StoppedAtRecursive = Unmatched->Recursive; |
| 4022 | // A calling function is known for the unmatched node only when a |
| 4023 | // call in that function was considered and rejected because its |
| 4024 | // callee did not match the node's callee. |
| 4025 | if (auto It = NodeToCallingFunc.find(Unmatched); |
| 4026 | It != NodeToCallingFunc.end()) { |
| 4027 | Info.StoppedAtMismatchedCallee = true; |
| 4028 | Info.UnmatchedCallingFunc = It->second; |
| 4029 | } |
| 4030 | break; |
| 4031 | } |
| 4032 | } |
| 4033 | |
| 4034 | // If the whole context matched, check whether its outermost node has |
| 4035 | // matched callers reached by other contexts: then this context was |
| 4036 | // trimmed below them, and clones of those callers can change which |
| 4037 | // allocation clone it reaches. |
| 4038 | if (Info.MaxMatchedFrame > 0 && |
| 4039 | Info.MaxMatchedFrame + 1 == Info.Chain.size()) { |
| 4040 | const ContextNode *OutermostMatchedNode = |
| 4041 | Info.Chain[Info.MaxMatchedFrame]; |
| 4042 | for (const auto &Edge : OutermostMatchedNode->CallerEdges) { |
| 4043 | if (!Edge->isRemoved() && Edge->Caller->hasCall()) { |
| 4044 | Info.OutermostMatchedNodeHasOtherCallers = true; |
| 4045 | break; |
| 4046 | } |
| 4047 | } |
| 4048 | } |
| 4049 | |
| 4050 | PreCloningContextInfoMap[Id] = std::move(Info); |
| 4051 | } |
| 4052 | } |
| 4053 | } |
| 4054 | |
| 4055 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 4056 | std::string |
| 4057 | CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::formatContextReportSuffix( |
| 4058 | const ContextNode *Node, uint32_t Id) const { |
| 4059 | auto InfoIt = PreCloningContextInfoMap.find(Id); |
| 4060 | if (InfoIt == PreCloningContextInfoMap.end()) |
| 4061 | return "" ; |
| 4062 | const PreCloningContextInfo &Info = InfoIt->second; |
| 4063 | |
| 4064 | // 1. The allocation function (clone) the context's allocation node ended up |
| 4065 | // in, with a qualifier when it is dead, or when the context has callers in |
| 4066 | // the profile but none of the allocation's callers matched (either because |
| 4067 | // nothing references the function at all, or because the references did not |
| 4068 | // match the profiled callsites). |
| 4069 | const ContextNode *OrigAlloc = Node->getOrigNode(); |
| 4070 | const FuncTy *AllocFunc = getCallingFunc(N: OrigAlloc); |
| 4071 | unsigned AllocCloneNo = Node->Call.cloneNo(); |
| 4072 | std::string AllocFuncStr = getFuncName(Func: AllocFunc, CloneNo: AllocCloneNo); |
| 4073 | if (isFuncDead(Func: AllocFunc)) { |
| 4074 | AllocFuncStr += " (dead)" ; |
| 4075 | } else if (!PreCloningAllocHasMatchedCaller.contains(OrigAlloc) && |
| 4076 | Info.Chain.size() > 1) { |
| 4077 | // Distinguish a function with no references in the IR/summary (it may |
| 4078 | // still be referenced from native objects) from one whose references did |
| 4079 | // not match. |
| 4080 | AllocFuncStr += AllocFunc && !FunctionsWithCallers.contains(AllocFunc) |
| 4081 | ? " (no references in IR/summary)" |
| 4082 | : " (no matched callers)" ; |
| 4083 | } |
| 4084 | |
| 4085 | // 2. How far the context matched: all of it ("matched N frames to F", or |
| 4086 | // "trimmed at frame N in F" when F has callers above that the context does |
| 4087 | // not cover), or up to an unmatched node (with the reason and stack id). |
| 4088 | unsigned TotalFrames = Info.Chain.size() > 0 ? Info.Chain.size() - 1 : 0; |
| 4089 | unsigned Matched = Info.MaxMatchedFrame; |
| 4090 | std::string StatusStr; |
| 4091 | if (TotalFrames == 0) { |
| 4092 | StatusStr = "inlined into leaf" ; |
| 4093 | } else if (Matched < TotalFrames) { |
| 4094 | // Where and why matching stopped: the first unmatched node was either a |
| 4095 | // recursive callsite (not matched when recursive callsites are disabled), |
| 4096 | // a callsite whose IR callee did not match the profiled callee, or a |
| 4097 | // callsite with no matching call at all. |
| 4098 | if (Info.StoppedAtRecursive) |
| 4099 | StatusStr = "unmatched recursive callsite" ; |
| 4100 | else if (Info.StoppedAtMismatchedCallee) |
| 4101 | StatusStr = "unmatched callsite (mismatched callee in " + |
| 4102 | getFuncName(Func: Info.UnmatchedCallingFunc, CloneNo: 0) + ")" ; |
| 4103 | else |
| 4104 | StatusStr = "unmatched callsite" ; |
| 4105 | StatusStr += " at frame " + std::to_string(val: Matched + 1) + "/" + |
| 4106 | std::to_string(val: TotalFrames) + " (stack id " + |
| 4107 | std::to_string(Info.UnmatchedStackId) + ")" ; |
| 4108 | } else { |
| 4109 | const ContextNode *OutermostMatchedNode = Info.Chain[Matched]; |
| 4110 | const FuncTy *OutermostMatchedFunc = getCallingFunc(N: OutermostMatchedNode); |
| 4111 | unsigned NumOutermostMatchedFuncClones = |
| 4112 | std::max<unsigned>(1, FuncToNumClones.lookup(OutermostMatchedFunc)); |
| 4113 | bool HasCallersAbove = |
| 4114 | NumOutermostMatchedFuncClones > 1 || |
| 4115 | Info.OutermostMatchedNodeHasOtherCallers || |
| 4116 | (OutermostMatchedFunc && |
| 4117 | FunctionsWithCallers.contains(OutermostMatchedFunc)); |
| 4118 | if (HasCallersAbove) { |
| 4119 | StatusStr = "trimmed at frame " + std::to_string(val: Matched) + " in " + |
| 4120 | getFuncName(Func: OutermostMatchedFunc, CloneNo: 0); |
| 4121 | } else { |
| 4122 | StatusStr = "matched " + std::to_string(val: Matched) + " frames to " + |
| 4123 | getFuncName(Func: OutermostMatchedFunc, CloneNo: 0); |
| 4124 | } |
| 4125 | } |
| 4126 | |
| 4127 | // 3. The post-cloning call paths: starting from each function clone of the |
| 4128 | // outermost matched caller, follow the callee clones assigned to the |
| 4129 | // context's callsites down to the allocation clone reached, and report the |
| 4130 | // hint that clone ended up with. Paths reaching the same allocation clone |
| 4131 | // are reported once. |
| 4132 | std::string ReachesStr; |
| 4133 | if (Matched >= 1) { |
| 4134 | const ContextNode *OutermostMatchedNode = Info.Chain[Matched]; |
| 4135 | const FuncTy *OutermostMatchedFunc = getCallingFunc(N: OutermostMatchedNode); |
| 4136 | unsigned NumOutermostMatchedFuncClones = |
| 4137 | std::max<unsigned>(1, FuncToNumClones.lookup(OutermostMatchedFunc)); |
| 4138 | ReachesStr = "; reaches " ; |
| 4139 | // Allocation clones (by clone number) already reported for this context. |
| 4140 | // Different function clones of the outermost matched caller often lead to |
| 4141 | // the same allocation clone; such paths are reported only once, for the |
| 4142 | // first caller clone reaching that allocation clone. |
| 4143 | SmallSet<unsigned, 4> ReportedAllocClones; |
| 4144 | bool FirstPath = true; |
| 4145 | for (unsigned OutermostMatchedFuncClone = 0; |
| 4146 | OutermostMatchedFuncClone < NumOutermostMatchedFuncClones; |
| 4147 | ++OutermostMatchedFuncClone) { |
| 4148 | // PathFuncs/PathClones[S] is the function and clone number at step S of |
| 4149 | // the path, from the outermost matched caller (S = 0) down to the |
| 4150 | // allocation function (S = Matched). |
| 4151 | std::vector<const FuncTy *> PathFuncs(Matched + 1); |
| 4152 | std::vector<unsigned> PathClones(Matched + 1); |
| 4153 | PathFuncs[0] = OutermostMatchedFunc; |
| 4154 | PathClones[0] = OutermostMatchedFuncClone; |
| 4155 | unsigned CurClone = OutermostMatchedFuncClone; |
| 4156 | for (unsigned Step = 0; Step < Matched; ++Step) { |
| 4157 | unsigned FrameIdx = Matched - Step; |
| 4158 | const ContextNode *CallNode = Info.Chain[FrameIdx]; |
| 4159 | unsigned NextClone = |
| 4160 | OrigNodeAndCloneToCalleeClone.lookup({CallNode, CurClone}); |
| 4161 | const FuncTy *CalleeFunc = getCallingFunc(N: Info.Chain[FrameIdx - 1]); |
| 4162 | PathFuncs[Step + 1] = CalleeFunc; |
| 4163 | PathClones[Step + 1] = NextClone; |
| 4164 | CurClone = NextClone; |
| 4165 | } |
| 4166 | |
| 4167 | if (!ReportedAllocClones.insert(V: CurClone).second) |
| 4168 | continue; |
| 4169 | if (!FirstPath) |
| 4170 | ReachesStr += ", " ; |
| 4171 | FirstPath = false; |
| 4172 | |
| 4173 | const ContextNode *ReachedAlloc = |
| 4174 | OrigAllocAndCloneToNode.lookup({OrigAlloc, CurClone}); |
| 4175 | // Report the hint actually applied to the reached allocation clone |
| 4176 | // (which accounts for allocTypeToUse and MinClonedColdBytePercent), |
| 4177 | // rather than the graph node's (possibly ambiguous) AllocTypes. |
| 4178 | std::string ReachedTypeStr; |
| 4179 | if (ReachedAlloc) |
| 4180 | ReachedTypeStr = getAllocTypeString( |
| 4181 | AllocTypes: (uint8_t)getAllocationCallType(Call: ReachedAlloc->Call)); |
| 4182 | else if (CurClone == 0) |
| 4183 | ReachedTypeStr = |
| 4184 | getAllocTypeString(AllocTypes: (uint8_t)getAllocationCallType(Call: OrigAlloc->Call)); |
| 4185 | else |
| 4186 | ReachedTypeStr = "None" ; |
| 4187 | |
| 4188 | // Print the path, eliding runs of uncloned (or cloned) functions in the |
| 4189 | // middle of paths longer than MemProfReportHintedSizesDetailMaxPathLen; |
| 4190 | // the ends and every transition between cloned and uncloned functions |
| 4191 | // are kept. |
| 4192 | std::string PathStr; |
| 4193 | bool SkippedPrev = false; |
| 4194 | bool PrintFullPath = |
| 4195 | Matched + 1 <= MemProfReportHintedSizesDetailMaxPathLen; |
| 4196 | for (unsigned S = 0; S <= Matched; ++S) { |
| 4197 | bool IsCloned = PathClones[S] > 0; |
| 4198 | bool PrevCloned = S > 0 ? (PathClones[S - 1] > 0) : IsCloned; |
| 4199 | bool NextCloned = S + 1 <= Matched ? (PathClones[S + 1] > 0) : IsCloned; |
| 4200 | bool Include = PrintFullPath || (S <= 1) || (S + 1 >= Matched) || |
| 4201 | (IsCloned != PrevCloned) || (IsCloned != NextCloned); |
| 4202 | if (!Include) { |
| 4203 | if (!SkippedPrev) { |
| 4204 | PathStr += " -> ..." ; |
| 4205 | SkippedPrev = true; |
| 4206 | } |
| 4207 | continue; |
| 4208 | } |
| 4209 | SkippedPrev = false; |
| 4210 | if (!PathStr.empty()) |
| 4211 | PathStr += " -> " ; |
| 4212 | PathStr += getFuncName(Func: PathFuncs[S], CloneNo: PathClones[S]); |
| 4213 | } |
| 4214 | ReachesStr += ReachedTypeStr + " via " + PathStr; |
| 4215 | } |
| 4216 | } |
| 4217 | |
| 4218 | return " [alloc in " + AllocFuncStr + "; " + StatusStr + ReachesStr + "]" ; |
| 4219 | } |
| 4220 | |
| 4221 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 4222 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::printTotalSizes( |
| 4223 | raw_ostream &OS, |
| 4224 | function_ref<void(StringRef, StringRef, const Twine &)> ) const { |
| 4225 | using GraphType = const CallsiteContextGraph<DerivedCCG, FuncTy, CallTy> *; |
| 4226 | for (const auto Node : nodes<GraphType>(this)) { |
| 4227 | if (Node->isRemoved()) |
| 4228 | continue; |
| 4229 | if (!Node->IsAllocation) |
| 4230 | continue; |
| 4231 | DenseSet<uint32_t> ContextIds = Node->getContextIds(); |
| 4232 | auto AllocTypeFromCall = getAllocationCallType(Call: Node->Call); |
| 4233 | std::vector<uint32_t> SortedIds(ContextIds.begin(), ContextIds.end()); |
| 4234 | std::sort(first: SortedIds.begin(), last: SortedIds.end()); |
| 4235 | for (auto Id : SortedIds) { |
| 4236 | auto TypeI = ContextIdToAllocationType.find(Val: Id); |
| 4237 | assert(TypeI != ContextIdToAllocationType.end()); |
| 4238 | std::string Suffix; |
| 4239 | bool ComputedSuffix = false; |
| 4240 | auto GetSuffix = [&]() -> const std::string & { |
| 4241 | if (!ComputedSuffix) { |
| 4242 | Suffix = MemProfReportHintedSizesDetail |
| 4243 | ? formatContextReportSuffix(Node, Id) |
| 4244 | : "" ; |
| 4245 | ComputedSuffix = true; |
| 4246 | } |
| 4247 | return Suffix; |
| 4248 | }; |
| 4249 | auto CSI = ContextIdToContextSizeInfos.find(Val: Id); |
| 4250 | if (CSI != ContextIdToContextSizeInfos.end()) { |
| 4251 | for (auto &Info : CSI->second) { |
| 4252 | std::string Msg = |
| 4253 | "MemProf hinting: " + getAllocTypeString(AllocTypes: (uint8_t)TypeI->second) + |
| 4254 | " full allocation context " + std::to_string(val: Info.FullStackId) + |
| 4255 | " with total size " + std::to_string(val: Info.TotalSize) + " is " + |
| 4256 | getAllocTypeString(Node->AllocTypes) + " after cloning" ; |
| 4257 | if (allocTypeToUse(Node->AllocTypes) != |
| 4258 | allocTypeToUse(AllocTypes: (uint8_t)AllocTypeFromCall)) |
| 4259 | Msg += " marked " + getAllocTypeString(AllocTypes: (uint8_t)AllocTypeFromCall) + |
| 4260 | " due to cold byte percent" ; |
| 4261 | // Print the internal context id to aid debugging and visualization. |
| 4262 | Msg += " (internal context id " + std::to_string(val: Id) + ")" ; |
| 4263 | if (MemProfReportHintedSizes) |
| 4264 | OS << Msg << GetSuffix() << "\n" ; |
| 4265 | if (EmitRemark) |
| 4266 | EmitRemark(DEBUG_TYPE, "MemProfReport" , Msg); |
| 4267 | } |
| 4268 | } else { |
| 4269 | // This is only emitted if the context size info is not present. |
| 4270 | std::string Msg = |
| 4271 | "MemProf hinting: " + getAllocTypeString(AllocTypes: (uint8_t)TypeI->second) + |
| 4272 | " context is " + getAllocTypeString(Node->AllocTypes) + |
| 4273 | " after cloning" ; |
| 4274 | if (allocTypeToUse(Node->AllocTypes) != |
| 4275 | allocTypeToUse(AllocTypes: (uint8_t)AllocTypeFromCall)) |
| 4276 | Msg += " marked " + getAllocTypeString(AllocTypes: (uint8_t)AllocTypeFromCall) + |
| 4277 | " due to cold byte percent" ; |
| 4278 | // Print the internal context id to aid debugging and visualization. |
| 4279 | Msg += " (internal context id " + std::to_string(val: Id) + ")" ; |
| 4280 | if (MemProfReportHintedSizes) |
| 4281 | OS << Msg << GetSuffix() << "\n" ; |
| 4282 | if (EmitRemark) |
| 4283 | EmitRemark(DEBUG_TYPE, "MemProfReport" , Msg); |
| 4284 | } |
| 4285 | } |
| 4286 | } |
| 4287 | } |
| 4288 | |
| 4289 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 4290 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::check() const { |
| 4291 | using GraphType = const CallsiteContextGraph<DerivedCCG, FuncTy, CallTy> *; |
| 4292 | for (const auto Node : nodes<GraphType>(this)) { |
| 4293 | checkNode<DerivedCCG, FuncTy, CallTy>(Node, /*CheckEdges=*/false); |
| 4294 | for (auto &Edge : Node->CallerEdges) |
| 4295 | checkEdge<DerivedCCG, FuncTy, CallTy>(Edge); |
| 4296 | } |
| 4297 | } |
| 4298 | |
| 4299 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 4300 | struct GraphTraits<const CallsiteContextGraph<DerivedCCG, FuncTy, CallTy> *> { |
| 4301 | using GraphType = const CallsiteContextGraph<DerivedCCG, FuncTy, CallTy> *; |
| 4302 | using NodeRef = const ContextNode<DerivedCCG, FuncTy, CallTy> *; |
| 4303 | |
| 4304 | using NodePtrTy = std::unique_ptr<ContextNode<DerivedCCG, FuncTy, CallTy>>; |
| 4305 | static NodeRef getNode(const NodePtrTy &P) { return P.get(); } |
| 4306 | |
| 4307 | using nodes_iterator = |
| 4308 | mapped_iterator<typename std::vector<NodePtrTy>::const_iterator, |
| 4309 | decltype(&getNode)>; |
| 4310 | |
| 4311 | static nodes_iterator nodes_begin(GraphType G) { |
| 4312 | return nodes_iterator(G->NodeOwner.begin(), &getNode); |
| 4313 | } |
| 4314 | |
| 4315 | static nodes_iterator nodes_end(GraphType G) { |
| 4316 | return nodes_iterator(G->NodeOwner.end(), &getNode); |
| 4317 | } |
| 4318 | |
| 4319 | static NodeRef getEntryNode(GraphType G) { |
| 4320 | return G->NodeOwner.begin()->get(); |
| 4321 | } |
| 4322 | |
| 4323 | using EdgePtrTy = std::shared_ptr<ContextEdge<DerivedCCG, FuncTy, CallTy>>; |
| 4324 | static const ContextNode<DerivedCCG, FuncTy, CallTy> * |
| 4325 | GetCallee(const EdgePtrTy &P) { |
| 4326 | return P->Callee; |
| 4327 | } |
| 4328 | |
| 4329 | using ChildIteratorType = |
| 4330 | mapped_iterator<typename std::vector<std::shared_ptr<ContextEdge< |
| 4331 | DerivedCCG, FuncTy, CallTy>>>::const_iterator, |
| 4332 | decltype(&GetCallee)>; |
| 4333 | |
| 4334 | static ChildIteratorType child_begin(NodeRef N) { |
| 4335 | return ChildIteratorType(N->CalleeEdges.begin(), &GetCallee); |
| 4336 | } |
| 4337 | |
| 4338 | static ChildIteratorType child_end(NodeRef N) { |
| 4339 | return ChildIteratorType(N->CalleeEdges.end(), &GetCallee); |
| 4340 | } |
| 4341 | }; |
| 4342 | |
| 4343 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 4344 | struct DOTGraphTraits<const CallsiteContextGraph<DerivedCCG, FuncTy, CallTy> *> |
| 4345 | : public DefaultDOTGraphTraits { |
| 4346 | DOTGraphTraits(bool IsSimple = false) : DefaultDOTGraphTraits(IsSimple) { |
| 4347 | // If the user requested the full graph to be exported, but provided an |
| 4348 | // allocation id, or if the user gave a context id and requested more than |
| 4349 | // just a specific context to be exported, note that highlighting is |
| 4350 | // enabled. |
| 4351 | DoHighlight = |
| 4352 | (AllocIdForDot.getNumOccurrences() && DotGraphScope == DotScope::All) || |
| 4353 | (ContextIdForDot.getNumOccurrences() && |
| 4354 | DotGraphScope != DotScope::Context); |
| 4355 | } |
| 4356 | |
| 4357 | using GraphType = const CallsiteContextGraph<DerivedCCG, FuncTy, CallTy> *; |
| 4358 | using GTraits = GraphTraits<GraphType>; |
| 4359 | using NodeRef = typename GTraits::NodeRef; |
| 4360 | using ChildIteratorType = typename GTraits::ChildIteratorType; |
| 4361 | |
| 4362 | static std::string getNodeLabel(NodeRef Node, GraphType G) { |
| 4363 | std::string LabelString = |
| 4364 | (Twine("OrigId: " ) + (Node->IsAllocation ? "Alloc" : "" ) + |
| 4365 | Twine(Node->OrigStackOrAllocId) + " NodeId: " + Twine(Node->NodeId)) |
| 4366 | .str(); |
| 4367 | LabelString += "\n" ; |
| 4368 | if (Node->hasCall()) { |
| 4369 | auto Func = G->NodeToCallingFunc.find(Node); |
| 4370 | assert(Func != G->NodeToCallingFunc.end()); |
| 4371 | LabelString += |
| 4372 | G->getLabel(Func->second, Node->Call.call(), Node->Call.cloneNo()); |
| 4373 | for (auto &MatchingCall : Node->MatchingCalls) { |
| 4374 | LabelString += "\n" ; |
| 4375 | LabelString += G->getLabel(Func->second, MatchingCall.call(), |
| 4376 | MatchingCall.cloneNo()); |
| 4377 | } |
| 4378 | } else { |
| 4379 | LabelString += "null call" ; |
| 4380 | if (Node->Recursive) |
| 4381 | LabelString += " (recursive)" ; |
| 4382 | else |
| 4383 | LabelString += " (external)" ; |
| 4384 | } |
| 4385 | return LabelString; |
| 4386 | } |
| 4387 | |
| 4388 | static std::string getNodeAttributes(NodeRef Node, GraphType G) { |
| 4389 | auto ContextIds = Node->getContextIds(); |
| 4390 | // If highlighting enabled, see if this node contains any of the context ids |
| 4391 | // of interest. If so, it will use a different color and a larger fontsize |
| 4392 | // (which makes the node larger as well). |
| 4393 | bool Highlight = false; |
| 4394 | if (DoHighlight) { |
| 4395 | assert(ContextIdForDot.getNumOccurrences() || |
| 4396 | AllocIdForDot.getNumOccurrences()); |
| 4397 | if (ContextIdForDot.getNumOccurrences()) |
| 4398 | Highlight = ContextIds.contains(ContextIdForDot); |
| 4399 | else |
| 4400 | Highlight = set_intersects(ContextIds, G->DotAllocContextIds); |
| 4401 | } |
| 4402 | std::string AttributeString = (Twine("tooltip=\"" ) + getNodeId(Node) + " " + |
| 4403 | getContextIds(ContextIds) + "\"" ) |
| 4404 | .str(); |
| 4405 | // Default fontsize is 14 |
| 4406 | if (Highlight) |
| 4407 | AttributeString += ",fontsize=\"30\"" ; |
| 4408 | AttributeString += |
| 4409 | (Twine(",fillcolor=\"" ) + getColor(AllocTypes: Node->AllocTypes, Highlight) + "\"" ) |
| 4410 | .str(); |
| 4411 | if (Node->CloneOf) { |
| 4412 | AttributeString += ",color=\"blue\"" ; |
| 4413 | AttributeString += ",style=\"filled,bold,dashed\"" ; |
| 4414 | } else |
| 4415 | AttributeString += ",style=\"filled\"" ; |
| 4416 | return AttributeString; |
| 4417 | } |
| 4418 | |
| 4419 | static std::string getEdgeAttributes(NodeRef, ChildIteratorType ChildIter, |
| 4420 | GraphType G) { |
| 4421 | auto &Edge = *(ChildIter.getCurrent()); |
| 4422 | // If highlighting enabled, see if this edge contains any of the context ids |
| 4423 | // of interest. If so, it will use a different color and a heavier arrow |
| 4424 | // size and weight (the larger weight makes the highlighted path |
| 4425 | // straighter). |
| 4426 | bool Highlight = false; |
| 4427 | if (DoHighlight) { |
| 4428 | assert(ContextIdForDot.getNumOccurrences() || |
| 4429 | AllocIdForDot.getNumOccurrences()); |
| 4430 | if (ContextIdForDot.getNumOccurrences()) |
| 4431 | Highlight = Edge->ContextIds.contains(ContextIdForDot); |
| 4432 | else |
| 4433 | Highlight = set_intersects(Edge->ContextIds, G->DotAllocContextIds); |
| 4434 | } |
| 4435 | auto Color = getColor(AllocTypes: Edge->AllocTypes, Highlight); |
| 4436 | std::string AttributeString = |
| 4437 | (Twine("tooltip=\"" ) + getContextIds(ContextIds: Edge->ContextIds) + "\"" + |
| 4438 | // fillcolor is the arrow head and color is the line |
| 4439 | Twine(",fillcolor=\"" ) + Color + "\"" + Twine(",color=\"" ) + Color + |
| 4440 | "\"" ) |
| 4441 | .str(); |
| 4442 | if (Edge->IsBackedge) |
| 4443 | AttributeString += ",style=\"dotted\"" ; |
| 4444 | // Default penwidth and weight are both 1. |
| 4445 | if (Highlight) |
| 4446 | AttributeString += ",penwidth=\"2.0\",weight=\"2\"" ; |
| 4447 | return AttributeString; |
| 4448 | } |
| 4449 | |
| 4450 | // Since the NodeOwners list includes nodes that are no longer connected to |
| 4451 | // the graph, skip them here. |
| 4452 | static bool isNodeHidden(NodeRef Node, GraphType G) { |
| 4453 | if (Node->isRemoved()) |
| 4454 | return true; |
| 4455 | // If a scope smaller than the full graph was requested, see if this node |
| 4456 | // contains any of the context ids of interest. |
| 4457 | if (DotGraphScope == DotScope::Alloc) |
| 4458 | return !set_intersects(Node->getContextIds(), G->DotAllocContextIds); |
| 4459 | if (DotGraphScope == DotScope::Context) |
| 4460 | return !Node->getContextIds().contains(ContextIdForDot); |
| 4461 | return false; |
| 4462 | } |
| 4463 | |
| 4464 | private: |
| 4465 | static std::string getContextIds(const DenseSet<uint32_t> &ContextIds) { |
| 4466 | std::string IdString = "ContextIds:" ; |
| 4467 | if (ContextIds.size() < 100) { |
| 4468 | std::vector<uint32_t> SortedIds(ContextIds.begin(), ContextIds.end()); |
| 4469 | std::sort(first: SortedIds.begin(), last: SortedIds.end()); |
| 4470 | for (auto Id : SortedIds) |
| 4471 | IdString += (" " + Twine(Id)).str(); |
| 4472 | } else { |
| 4473 | IdString += (" (" + Twine(ContextIds.size()) + " ids)" ).str(); |
| 4474 | } |
| 4475 | return IdString; |
| 4476 | } |
| 4477 | |
| 4478 | static std::string getColor(uint8_t AllocTypes, bool Highlight) { |
| 4479 | // If DoHighlight is not enabled, we want to use the highlight colors for |
| 4480 | // NotCold and Cold, and the non-highlight color for NotCold+Cold. This is |
| 4481 | // both compatible with the color scheme before highlighting was supported, |
| 4482 | // and for the NotCold+Cold color the non-highlight color is a bit more |
| 4483 | // readable. |
| 4484 | if (AllocTypes == (uint8_t)AllocationType::NotCold) |
| 4485 | // Color "brown1" actually looks like a lighter red. |
| 4486 | return !DoHighlight || Highlight ? "brown1" : "lightpink" ; |
| 4487 | if (AllocTypes == (uint8_t)AllocationType::Cold) |
| 4488 | return !DoHighlight || Highlight ? "cyan" : "lightskyblue" ; |
| 4489 | if (AllocTypes == |
| 4490 | ((uint8_t)AllocationType::NotCold | (uint8_t)AllocationType::Cold)) |
| 4491 | return Highlight ? "magenta" : "mediumorchid1" ; |
| 4492 | return "gray" ; |
| 4493 | } |
| 4494 | |
| 4495 | static std::string getNodeId(NodeRef Node) { |
| 4496 | std::stringstream SStream; |
| 4497 | SStream << std::hex << "N0x" << (unsigned long long)Node; |
| 4498 | std::string Result = SStream.str(); |
| 4499 | return Result; |
| 4500 | } |
| 4501 | |
| 4502 | // True if we should highlight a specific context or allocation's contexts in |
| 4503 | // the emitted graph. |
| 4504 | static bool DoHighlight; |
| 4505 | }; |
| 4506 | |
| 4507 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 4508 | bool DOTGraphTraits< |
| 4509 | const CallsiteContextGraph<DerivedCCG, FuncTy, CallTy> *>::DoHighlight = |
| 4510 | false; |
| 4511 | |
| 4512 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 4513 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::exportToDot( |
| 4514 | std::string Label) const { |
| 4515 | WriteGraph(this, "" , false, Label, |
| 4516 | DotFilePathPrefix + "ccg." + Label + ".dot" ); |
| 4517 | } |
| 4518 | |
| 4519 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 4520 | typename CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextNode * |
| 4521 | CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::moveEdgeToNewCalleeClone( |
| 4522 | const std::shared_ptr<ContextEdge> &Edge, |
| 4523 | DenseSet<uint32_t> ContextIdsToMove) { |
| 4524 | ContextNode *Node = Edge->Callee; |
| 4525 | assert(NodeToCallingFunc.count(Node)); |
| 4526 | ContextNode *Clone = |
| 4527 | createNewNode(IsAllocation: Node->IsAllocation, F: NodeToCallingFunc[Node], C: Node->Call); |
| 4528 | Node->addClone(Clone); |
| 4529 | Clone->MatchingCalls = Node->MatchingCalls; |
| 4530 | moveEdgeToExistingCalleeClone(Edge, NewCallee: Clone, /*NewClone=*/true, |
| 4531 | ContextIdsToMove); |
| 4532 | return Clone; |
| 4533 | } |
| 4534 | |
| 4535 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 4536 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>:: |
| 4537 | moveEdgeToExistingCalleeClone(const std::shared_ptr<ContextEdge> &Edge, |
| 4538 | ContextNode *NewCallee, bool NewClone, |
| 4539 | DenseSet<uint32_t> ContextIdsToMove) { |
| 4540 | // NewCallee and Edge's current callee must be clones of the same original |
| 4541 | // node (Edge's current callee may be the original node too). |
| 4542 | assert(NewCallee->getOrigNode() == Edge->Callee->getOrigNode()); |
| 4543 | |
| 4544 | bool EdgeIsRecursive = Edge->Callee == Edge->Caller; |
| 4545 | |
| 4546 | ContextNode *OldCallee = Edge->Callee; |
| 4547 | |
| 4548 | // We might already have an edge to the new callee from earlier cloning for a |
| 4549 | // different allocation. If one exists we will reuse it. |
| 4550 | auto ExistingEdgeToNewCallee = NewCallee->findEdgeFromCaller(Edge->Caller); |
| 4551 | |
| 4552 | // Callers will pass an empty ContextIdsToMove set when they want to move the |
| 4553 | // edge. Copy in Edge's ids for simplicity. |
| 4554 | if (ContextIdsToMove.empty()) |
| 4555 | ContextIdsToMove = Edge->getContextIds(); |
| 4556 | |
| 4557 | // If we are moving all of Edge's ids, then just move the whole Edge. |
| 4558 | // Otherwise only move the specified subset, to a new edge if needed. |
| 4559 | if (Edge->getContextIds().size() == ContextIdsToMove.size()) { |
| 4560 | // First, update the alloc types on New Callee from Edge. |
| 4561 | // Do this before we potentially clear Edge's fields below! |
| 4562 | NewCallee->AllocTypes |= Edge->AllocTypes; |
| 4563 | // Moving the whole Edge. |
| 4564 | if (ExistingEdgeToNewCallee) { |
| 4565 | // Since we already have an edge to NewCallee, simply move the ids |
| 4566 | // onto it, and remove the existing Edge. |
| 4567 | ExistingEdgeToNewCallee->getContextIds().insert_range(ContextIdsToMove); |
| 4568 | ExistingEdgeToNewCallee->AllocTypes |= Edge->AllocTypes; |
| 4569 | assert(Edge->ContextIds == ContextIdsToMove); |
| 4570 | removeEdgeFromGraph(Edge: Edge.get()); |
| 4571 | } else { |
| 4572 | // Otherwise just reconnect Edge to NewCallee. |
| 4573 | Edge->Callee = NewCallee; |
| 4574 | NewCallee->CallerEdges.push_back(Edge); |
| 4575 | // Remove it from callee where it was previously connected. |
| 4576 | OldCallee->eraseCallerEdge(Edge.get()); |
| 4577 | // Don't need to update Edge's context ids since we are simply |
| 4578 | // reconnecting it. |
| 4579 | } |
| 4580 | } else { |
| 4581 | // Only moving a subset of Edge's ids. |
| 4582 | // Compute the alloc type of the subset of ids being moved. |
| 4583 | auto CallerEdgeAllocType = computeAllocType(ContextIds&: ContextIdsToMove); |
| 4584 | if (ExistingEdgeToNewCallee) { |
| 4585 | // Since we already have an edge to NewCallee, simply move the ids |
| 4586 | // onto it. |
| 4587 | ExistingEdgeToNewCallee->getContextIds().insert_range(ContextIdsToMove); |
| 4588 | ExistingEdgeToNewCallee->AllocTypes |= CallerEdgeAllocType; |
| 4589 | } else { |
| 4590 | // Otherwise, create a new edge to NewCallee for the ids being moved. |
| 4591 | auto NewEdge = std::make_shared<ContextEdge>( |
| 4592 | NewCallee, Edge->Caller, CallerEdgeAllocType, ContextIdsToMove); |
| 4593 | Edge->Caller->CalleeEdges.push_back(NewEdge); |
| 4594 | NewCallee->CallerEdges.push_back(NewEdge); |
| 4595 | } |
| 4596 | // In either case, need to update the alloc types on NewCallee, and remove |
| 4597 | // those ids and update the alloc type on the original Edge. |
| 4598 | NewCallee->AllocTypes |= CallerEdgeAllocType; |
| 4599 | set_subtract(Edge->ContextIds, ContextIdsToMove); |
| 4600 | Edge->AllocTypes = computeAllocType(ContextIds&: Edge->ContextIds); |
| 4601 | } |
| 4602 | // Now walk the old callee node's callee edges and move Edge's context ids |
| 4603 | // over to the corresponding edge into the clone (which is created here if |
| 4604 | // this is a newly created clone). |
| 4605 | for (auto &OldCalleeEdge : OldCallee->CalleeEdges) { |
| 4606 | ContextNode *CalleeToUse = OldCalleeEdge->Callee; |
| 4607 | // If this is a direct recursion edge, use NewCallee (the clone) as the |
| 4608 | // callee as well, so that any edge updated/created here is also direct |
| 4609 | // recursive. |
| 4610 | if (CalleeToUse == OldCallee) { |
| 4611 | // If this is a recursive edge, see if we already moved a recursive edge |
| 4612 | // (which would have to have been this one) - if we were only moving a |
| 4613 | // subset of context ids it would still be on OldCallee. |
| 4614 | if (EdgeIsRecursive) { |
| 4615 | assert(OldCalleeEdge == Edge); |
| 4616 | continue; |
| 4617 | } |
| 4618 | CalleeToUse = NewCallee; |
| 4619 | } |
| 4620 | // The context ids moving to the new callee are the subset of this edge's |
| 4621 | // context ids and the context ids on the caller edge being moved. |
| 4622 | DenseSet<uint32_t> EdgeContextIdsToMove = |
| 4623 | set_intersection(OldCalleeEdge->getContextIds(), ContextIdsToMove); |
| 4624 | set_subtract(OldCalleeEdge->getContextIds(), EdgeContextIdsToMove); |
| 4625 | OldCalleeEdge->AllocTypes = |
| 4626 | computeAllocType(ContextIds&: OldCalleeEdge->getContextIds()); |
| 4627 | if (!NewClone) { |
| 4628 | // Update context ids / alloc type on corresponding edge to NewCallee. |
| 4629 | // There is a chance this may not exist if we are reusing an existing |
| 4630 | // clone, specifically during function assignment, where we would have |
| 4631 | // removed none type edges after creating the clone. If we can't find |
| 4632 | // a corresponding edge there, fall through to the cloning below. |
| 4633 | if (auto *NewCalleeEdge = NewCallee->findEdgeFromCallee(CalleeToUse)) { |
| 4634 | NewCalleeEdge->getContextIds().insert_range(EdgeContextIdsToMove); |
| 4635 | NewCalleeEdge->AllocTypes |= computeAllocType(ContextIds&: EdgeContextIdsToMove); |
| 4636 | continue; |
| 4637 | } |
| 4638 | } |
| 4639 | auto NewEdge = std::make_shared<ContextEdge>( |
| 4640 | CalleeToUse, NewCallee, computeAllocType(ContextIds&: EdgeContextIdsToMove), |
| 4641 | EdgeContextIdsToMove); |
| 4642 | NewCallee->CalleeEdges.push_back(NewEdge); |
| 4643 | NewEdge->Callee->CallerEdges.push_back(NewEdge); |
| 4644 | } |
| 4645 | // Recompute the node alloc type now that its callee edges have been |
| 4646 | // updated (since we will compute from those edges). |
| 4647 | OldCallee->AllocTypes = OldCallee->computeAllocType(); |
| 4648 | // OldCallee alloc type should be None iff its context id set is now empty. |
| 4649 | assert((OldCallee->AllocTypes == (uint8_t)AllocationType::None) == |
| 4650 | OldCallee->emptyContextIds()); |
| 4651 | if (VerifyCCG) { |
| 4652 | checkNode<DerivedCCG, FuncTy, CallTy>(OldCallee, /*CheckEdges=*/false); |
| 4653 | checkNode<DerivedCCG, FuncTy, CallTy>(NewCallee, /*CheckEdges=*/false); |
| 4654 | for (const auto &OldCalleeEdge : OldCallee->CalleeEdges) |
| 4655 | checkNode<DerivedCCG, FuncTy, CallTy>(OldCalleeEdge->Callee, |
| 4656 | /*CheckEdges=*/false); |
| 4657 | for (const auto &NewCalleeEdge : NewCallee->CalleeEdges) |
| 4658 | checkNode<DerivedCCG, FuncTy, CallTy>(NewCalleeEdge->Callee, |
| 4659 | /*CheckEdges=*/false); |
| 4660 | } |
| 4661 | } |
| 4662 | |
| 4663 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 4664 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>:: |
| 4665 | moveCalleeEdgeToNewCaller(const std::shared_ptr<ContextEdge> &Edge, |
| 4666 | ContextNode *NewCaller) { |
| 4667 | auto *OldCallee = Edge->Callee; |
| 4668 | auto *NewCallee = OldCallee; |
| 4669 | // If this edge was direct recursive, make any new/updated edge also direct |
| 4670 | // recursive to NewCaller. |
| 4671 | bool Recursive = Edge->Caller == Edge->Callee; |
| 4672 | if (Recursive) |
| 4673 | NewCallee = NewCaller; |
| 4674 | |
| 4675 | ContextNode *OldCaller = Edge->Caller; |
| 4676 | OldCaller->eraseCalleeEdge(Edge.get()); |
| 4677 | |
| 4678 | // We might already have an edge to the new caller. If one exists we will |
| 4679 | // reuse it. |
| 4680 | auto ExistingEdgeToNewCaller = NewCaller->findEdgeFromCallee(NewCallee); |
| 4681 | |
| 4682 | if (ExistingEdgeToNewCaller) { |
| 4683 | // Since we already have an edge to NewCaller, simply move the ids |
| 4684 | // onto it, and remove the existing Edge. |
| 4685 | ExistingEdgeToNewCaller->getContextIds().insert_range( |
| 4686 | Edge->getContextIds()); |
| 4687 | ExistingEdgeToNewCaller->AllocTypes |= Edge->AllocTypes; |
| 4688 | Edge->ContextIds.clear(); |
| 4689 | Edge->AllocTypes = (uint8_t)AllocationType::None; |
| 4690 | OldCallee->eraseCallerEdge(Edge.get()); |
| 4691 | } else { |
| 4692 | // Otherwise just reconnect Edge to NewCaller. |
| 4693 | Edge->Caller = NewCaller; |
| 4694 | NewCaller->CalleeEdges.push_back(Edge); |
| 4695 | if (Recursive) { |
| 4696 | assert(NewCallee == NewCaller); |
| 4697 | // In the case of (direct) recursive edges, we update the callee as well |
| 4698 | // so that it becomes recursive on the new caller. |
| 4699 | Edge->Callee = NewCallee; |
| 4700 | NewCallee->CallerEdges.push_back(Edge); |
| 4701 | OldCallee->eraseCallerEdge(Edge.get()); |
| 4702 | } |
| 4703 | // Don't need to update Edge's context ids since we are simply |
| 4704 | // reconnecting it. |
| 4705 | } |
| 4706 | // In either case, need to update the alloc types on New Caller. |
| 4707 | NewCaller->AllocTypes |= Edge->AllocTypes; |
| 4708 | |
| 4709 | // Now walk the old caller node's caller edges and move Edge's context ids |
| 4710 | // over to the corresponding edge into the node (which is created here if |
| 4711 | // this is a newly created node). We can tell whether this is a newly created |
| 4712 | // node by seeing if it has any caller edges yet. |
| 4713 | #ifndef NDEBUG |
| 4714 | bool IsNewNode = NewCaller->CallerEdges.empty(); |
| 4715 | #endif |
| 4716 | // If we just moved a direct recursive edge, presumably its context ids should |
| 4717 | // also flow out of OldCaller via some other non-recursive callee edge. We |
| 4718 | // don't want to remove the recursive context ids from other caller edges yet, |
| 4719 | // otherwise the context ids get into an inconsistent state on OldCaller. |
| 4720 | // We will update these context ids on the non-recursive caller edge when and |
| 4721 | // if they are updated on the non-recursive callee. |
| 4722 | if (!Recursive) { |
| 4723 | for (auto &OldCallerEdge : OldCaller->CallerEdges) { |
| 4724 | auto OldCallerCaller = OldCallerEdge->Caller; |
| 4725 | // The context ids moving to the new caller are the subset of this edge's |
| 4726 | // context ids and the context ids on the callee edge being moved. |
| 4727 | DenseSet<uint32_t> EdgeContextIdsToMove = set_intersection( |
| 4728 | OldCallerEdge->getContextIds(), Edge->getContextIds()); |
| 4729 | if (OldCaller == OldCallerCaller) { |
| 4730 | OldCallerCaller = NewCaller; |
| 4731 | // Don't actually move this one. The caller will move it directly via a |
| 4732 | // call to this function with this as the Edge if it is appropriate to |
| 4733 | // move to a diff node that has a matching callee (itself). |
| 4734 | continue; |
| 4735 | } |
| 4736 | set_subtract(OldCallerEdge->getContextIds(), EdgeContextIdsToMove); |
| 4737 | OldCallerEdge->AllocTypes = |
| 4738 | computeAllocType(ContextIds&: OldCallerEdge->getContextIds()); |
| 4739 | // In this function we expect that any pre-existing node already has edges |
| 4740 | // from the same callers as the old node. That should be true in the |
| 4741 | // current use case, where we will remove None-type edges after copying |
| 4742 | // over all caller edges from the callee. |
| 4743 | auto *ExistingCallerEdge = NewCaller->findEdgeFromCaller(OldCallerCaller); |
| 4744 | // Since we would have skipped caller edges when moving a direct recursive |
| 4745 | // edge, this may not hold true when recursive handling enabled. |
| 4746 | assert(IsNewNode || ExistingCallerEdge || AllowRecursiveCallsites); |
| 4747 | if (ExistingCallerEdge) { |
| 4748 | ExistingCallerEdge->getContextIds().insert_range(EdgeContextIdsToMove); |
| 4749 | ExistingCallerEdge->AllocTypes |= |
| 4750 | computeAllocType(ContextIds&: EdgeContextIdsToMove); |
| 4751 | continue; |
| 4752 | } |
| 4753 | auto NewEdge = std::make_shared<ContextEdge>( |
| 4754 | NewCaller, OldCallerCaller, computeAllocType(ContextIds&: EdgeContextIdsToMove), |
| 4755 | EdgeContextIdsToMove); |
| 4756 | NewCaller->CallerEdges.push_back(NewEdge); |
| 4757 | NewEdge->Caller->CalleeEdges.push_back(NewEdge); |
| 4758 | } |
| 4759 | } |
| 4760 | // Recompute the node alloc type now that its caller edges have been |
| 4761 | // updated (since we will compute from those edges). |
| 4762 | OldCaller->AllocTypes = OldCaller->computeAllocType(); |
| 4763 | // OldCaller alloc type should be None iff its context id set is now empty. |
| 4764 | assert((OldCaller->AllocTypes == (uint8_t)AllocationType::None) == |
| 4765 | OldCaller->emptyContextIds()); |
| 4766 | if (VerifyCCG) { |
| 4767 | checkNode<DerivedCCG, FuncTy, CallTy>(OldCaller, /*CheckEdges=*/false); |
| 4768 | checkNode<DerivedCCG, FuncTy, CallTy>(NewCaller, /*CheckEdges=*/false); |
| 4769 | for (const auto &OldCallerEdge : OldCaller->CallerEdges) |
| 4770 | checkNode<DerivedCCG, FuncTy, CallTy>(OldCallerEdge->Caller, |
| 4771 | /*CheckEdges=*/false); |
| 4772 | for (const auto &NewCallerEdge : NewCaller->CallerEdges) |
| 4773 | checkNode<DerivedCCG, FuncTy, CallTy>(NewCallerEdge->Caller, |
| 4774 | /*CheckEdges=*/false); |
| 4775 | } |
| 4776 | } |
| 4777 | |
| 4778 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 4779 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>:: |
| 4780 | recursivelyRemoveNoneTypeCalleeEdges( |
| 4781 | ContextNode *Node, DenseSet<const ContextNode *> &Visited) { |
| 4782 | auto Inserted = Visited.insert(Node); |
| 4783 | if (!Inserted.second) |
| 4784 | return; |
| 4785 | |
| 4786 | removeNoneTypeCalleeEdges(Node); |
| 4787 | |
| 4788 | for (auto *Clone : Node->Clones) |
| 4789 | recursivelyRemoveNoneTypeCalleeEdges(Node: Clone, Visited); |
| 4790 | |
| 4791 | // The recursive call may remove some of this Node's caller edges. |
| 4792 | // Iterate over a copy and skip any that were removed. |
| 4793 | auto CallerEdges = Node->CallerEdges; |
| 4794 | for (auto &Edge : CallerEdges) { |
| 4795 | // Skip any that have been removed by an earlier recursive call. |
| 4796 | if (Edge->isRemoved()) { |
| 4797 | assert(!is_contained(Node->CallerEdges, Edge)); |
| 4798 | continue; |
| 4799 | } |
| 4800 | recursivelyRemoveNoneTypeCalleeEdges(Node: Edge->Caller, Visited); |
| 4801 | } |
| 4802 | } |
| 4803 | |
| 4804 | // This is the standard DFS based backedge discovery algorithm. |
| 4805 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 4806 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::markBackedges() { |
| 4807 | // If we are cloning recursive contexts, find and mark backedges from all root |
| 4808 | // callers, using the typical DFS based backedge analysis. |
| 4809 | if (!CloneRecursiveContexts) |
| 4810 | return; |
| 4811 | DenseSet<const ContextNode *> Visited; |
| 4812 | DenseSet<const ContextNode *> CurrentStack; |
| 4813 | for (auto &Entry : NonAllocationCallToContextNodeMap) { |
| 4814 | auto *Node = Entry.second; |
| 4815 | if (Node->isRemoved()) |
| 4816 | continue; |
| 4817 | // It is a root if it doesn't have callers. |
| 4818 | if (!Node->CallerEdges.empty()) |
| 4819 | continue; |
| 4820 | markBackedges(Node, Visited, CurrentStack); |
| 4821 | assert(CurrentStack.empty()); |
| 4822 | } |
| 4823 | } |
| 4824 | |
| 4825 | // Recursive helper for above markBackedges method. |
| 4826 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 4827 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::markBackedges( |
| 4828 | ContextNode *Node, DenseSet<const ContextNode *> &Visited, |
| 4829 | DenseSet<const ContextNode *> &CurrentStack) { |
| 4830 | auto I = Visited.insert(Node); |
| 4831 | // We should only call this for unvisited nodes. |
| 4832 | assert(I.second); |
| 4833 | (void)I; |
| 4834 | for (auto &CalleeEdge : Node->CalleeEdges) { |
| 4835 | auto *Callee = CalleeEdge->Callee; |
| 4836 | if (Visited.count(Callee)) { |
| 4837 | // Since this was already visited we need to check if it is currently on |
| 4838 | // the recursive stack in which case it is a backedge. |
| 4839 | if (CurrentStack.count(Callee)) |
| 4840 | CalleeEdge->IsBackedge = true; |
| 4841 | continue; |
| 4842 | } |
| 4843 | CurrentStack.insert(Callee); |
| 4844 | markBackedges(Callee, Visited, CurrentStack); |
| 4845 | CurrentStack.erase(Callee); |
| 4846 | } |
| 4847 | } |
| 4848 | |
| 4849 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 4850 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::identifyClones() { |
| 4851 | DenseSet<const ContextNode *> Visited; |
| 4852 | for (auto &Entry : AllocationCallToContextNodeMap) { |
| 4853 | Visited.clear(); |
| 4854 | identifyClones(Entry.second, Visited, Entry.second->getContextIds()); |
| 4855 | } |
| 4856 | Visited.clear(); |
| 4857 | for (auto &Entry : AllocationCallToContextNodeMap) |
| 4858 | recursivelyRemoveNoneTypeCalleeEdges(Node: Entry.second, Visited); |
| 4859 | if (VerifyCCG) |
| 4860 | check(); |
| 4861 | } |
| 4862 | |
| 4863 | // helper function to check an AllocType is cold or notcold or both. |
| 4864 | bool checkColdOrNotCold(uint8_t AllocType) { |
| 4865 | return (AllocType == (uint8_t)AllocationType::Cold) || |
| 4866 | (AllocType == (uint8_t)AllocationType::NotCold) || |
| 4867 | (AllocType == |
| 4868 | ((uint8_t)AllocationType::Cold | (uint8_t)AllocationType::NotCold)); |
| 4869 | } |
| 4870 | |
| 4871 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 4872 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::identifyClones( |
| 4873 | ContextNode *Node, DenseSet<const ContextNode *> &Visited, |
| 4874 | const DenseSet<uint32_t> &AllocContextIds) { |
| 4875 | if (VerifyNodes) |
| 4876 | checkNode<DerivedCCG, FuncTy, CallTy>(Node, /*CheckEdges=*/false); |
| 4877 | assert(!Node->CloneOf); |
| 4878 | |
| 4879 | // If Node as a null call, then either it wasn't found in the module (regular |
| 4880 | // LTO) or summary index (ThinLTO), or there were other conditions blocking |
| 4881 | // cloning (e.g. recursion, calls multiple targets, etc). |
| 4882 | // Do this here so that we don't try to recursively clone callers below, which |
| 4883 | // isn't useful at least for this node. |
| 4884 | if (!Node->hasCall()) |
| 4885 | return; |
| 4886 | |
| 4887 | // No need to look at any callers if allocation type already unambiguous. |
| 4888 | if (hasSingleAllocType(Node->AllocTypes)) |
| 4889 | return; |
| 4890 | |
| 4891 | #ifndef NDEBUG |
| 4892 | auto Insert = |
| 4893 | #endif |
| 4894 | Visited.insert(Node); |
| 4895 | // We should not have visited this node yet. |
| 4896 | assert(Insert.second); |
| 4897 | // The recursive call to identifyClones may delete the current edge from the |
| 4898 | // CallerEdges vector. Make a copy and iterate on that, simpler than passing |
| 4899 | // in an iterator and having recursive call erase from it. Other edges may |
| 4900 | // also get removed during the recursion, which will have null Callee and |
| 4901 | // Caller pointers (and are deleted later), so we skip those below. |
| 4902 | { |
| 4903 | auto CallerEdges = Node->CallerEdges; |
| 4904 | for (auto &Edge : CallerEdges) { |
| 4905 | // Skip any that have been removed by an earlier recursive call. |
| 4906 | if (Edge->isRemoved()) { |
| 4907 | assert(!is_contained(Node->CallerEdges, Edge)); |
| 4908 | continue; |
| 4909 | } |
| 4910 | // Defer backedges. See comments further below where these edges are |
| 4911 | // handled during the cloning of this Node. |
| 4912 | if (Edge->IsBackedge) { |
| 4913 | // We should only mark these if cloning recursive contexts, where we |
| 4914 | // need to do this deferral. |
| 4915 | assert(CloneRecursiveContexts); |
| 4916 | continue; |
| 4917 | } |
| 4918 | // Ignore any caller we previously visited via another edge. |
| 4919 | if (!Visited.count(Edge->Caller) && !Edge->Caller->CloneOf) { |
| 4920 | identifyClones(Edge->Caller, Visited, AllocContextIds); |
| 4921 | } |
| 4922 | } |
| 4923 | } |
| 4924 | |
| 4925 | // Check if we reached an unambiguous call or have have only a single caller. |
| 4926 | if (hasSingleAllocType(Node->AllocTypes) || Node->CallerEdges.size() <= 1) |
| 4927 | return; |
| 4928 | |
| 4929 | // We need to clone. |
| 4930 | |
| 4931 | // Try to keep the original version as alloc type NotCold. This will make |
| 4932 | // cases with indirect calls or any other situation with an unknown call to |
| 4933 | // the original function get the default behavior. We do this by sorting the |
| 4934 | // CallerEdges of the Node we will clone by alloc type. |
| 4935 | // |
| 4936 | // Give NotCold edge the lowest sort priority so those edges are at the end of |
| 4937 | // the caller edges vector, and stay on the original version (since the below |
| 4938 | // code clones greedily until it finds all remaining edges have the same type |
| 4939 | // and leaves the remaining ones on the original Node). |
| 4940 | // |
| 4941 | // We shouldn't actually have any None type edges, so the sorting priority for |
| 4942 | // that is arbitrary, and we assert in that case below. |
| 4943 | const unsigned AllocTypeCloningPriority[] = {/*None*/ 3, /*NotCold*/ 4, |
| 4944 | /*Cold*/ 1, |
| 4945 | /*NotColdCold*/ 2}; |
| 4946 | llvm::stable_sort(Node->CallerEdges, |
| 4947 | [&](const std::shared_ptr<ContextEdge> &A, |
| 4948 | const std::shared_ptr<ContextEdge> &B) { |
| 4949 | // Nodes with non-empty context ids should be sorted |
| 4950 | // before those with empty context ids. |
| 4951 | if (A->ContextIds.empty()) |
| 4952 | // Either B ContextIds are non-empty (in which case we |
| 4953 | // should return false because B < A), or B ContextIds |
| 4954 | // are empty, in which case they are equal, and we |
| 4955 | // should maintain the original relative ordering. |
| 4956 | return false; |
| 4957 | if (B->ContextIds.empty()) |
| 4958 | return true; |
| 4959 | |
| 4960 | if (A->AllocTypes == B->AllocTypes) |
| 4961 | // Use the caller node id as a deterministic |
| 4962 | // tie-breaker. |
| 4963 | return A->Caller->NodeId < B->Caller->NodeId; |
| 4964 | return AllocTypeCloningPriority[A->AllocTypes] < |
| 4965 | AllocTypeCloningPriority[B->AllocTypes]; |
| 4966 | }); |
| 4967 | |
| 4968 | assert(Node->AllocTypes != (uint8_t)AllocationType::None); |
| 4969 | |
| 4970 | DenseSet<uint32_t> RecursiveContextIds; |
| 4971 | assert(AllowRecursiveContexts || !CloneRecursiveContexts); |
| 4972 | // If we are allowing recursive callsites, but have also disabled recursive |
| 4973 | // contexts, look for context ids that show up in multiple caller edges. |
| 4974 | if (AllowRecursiveCallsites && !AllowRecursiveContexts) { |
| 4975 | DenseSet<uint32_t> AllCallerContextIds; |
| 4976 | for (auto &CE : Node->CallerEdges) { |
| 4977 | // Resize to the largest set of caller context ids, since we know the |
| 4978 | // final set will be at least that large. |
| 4979 | AllCallerContextIds.reserve(Size: CE->getContextIds().size()); |
| 4980 | for (auto Id : CE->getContextIds()) |
| 4981 | if (!AllCallerContextIds.insert(Id).second) |
| 4982 | RecursiveContextIds.insert(Id); |
| 4983 | } |
| 4984 | } |
| 4985 | |
| 4986 | // Iterate until we find no more opportunities for disambiguating the alloc |
| 4987 | // types via cloning. In most cases this loop will terminate once the Node |
| 4988 | // has a single allocation type, in which case no more cloning is needed. |
| 4989 | // Iterate over a copy of Node's caller edges, since we may need to remove |
| 4990 | // edges in the moveEdgeTo* methods, and this simplifies the handling and |
| 4991 | // makes it less error-prone. |
| 4992 | auto CallerEdges = Node->CallerEdges; |
| 4993 | for (auto &CallerEdge : CallerEdges) { |
| 4994 | // Skip any that have been removed by an earlier recursive call. |
| 4995 | if (CallerEdge->isRemoved()) { |
| 4996 | assert(!is_contained(Node->CallerEdges, CallerEdge)); |
| 4997 | continue; |
| 4998 | } |
| 4999 | assert(CallerEdge->Callee == Node); |
| 5000 | |
| 5001 | // See if cloning the prior caller edge left this node with a single alloc |
| 5002 | // type or a single caller. In that case no more cloning of Node is needed. |
| 5003 | if (hasSingleAllocType(Node->AllocTypes) || Node->CallerEdges.size() <= 1) |
| 5004 | break; |
| 5005 | |
| 5006 | // If the caller was not successfully matched to a call in the IR/summary, |
| 5007 | // there is no point in trying to clone for it as we can't update that call. |
| 5008 | if (!CallerEdge->Caller->hasCall()) |
| 5009 | continue; |
| 5010 | |
| 5011 | // Only need to process the ids along this edge pertaining to the given |
| 5012 | // allocation. |
| 5013 | auto CallerEdgeContextsForAlloc = |
| 5014 | set_intersection(CallerEdge->getContextIds(), AllocContextIds); |
| 5015 | if (!RecursiveContextIds.empty()) |
| 5016 | CallerEdgeContextsForAlloc = |
| 5017 | set_difference(CallerEdgeContextsForAlloc, RecursiveContextIds); |
| 5018 | if (CallerEdgeContextsForAlloc.empty()) |
| 5019 | continue; |
| 5020 | |
| 5021 | auto CallerAllocTypeForAlloc = computeAllocType(ContextIds&: CallerEdgeContextsForAlloc); |
| 5022 | |
| 5023 | // Compute the node callee edge alloc types corresponding to the context ids |
| 5024 | // for this caller edge. |
| 5025 | std::vector<uint8_t> CalleeEdgeAllocTypesForCallerEdge; |
| 5026 | CalleeEdgeAllocTypesForCallerEdge.reserve(n: Node->CalleeEdges.size()); |
| 5027 | for (auto &CalleeEdge : Node->CalleeEdges) |
| 5028 | CalleeEdgeAllocTypesForCallerEdge.push_back(intersectAllocTypes( |
| 5029 | Node1Ids: CalleeEdge->getContextIds(), Node2Ids: CallerEdgeContextsForAlloc)); |
| 5030 | |
| 5031 | // Don't clone if doing so will not disambiguate any alloc types amongst |
| 5032 | // caller edges (including the callee edges that would be cloned). |
| 5033 | // Otherwise we will simply move all edges to the clone. |
| 5034 | // |
| 5035 | // First check if by cloning we will disambiguate the caller allocation |
| 5036 | // type from node's allocation type. Query allocTypeToUse so that we don't |
| 5037 | // bother cloning to distinguish NotCold+Cold from NotCold. Note that |
| 5038 | // neither of these should be None type. |
| 5039 | // |
| 5040 | // Then check if by cloning node at least one of the callee edges will be |
| 5041 | // disambiguated by splitting out different context ids. |
| 5042 | // |
| 5043 | // However, always do the cloning if this is a backedge, in which case we |
| 5044 | // have not yet cloned along this caller edge. |
| 5045 | assert(CallerEdge->AllocTypes != (uint8_t)AllocationType::None); |
| 5046 | assert(Node->AllocTypes != (uint8_t)AllocationType::None); |
| 5047 | if (!CallerEdge->IsBackedge && |
| 5048 | allocTypeToUse(CallerAllocTypeForAlloc) == |
| 5049 | allocTypeToUse(Node->AllocTypes) && |
| 5050 | allocTypesMatch<DerivedCCG, FuncTy, CallTy>( |
| 5051 | CalleeEdgeAllocTypesForCallerEdge, Node->CalleeEdges)) { |
| 5052 | continue; |
| 5053 | } |
| 5054 | |
| 5055 | if (CallerEdge->IsBackedge) { |
| 5056 | // We should only mark these if cloning recursive contexts, where we |
| 5057 | // need to do this deferral. |
| 5058 | assert(CloneRecursiveContexts); |
| 5059 | DeferredBackedges++; |
| 5060 | } |
| 5061 | |
| 5062 | // If this is a backedge, we now do recursive cloning starting from its |
| 5063 | // caller since we may have moved unambiguous caller contexts to a clone |
| 5064 | // of this Node in a previous iteration of the current loop, giving more |
| 5065 | // opportunity for cloning through the backedge. Because we sorted the |
| 5066 | // caller edges earlier so that cold caller edges are first, we would have |
| 5067 | // visited and cloned this node for any unamibiguously cold non-recursive |
| 5068 | // callers before any ambiguous backedge callers. Note that we don't do this |
| 5069 | // if the caller is already cloned or visited during cloning (e.g. via a |
| 5070 | // different context path from the allocation). |
| 5071 | // TODO: Can we do better in the case where the caller was already visited? |
| 5072 | if (CallerEdge->IsBackedge && !CallerEdge->Caller->CloneOf && |
| 5073 | !Visited.count(CallerEdge->Caller)) { |
| 5074 | const auto OrigIdCount = CallerEdge->getContextIds().size(); |
| 5075 | // Now do the recursive cloning of this backedge's caller, which was |
| 5076 | // deferred earlier. |
| 5077 | identifyClones(CallerEdge->Caller, Visited, CallerEdgeContextsForAlloc); |
| 5078 | removeNoneTypeCalleeEdges(Node: CallerEdge->Caller); |
| 5079 | // See if the recursive call to identifyClones moved the context ids to a |
| 5080 | // new edge from this node to a clone of caller, and switch to looking at |
| 5081 | // that new edge so that we clone Node for the new caller clone. |
| 5082 | bool UpdatedEdge = false; |
| 5083 | if (OrigIdCount > CallerEdge->getContextIds().size()) { |
| 5084 | for (auto E : Node->CallerEdges) { |
| 5085 | // Only interested in clones of the current edges caller. |
| 5086 | if (E->Caller->CloneOf != CallerEdge->Caller) |
| 5087 | continue; |
| 5088 | // See if this edge contains any of the context ids originally on the |
| 5089 | // current caller edge. |
| 5090 | auto CallerEdgeContextsForAllocNew = |
| 5091 | set_intersection(CallerEdgeContextsForAlloc, E->getContextIds()); |
| 5092 | if (CallerEdgeContextsForAllocNew.empty()) |
| 5093 | continue; |
| 5094 | // Make sure we don't pick a previously existing caller edge of this |
| 5095 | // Node, which would be processed on a different iteration of the |
| 5096 | // outer loop over the saved CallerEdges. |
| 5097 | if (llvm::is_contained(CallerEdges, E)) |
| 5098 | continue; |
| 5099 | // The CallerAllocTypeForAlloc and CalleeEdgeAllocTypesForCallerEdge |
| 5100 | // are updated further below for all cases where we just invoked |
| 5101 | // identifyClones recursively. |
| 5102 | CallerEdgeContextsForAlloc.swap(CallerEdgeContextsForAllocNew); |
| 5103 | CallerEdge = E; |
| 5104 | UpdatedEdge = true; |
| 5105 | break; |
| 5106 | } |
| 5107 | } |
| 5108 | // If cloning removed this edge (and we didn't update it to a new edge |
| 5109 | // above), we're done with this edge. It's possible we moved all of the |
| 5110 | // context ids to an existing clone, in which case there's no need to do |
| 5111 | // further processing for them. |
| 5112 | if (CallerEdge->isRemoved()) |
| 5113 | continue; |
| 5114 | |
| 5115 | // Now we need to update the information used for the cloning decisions |
| 5116 | // further below, as we may have modified edges and their context ids. |
| 5117 | |
| 5118 | // Note if we changed the CallerEdge above we would have already updated |
| 5119 | // the context ids. |
| 5120 | if (!UpdatedEdge) { |
| 5121 | CallerEdgeContextsForAlloc = set_intersection( |
| 5122 | CallerEdgeContextsForAlloc, CallerEdge->getContextIds()); |
| 5123 | if (CallerEdgeContextsForAlloc.empty()) |
| 5124 | continue; |
| 5125 | } |
| 5126 | // Update the other information that depends on the edges and on the now |
| 5127 | // updated CallerEdgeContextsForAlloc. |
| 5128 | CallerAllocTypeForAlloc = computeAllocType(ContextIds&: CallerEdgeContextsForAlloc); |
| 5129 | CalleeEdgeAllocTypesForCallerEdge.clear(); |
| 5130 | for (auto &CalleeEdge : Node->CalleeEdges) { |
| 5131 | CalleeEdgeAllocTypesForCallerEdge.push_back(intersectAllocTypes( |
| 5132 | Node1Ids: CalleeEdge->getContextIds(), Node2Ids: CallerEdgeContextsForAlloc)); |
| 5133 | } |
| 5134 | } |
| 5135 | |
| 5136 | // First see if we can use an existing clone. Check each clone and its |
| 5137 | // callee edges for matching alloc types. |
| 5138 | ContextNode *Clone = nullptr; |
| 5139 | for (auto *CurClone : Node->Clones) { |
| 5140 | if (allocTypeToUse(CurClone->AllocTypes) != |
| 5141 | allocTypeToUse(CallerAllocTypeForAlloc)) |
| 5142 | continue; |
| 5143 | |
| 5144 | bool BothSingleAlloc = hasSingleAllocType(CurClone->AllocTypes) && |
| 5145 | hasSingleAllocType(CallerAllocTypeForAlloc); |
| 5146 | // The above check should mean that if both have single alloc types that |
| 5147 | // they should be equal. |
| 5148 | assert(!BothSingleAlloc || |
| 5149 | CurClone->AllocTypes == CallerAllocTypeForAlloc); |
| 5150 | |
| 5151 | // If either both have a single alloc type (which are the same), or if the |
| 5152 | // clone's callee edges have the same alloc types as those for the current |
| 5153 | // allocation on Node's callee edges (CalleeEdgeAllocTypesForCallerEdge), |
| 5154 | // then we can reuse this clone. |
| 5155 | if (BothSingleAlloc || allocTypesMatchClone<DerivedCCG, FuncTy, CallTy>( |
| 5156 | CalleeEdgeAllocTypesForCallerEdge, CurClone)) { |
| 5157 | Clone = CurClone; |
| 5158 | break; |
| 5159 | } |
| 5160 | } |
| 5161 | |
| 5162 | // The edge iterator is adjusted when we move the CallerEdge to the clone. |
| 5163 | if (Clone) |
| 5164 | moveEdgeToExistingCalleeClone(Edge: CallerEdge, NewCallee: Clone, /*NewClone=*/false, |
| 5165 | ContextIdsToMove: CallerEdgeContextsForAlloc); |
| 5166 | else |
| 5167 | Clone = moveEdgeToNewCalleeClone(Edge: CallerEdge, ContextIdsToMove: CallerEdgeContextsForAlloc); |
| 5168 | |
| 5169 | // Sanity check that no alloc types on clone or its edges are None. |
| 5170 | assert(Clone->AllocTypes != (uint8_t)AllocationType::None); |
| 5171 | } |
| 5172 | |
| 5173 | // We should still have some context ids on the original Node. |
| 5174 | assert(!Node->emptyContextIds()); |
| 5175 | |
| 5176 | // Sanity check that no alloc types on node or edges are None. |
| 5177 | assert(Node->AllocTypes != (uint8_t)AllocationType::None); |
| 5178 | |
| 5179 | if (VerifyNodes) |
| 5180 | checkNode<DerivedCCG, FuncTy, CallTy>(Node, /*CheckEdges=*/false); |
| 5181 | } |
| 5182 | |
| 5183 | void ModuleCallsiteContextGraph::updateAllocationCall( |
| 5184 | CallInfo &Call, AllocationType AllocType) { |
| 5185 | std::string AllocTypeString = getAllocTypeAttributeString(Type: AllocType); |
| 5186 | auto *CB = cast<CallBase>(Val: Call.call()); |
| 5187 | if (isAmbiguousAllocType(AllocTypes: (uint8_t)AllocType)) { |
| 5188 | // The contexts reaching this allocation could not be disambiguated, so it |
| 5189 | // keeps (or gets) the ambiguous attribute added during matching rather than |
| 5190 | // the default not cold attribute. |
| 5191 | addAmbiguousAttribute(CB); |
| 5192 | } else { |
| 5193 | removeAnyExistingAmbiguousAttribute(CB); |
| 5194 | auto A = llvm::Attribute::get(Context&: CB->getFunction()->getContext(), Kind: "memprof" , |
| 5195 | Val: AllocTypeString); |
| 5196 | CB->addFnAttr(Attr: A); |
| 5197 | } |
| 5198 | OREGetter(Call.call()->getFunction()) |
| 5199 | .emit(OptDiag: OptimizationRemark(DEBUG_TYPE, "MemprofAttribute" , Call.call()) |
| 5200 | << ore::NV("AllocationCall" , Call.call()) << " in clone " |
| 5201 | << ore::NV("Caller" , Call.call()->getFunction()) |
| 5202 | << " marked with memprof allocation attribute " |
| 5203 | << ore::NV("Attribute" , AllocTypeString)); |
| 5204 | } |
| 5205 | |
| 5206 | void IndexCallsiteContextGraph::updateAllocationCall(CallInfo &Call, |
| 5207 | AllocationType AllocType) { |
| 5208 | auto *AI = cast<AllocInfo *>(Val: Call.call()); |
| 5209 | assert(AI); |
| 5210 | assert(AI->Versions.size() > Call.cloneNo()); |
| 5211 | AI->Versions[Call.cloneNo()] = (uint8_t)AllocType; |
| 5212 | } |
| 5213 | |
| 5214 | AllocationType |
| 5215 | ModuleCallsiteContextGraph::getAllocationCallType(const CallInfo &Call) const { |
| 5216 | const auto *CB = cast<CallBase>(Val: Call.call()); |
| 5217 | if (!CB->getAttributes().hasFnAttr(Kind: "memprof" )) |
| 5218 | return AllocationType::None; |
| 5219 | StringRef AllocTypeString = |
| 5220 | CB->getAttributes().getFnAttr(Kind: "memprof" ).getValueAsString(); |
| 5221 | if (AllocTypeString == "cold" ) |
| 5222 | return AllocationType::Cold; |
| 5223 | if (AllocTypeString == "ambiguous" ) |
| 5224 | return (AllocationType)((uint8_t)AllocationType::NotCold | |
| 5225 | (uint8_t)AllocationType::Cold); |
| 5226 | return AllocationType::NotCold; |
| 5227 | } |
| 5228 | |
| 5229 | AllocationType |
| 5230 | IndexCallsiteContextGraph::getAllocationCallType(const CallInfo &Call) const { |
| 5231 | const auto *AI = cast<AllocInfo *>(Val: Call.call()); |
| 5232 | assert(AI->Versions.size() > Call.cloneNo()); |
| 5233 | return (AllocationType)AI->Versions[Call.cloneNo()]; |
| 5234 | } |
| 5235 | |
| 5236 | void ModuleCallsiteContextGraph::updateCall(CallInfo &CallerCall, |
| 5237 | FuncInfo CalleeFunc) { |
| 5238 | auto *CurF = getCalleeFunc(Call: CallerCall.call()); |
| 5239 | auto NewCalleeCloneNo = CalleeFunc.cloneNo(); |
| 5240 | if (isMemProfClone(GV: *CurF)) { |
| 5241 | // If we already assigned this callsite to call a specific non-default |
| 5242 | // clone (i.e. not the original function which is clone 0), ensure that we |
| 5243 | // aren't trying to now update it to call a different clone, which is |
| 5244 | // indicative of a bug in the graph or function assignment. |
| 5245 | auto CurCalleeCloneNo = getMemProfCloneNum(F: *CurF); |
| 5246 | if (CurCalleeCloneNo != NewCalleeCloneNo) { |
| 5247 | LLVM_DEBUG(dbgs() << "Mismatch in call clone assignment: was " |
| 5248 | << CurCalleeCloneNo << " now " << NewCalleeCloneNo |
| 5249 | << "\n" ); |
| 5250 | MismatchedCloneAssignments++; |
| 5251 | } |
| 5252 | } |
| 5253 | if (NewCalleeCloneNo > 0) |
| 5254 | cast<CallBase>(Val: CallerCall.call())->setCalledFunction(CalleeFunc.func()); |
| 5255 | OREGetter(CallerCall.call()->getFunction()) |
| 5256 | .emit(OptDiag: OptimizationRemark(DEBUG_TYPE, "MemprofCall" , CallerCall.call()) |
| 5257 | << ore::NV("Call" , CallerCall.call()) << " in clone " |
| 5258 | << ore::NV("Caller" , CallerCall.call()->getFunction()) |
| 5259 | << " assigned to call function clone " |
| 5260 | << ore::NV("Callee" , CalleeFunc.func())); |
| 5261 | } |
| 5262 | |
| 5263 | void IndexCallsiteContextGraph::updateCall(CallInfo &CallerCall, |
| 5264 | FuncInfo CalleeFunc) { |
| 5265 | auto *CI = cast<CallsiteInfo *>(Val: CallerCall.call()); |
| 5266 | assert(CI && |
| 5267 | "Caller cannot be an allocation which should not have profiled calls" ); |
| 5268 | assert(CI->Clones.size() > CallerCall.cloneNo()); |
| 5269 | auto NewCalleeCloneNo = CalleeFunc.cloneNo(); |
| 5270 | auto &CurCalleeCloneNo = CI->Clones[CallerCall.cloneNo()]; |
| 5271 | // If we already assigned this callsite to call a specific non-default |
| 5272 | // clone (i.e. not the original function which is clone 0), ensure that we |
| 5273 | // aren't trying to now update it to call a different clone, which is |
| 5274 | // indicative of a bug in the graph or function assignment. |
| 5275 | if (CurCalleeCloneNo != 0 && CurCalleeCloneNo != NewCalleeCloneNo) { |
| 5276 | LLVM_DEBUG(dbgs() << "Mismatch in call clone assignment: was " |
| 5277 | << CurCalleeCloneNo << " now " << NewCalleeCloneNo |
| 5278 | << "\n" ); |
| 5279 | MismatchedCloneAssignments++; |
| 5280 | } |
| 5281 | CurCalleeCloneNo = NewCalleeCloneNo; |
| 5282 | } |
| 5283 | |
| 5284 | // Update the debug information attached to NewFunc to use the clone Name. Note |
| 5285 | // this needs to be done for both any existing DISubprogram for the definition, |
| 5286 | // as well as any separate declaration DISubprogram. |
| 5287 | static void updateSubprogramLinkageName(Function *NewFunc, StringRef Name) { |
| 5288 | assert(Name == NewFunc->getName()); |
| 5289 | auto *SP = NewFunc->getSubprogram(); |
| 5290 | if (!SP) |
| 5291 | return; |
| 5292 | auto *MDName = MDString::get(Context&: NewFunc->getParent()->getContext(), Str: Name); |
| 5293 | SP->replaceLinkageName(LN: MDName); |
| 5294 | DISubprogram *Decl = SP->getDeclaration(); |
| 5295 | if (!Decl) |
| 5296 | return; |
| 5297 | TempDISubprogram NewDecl = Decl->clone(); |
| 5298 | NewDecl->replaceLinkageName(LN: MDName); |
| 5299 | SP->replaceDeclaration(Decl: MDNode::replaceWithUniqued(N: std::move(NewDecl))); |
| 5300 | } |
| 5301 | |
| 5302 | CallsiteContextGraph<ModuleCallsiteContextGraph, Function, |
| 5303 | Instruction *>::FuncInfo |
| 5304 | ModuleCallsiteContextGraph::cloneFunctionForCallsite( |
| 5305 | FuncInfo &Func, CallInfo &Call, DenseMap<CallInfo, CallInfo> &CallMap, |
| 5306 | std::vector<CallInfo> &CallsWithMetadataInFunc, unsigned CloneNo) { |
| 5307 | // Use existing LLVM facilities for cloning and obtaining Call in clone |
| 5308 | ValueToValueMapTy VMap; |
| 5309 | auto *NewFunc = CloneFunction(F: Func.func(), VMap); |
| 5310 | // Clones are only called directly and never have their address taken, so |
| 5311 | // strip !kcfi_type to avoid emitting KCFI type preambles/symbols on them or |
| 5312 | // making them valid KCFI indirect-call targets. |
| 5313 | NewFunc->setMetadata(KindID: LLVMContext::MD_kcfi_type, Node: nullptr); |
| 5314 | std::string Name = getMemProfFuncName(Base: Func.func()->getName(), CloneNo); |
| 5315 | assert(!Func.func()->getParent()->getFunction(Name)); |
| 5316 | NewFunc->setName(Name); |
| 5317 | updateSubprogramLinkageName(NewFunc, Name); |
| 5318 | for (auto &Inst : CallsWithMetadataInFunc) { |
| 5319 | // This map always has the initial version in it. |
| 5320 | assert(Inst.cloneNo() == 0); |
| 5321 | CallMap[Inst] = {cast<Instruction>(Val&: VMap[Inst.call()]), CloneNo}; |
| 5322 | } |
| 5323 | OREGetter(Func.func()) |
| 5324 | .emit(OptDiag: OptimizationRemark(DEBUG_TYPE, "MemprofClone" , Func.func()) |
| 5325 | << "created clone " << ore::NV("NewFunction" , NewFunc)); |
| 5326 | return {NewFunc, CloneNo}; |
| 5327 | } |
| 5328 | |
| 5329 | CallsiteContextGraph<IndexCallsiteContextGraph, FunctionSummary, |
| 5330 | IndexCall>::FuncInfo |
| 5331 | IndexCallsiteContextGraph::cloneFunctionForCallsite( |
| 5332 | FuncInfo &Func, CallInfo &Call, DenseMap<CallInfo, CallInfo> &CallMap, |
| 5333 | std::vector<CallInfo> &CallsWithMetadataInFunc, unsigned CloneNo) { |
| 5334 | // Check how many clones we have of Call (and therefore function). |
| 5335 | // The next clone number is the current size of versions array. |
| 5336 | // Confirm this matches the CloneNo provided by the caller, which is based on |
| 5337 | // the number of function clones we have. |
| 5338 | assert(CloneNo == (isa<AllocInfo *>(Call.call()) |
| 5339 | ? cast<AllocInfo *>(Call.call())->Versions.size() |
| 5340 | : cast<CallsiteInfo *>(Call.call())->Clones.size())); |
| 5341 | // Walk all the instructions in this function. Create a new version for |
| 5342 | // each (by adding an entry to the Versions/Clones summary array), and copy |
| 5343 | // over the version being called for the function clone being cloned here. |
| 5344 | // Additionally, add an entry to the CallMap for the new function clone, |
| 5345 | // mapping the original call (clone 0, what is in CallsWithMetadataInFunc) |
| 5346 | // to the new call clone. |
| 5347 | for (auto &Inst : CallsWithMetadataInFunc) { |
| 5348 | // This map always has the initial version in it. |
| 5349 | assert(Inst.cloneNo() == 0); |
| 5350 | if (auto *AI = dyn_cast<AllocInfo *>(Val: Inst.call())) { |
| 5351 | assert(AI->Versions.size() == CloneNo); |
| 5352 | // We assign the allocation type later (in updateAllocationCall), just add |
| 5353 | // an entry for it here. |
| 5354 | AI->Versions.push_back(Elt: 0); |
| 5355 | } else { |
| 5356 | auto *CI = cast<CallsiteInfo *>(Val: Inst.call()); |
| 5357 | assert(CI && CI->Clones.size() == CloneNo); |
| 5358 | // We assign the clone number later (in updateCall), just add an entry for |
| 5359 | // it here. |
| 5360 | CI->Clones.push_back(Elt: 0); |
| 5361 | } |
| 5362 | CallMap[Inst] = {Inst.call(), CloneNo}; |
| 5363 | } |
| 5364 | return {Func.func(), CloneNo}; |
| 5365 | } |
| 5366 | |
| 5367 | // We perform cloning for each allocation node separately. However, this |
| 5368 | // sometimes results in a situation where the same node calls multiple |
| 5369 | // clones of the same callee, created for different allocations. This |
| 5370 | // causes issues when assigning functions to these clones, as each node can |
| 5371 | // in reality only call a single callee clone. |
| 5372 | // |
| 5373 | // To address this, before assigning functions, merge callee clone nodes as |
| 5374 | // needed using a post order traversal from the allocations. We attempt to |
| 5375 | // use existing clones as the merge node when legal, and to share them |
| 5376 | // among callers with the same properties (callers calling the same set of |
| 5377 | // callee clone nodes for the same allocations). |
| 5378 | // |
| 5379 | // Without this fix, in some cases incorrect function assignment will lead |
| 5380 | // to calling the wrong allocation clone. |
| 5381 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 5382 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::mergeClones() { |
| 5383 | if (!MergeClones) |
| 5384 | return; |
| 5385 | |
| 5386 | // Generate a map from context id to the associated allocation node for use |
| 5387 | // when merging clones. |
| 5388 | DenseMap<uint32_t, ContextNode *> ContextIdToAllocationNode; |
| 5389 | for (auto &Entry : AllocationCallToContextNodeMap) { |
| 5390 | auto *Node = Entry.second; |
| 5391 | for (auto Id : Node->getContextIds()) |
| 5392 | ContextIdToAllocationNode[Id] = Node->getOrigNode(); |
| 5393 | for (auto *Clone : Node->Clones) { |
| 5394 | for (auto Id : Clone->getContextIds()) |
| 5395 | ContextIdToAllocationNode[Id] = Clone->getOrigNode(); |
| 5396 | } |
| 5397 | } |
| 5398 | |
| 5399 | // Post order traversal starting from allocations to ensure each callsite |
| 5400 | // calls a single clone of its callee. Callee nodes that are clones of each |
| 5401 | // other are merged (via new merge nodes if needed) to achieve this. |
| 5402 | DenseSet<const ContextNode *> Visited; |
| 5403 | for (auto &Entry : AllocationCallToContextNodeMap) { |
| 5404 | auto *Node = Entry.second; |
| 5405 | |
| 5406 | mergeClones(Node, Visited, ContextIdToAllocationNode); |
| 5407 | |
| 5408 | // Make a copy so the recursive post order traversal that may create new |
| 5409 | // clones doesn't mess up iteration. Note that the recursive traversal |
| 5410 | // itself does not call mergeClones on any of these nodes, which are all |
| 5411 | // (clones of) allocations. |
| 5412 | auto Clones = Node->Clones; |
| 5413 | for (auto *Clone : Clones) |
| 5414 | mergeClones(Clone, Visited, ContextIdToAllocationNode); |
| 5415 | } |
| 5416 | |
| 5417 | if (DumpCCG) { |
| 5418 | dbgs() << "CCG after merging:\n" ; |
| 5419 | dbgs() << *this; |
| 5420 | } |
| 5421 | if (ExportToDot) |
| 5422 | exportToDot(Label: "aftermerge" ); |
| 5423 | |
| 5424 | if (VerifyCCG) { |
| 5425 | check(); |
| 5426 | } |
| 5427 | } |
| 5428 | |
| 5429 | // Recursive helper for above mergeClones method. |
| 5430 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 5431 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::mergeClones( |
| 5432 | ContextNode *Node, DenseSet<const ContextNode *> &Visited, |
| 5433 | DenseMap<uint32_t, ContextNode *> &ContextIdToAllocationNode) { |
| 5434 | auto Inserted = Visited.insert(Node); |
| 5435 | if (!Inserted.second) |
| 5436 | return; |
| 5437 | |
| 5438 | // Iteratively perform merging on this node to handle new caller nodes created |
| 5439 | // during the recursive traversal. We could do something more elegant such as |
| 5440 | // maintain a worklist, but this is a simple approach that doesn't cause a |
| 5441 | // measureable compile time effect, as most nodes don't have many caller |
| 5442 | // edges to check. |
| 5443 | bool FoundUnvisited = true; |
| 5444 | unsigned Iters = 0; |
| 5445 | while (FoundUnvisited) { |
| 5446 | Iters++; |
| 5447 | FoundUnvisited = false; |
| 5448 | // Make a copy since the recursive call may move a caller edge to a new |
| 5449 | // callee, messing up the iterator. |
| 5450 | auto CallerEdges = Node->CallerEdges; |
| 5451 | for (auto CallerEdge : CallerEdges) { |
| 5452 | // Skip any caller edge moved onto a different callee during recursion. |
| 5453 | if (CallerEdge->Callee != Node) |
| 5454 | continue; |
| 5455 | // If we found an unvisited caller, note that we should check the caller |
| 5456 | // edges again as mergeClones may add or change caller nodes. |
| 5457 | if (DoMergeIteration && !Visited.contains(CallerEdge->Caller)) |
| 5458 | FoundUnvisited = true; |
| 5459 | mergeClones(CallerEdge->Caller, Visited, ContextIdToAllocationNode); |
| 5460 | } |
| 5461 | } |
| 5462 | |
| 5463 | TotalMergeInvokes++; |
| 5464 | TotalMergeIters += Iters; |
| 5465 | if (Iters > MaxMergeIters) |
| 5466 | MaxMergeIters = Iters; |
| 5467 | |
| 5468 | // Merge for this node after we handle its callers. |
| 5469 | mergeNodeCalleeClones(Node, Visited, ContextIdToAllocationNode); |
| 5470 | } |
| 5471 | |
| 5472 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 5473 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::mergeNodeCalleeClones( |
| 5474 | ContextNode *Node, DenseSet<const ContextNode *> &Visited, |
| 5475 | DenseMap<uint32_t, ContextNode *> &ContextIdToAllocationNode) { |
| 5476 | // Ignore Node if we moved all of its contexts to clones. |
| 5477 | if (Node->emptyContextIds()) |
| 5478 | return; |
| 5479 | |
| 5480 | // First identify groups of clones among Node's callee edges, by building |
| 5481 | // a map from each callee base node to the associated callee edges from Node. |
| 5482 | MapVector<ContextNode *, std::vector<std::shared_ptr<ContextEdge>>> |
| 5483 | OrigNodeToCloneEdges; |
| 5484 | for (const auto &E : Node->CalleeEdges) { |
| 5485 | auto *Callee = E->Callee; |
| 5486 | if (!Callee->CloneOf && Callee->Clones.empty()) |
| 5487 | continue; |
| 5488 | ContextNode *Base = Callee->getOrigNode(); |
| 5489 | OrigNodeToCloneEdges[Base].push_back(E); |
| 5490 | } |
| 5491 | |
| 5492 | // Helper for callee edge sorting below. Return true if A's callee has fewer |
| 5493 | // caller edges than B, or if A is a clone and B is not, or if A's callee |
| 5494 | // node id is smaller than B's. |
| 5495 | auto CalleeCallerEdgeLessThan = [](const std::shared_ptr<ContextEdge> &A, |
| 5496 | const std::shared_ptr<ContextEdge> &B) { |
| 5497 | if (A->Callee->CallerEdges.size() != B->Callee->CallerEdges.size()) |
| 5498 | return A->Callee->CallerEdges.size() < B->Callee->CallerEdges.size(); |
| 5499 | if (A->Callee->CloneOf && !B->Callee->CloneOf) |
| 5500 | return true; |
| 5501 | else if (!A->Callee->CloneOf && B->Callee->CloneOf) |
| 5502 | return false; |
| 5503 | // Use the callee node id as a deterministic tie-breaker. |
| 5504 | return A->Callee->NodeId < B->Callee->NodeId; |
| 5505 | }; |
| 5506 | |
| 5507 | // Process each set of callee clones called by Node, performing the needed |
| 5508 | // merging. |
| 5509 | for (auto Entry : OrigNodeToCloneEdges) { |
| 5510 | // CalleeEdges is the set of edges from Node reaching callees that are |
| 5511 | // mutual clones of each other. |
| 5512 | auto &CalleeEdges = Entry.second; |
| 5513 | auto NumCalleeClones = CalleeEdges.size(); |
| 5514 | // A single edge means there is no merging needed. |
| 5515 | if (NumCalleeClones == 1) |
| 5516 | continue; |
| 5517 | // Sort the CalleeEdges calling this group of clones in ascending order of |
| 5518 | // their caller edge counts, putting the original non-clone node first in |
| 5519 | // cases of a tie. This simplifies finding an existing node to use as the |
| 5520 | // merge node. |
| 5521 | llvm::stable_sort(CalleeEdges, CalleeCallerEdgeLessThan); |
| 5522 | |
| 5523 | /// Find other callers of the given set of callee edges that can |
| 5524 | /// share the same callee merge node. See the comments at this method |
| 5525 | /// definition for details. |
| 5526 | DenseSet<ContextNode *> OtherCallersToShareMerge; |
| 5527 | findOtherCallersToShareMerge(Node, CalleeEdges, ContextIdToAllocationNode, |
| 5528 | OtherCallersToShareMerge); |
| 5529 | |
| 5530 | // Now do the actual merging. Identify existing or create a new MergeNode |
| 5531 | // during the first iteration. Move each callee over, along with edges from |
| 5532 | // other callers we've determined above can share the same merge node. |
| 5533 | ContextNode *MergeNode = nullptr; |
| 5534 | DenseMap<ContextNode *, unsigned> CallerToMoveCount; |
| 5535 | for (auto CalleeEdge : CalleeEdges) { |
| 5536 | auto *OrigCallee = CalleeEdge->Callee; |
| 5537 | // If we don't have a MergeNode yet (only happens on the first iteration, |
| 5538 | // as a new one will be created when we go to move the first callee edge |
| 5539 | // over as needed), see if we can use this callee. |
| 5540 | if (!MergeNode) { |
| 5541 | // If there are no other callers, simply use this callee. |
| 5542 | if (CalleeEdge->Callee->CallerEdges.size() == 1) { |
| 5543 | MergeNode = OrigCallee; |
| 5544 | NonNewMergedNodes++; |
| 5545 | continue; |
| 5546 | } |
| 5547 | // Otherwise, if we have identified other caller nodes that can share |
| 5548 | // the merge node with Node, see if all of OrigCallee's callers are |
| 5549 | // going to share the same merge node. In that case we can use callee |
| 5550 | // (since all of its callers would move to the new merge node). |
| 5551 | if (!OtherCallersToShareMerge.empty()) { |
| 5552 | bool MoveAllCallerEdges = true; |
| 5553 | for (auto CalleeCallerE : OrigCallee->CallerEdges) { |
| 5554 | if (CalleeCallerE == CalleeEdge) |
| 5555 | continue; |
| 5556 | if (!OtherCallersToShareMerge.contains(CalleeCallerE->Caller)) { |
| 5557 | MoveAllCallerEdges = false; |
| 5558 | break; |
| 5559 | } |
| 5560 | } |
| 5561 | // If we are going to move all callers over, we can use this callee as |
| 5562 | // the MergeNode. |
| 5563 | if (MoveAllCallerEdges) { |
| 5564 | MergeNode = OrigCallee; |
| 5565 | NonNewMergedNodes++; |
| 5566 | continue; |
| 5567 | } |
| 5568 | } |
| 5569 | } |
| 5570 | // Move this callee edge, creating a new merge node if necessary. |
| 5571 | if (MergeNode) { |
| 5572 | assert(MergeNode != OrigCallee); |
| 5573 | moveEdgeToExistingCalleeClone(Edge: CalleeEdge, NewCallee: MergeNode, |
| 5574 | /*NewClone*/ false); |
| 5575 | } else { |
| 5576 | MergeNode = moveEdgeToNewCalleeClone(Edge: CalleeEdge); |
| 5577 | NewMergedNodes++; |
| 5578 | } |
| 5579 | // Now move all identified edges from other callers over to the merge node |
| 5580 | // as well. |
| 5581 | if (!OtherCallersToShareMerge.empty()) { |
| 5582 | // Make and iterate over a copy of OrigCallee's caller edges because |
| 5583 | // some of these will be moved off of the OrigCallee and that would mess |
| 5584 | // up the iteration from OrigCallee. |
| 5585 | auto OrigCalleeCallerEdges = OrigCallee->CallerEdges; |
| 5586 | for (auto &CalleeCallerE : OrigCalleeCallerEdges) { |
| 5587 | if (CalleeCallerE == CalleeEdge) |
| 5588 | continue; |
| 5589 | if (!OtherCallersToShareMerge.contains(CalleeCallerE->Caller)) |
| 5590 | continue; |
| 5591 | CallerToMoveCount[CalleeCallerE->Caller]++; |
| 5592 | moveEdgeToExistingCalleeClone(Edge: CalleeCallerE, NewCallee: MergeNode, |
| 5593 | /*NewClone*/ false); |
| 5594 | } |
| 5595 | } |
| 5596 | removeNoneTypeCalleeEdges(Node: OrigCallee); |
| 5597 | removeNoneTypeCalleeEdges(Node: MergeNode); |
| 5598 | } |
| 5599 | } |
| 5600 | } |
| 5601 | |
| 5602 | // Look for other nodes that have edges to the same set of callee |
| 5603 | // clones as the current Node. Those can share the eventual merge node |
| 5604 | // (reducing cloning and binary size overhead) iff: |
| 5605 | // - they have edges to the same set of callee clones |
| 5606 | // - each callee edge reaches a subset of the same allocations as Node's |
| 5607 | // corresponding edge to the same callee clone. |
| 5608 | // The second requirement is to ensure that we don't undo any of the |
| 5609 | // necessary cloning to distinguish contexts with different allocation |
| 5610 | // behavior. |
| 5611 | // FIXME: This is somewhat conservative, as we really just need to ensure |
| 5612 | // that they don't reach the same allocations as contexts on edges from Node |
| 5613 | // going to any of the *other* callee clones being merged. However, that |
| 5614 | // requires more tracking and checking to get right. |
| 5615 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 5616 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>:: |
| 5617 | findOtherCallersToShareMerge( |
| 5618 | ContextNode *Node, |
| 5619 | std::vector<std::shared_ptr<ContextEdge>> &CalleeEdges, |
| 5620 | DenseMap<uint32_t, ContextNode *> &ContextIdToAllocationNode, |
| 5621 | DenseSet<ContextNode *> &OtherCallersToShareMerge) { |
| 5622 | auto NumCalleeClones = CalleeEdges.size(); |
| 5623 | // This map counts how many edges to the same callee clone exist for other |
| 5624 | // caller nodes of each callee clone. |
| 5625 | DenseMap<ContextNode *, unsigned> OtherCallersToSharedCalleeEdgeCount; |
| 5626 | // Counts the number of other caller nodes that have edges to all callee |
| 5627 | // clones that don't violate the allocation context checking. |
| 5628 | unsigned PossibleOtherCallerNodes = 0; |
| 5629 | |
| 5630 | // We only need to look at other Caller nodes if the first callee edge has |
| 5631 | // multiple callers (recall they are sorted in ascending order above). |
| 5632 | if (CalleeEdges[0]->Callee->CallerEdges.size() < 2) |
| 5633 | return; |
| 5634 | |
| 5635 | // For each callee edge: |
| 5636 | // - Collect the count of other caller nodes calling the same callees. |
| 5637 | // - Collect the alloc nodes reached by contexts on each callee edge. |
| 5638 | DenseMap<ContextEdge *, DenseSet<ContextNode *>> CalleeEdgeToAllocNodes; |
| 5639 | for (auto CalleeEdge : CalleeEdges) { |
| 5640 | assert(CalleeEdge->Callee->CallerEdges.size() > 1); |
| 5641 | // For each other caller of the same callee, increment the count of |
| 5642 | // edges reaching the same callee clone. |
| 5643 | for (auto CalleeCallerEdges : CalleeEdge->Callee->CallerEdges) { |
| 5644 | if (CalleeCallerEdges->Caller == Node) { |
| 5645 | assert(CalleeCallerEdges == CalleeEdge); |
| 5646 | continue; |
| 5647 | } |
| 5648 | OtherCallersToSharedCalleeEdgeCount[CalleeCallerEdges->Caller]++; |
| 5649 | // If this caller edge now reaches all of the same callee clones, |
| 5650 | // increment the count of candidate other caller nodes. |
| 5651 | if (OtherCallersToSharedCalleeEdgeCount[CalleeCallerEdges->Caller] == |
| 5652 | NumCalleeClones) |
| 5653 | PossibleOtherCallerNodes++; |
| 5654 | } |
| 5655 | // Collect the alloc nodes reached by contexts on each callee edge, for |
| 5656 | // later analysis. |
| 5657 | for (auto Id : CalleeEdge->getContextIds()) { |
| 5658 | auto *Alloc = ContextIdToAllocationNode.lookup(Id); |
| 5659 | if (!Alloc) { |
| 5660 | // FIXME: unclear why this happens occasionally, presumably |
| 5661 | // imperfect graph updates possibly with recursion. |
| 5662 | MissingAllocForContextId++; |
| 5663 | continue; |
| 5664 | } |
| 5665 | CalleeEdgeToAllocNodes[CalleeEdge.get()].insert(Alloc); |
| 5666 | } |
| 5667 | } |
| 5668 | |
| 5669 | // Now walk the callee edges again, and make sure that for each candidate |
| 5670 | // caller node all of its edges to the callees reach the same allocs (or |
| 5671 | // a subset) as those along the corresponding callee edge from Node. |
| 5672 | for (auto CalleeEdge : CalleeEdges) { |
| 5673 | assert(CalleeEdge->Callee->CallerEdges.size() > 1); |
| 5674 | // Stop if we do not have any (more) candidate other caller nodes. |
| 5675 | if (!PossibleOtherCallerNodes) |
| 5676 | break; |
| 5677 | auto &CurCalleeAllocNodes = CalleeEdgeToAllocNodes[CalleeEdge.get()]; |
| 5678 | // Check each other caller of this callee clone. |
| 5679 | for (auto &CalleeCallerE : CalleeEdge->Callee->CallerEdges) { |
| 5680 | // Not interested in the callee edge from Node itself. |
| 5681 | if (CalleeCallerE == CalleeEdge) |
| 5682 | continue; |
| 5683 | // Skip any callers that didn't have callee edges to all the same |
| 5684 | // callee clones. |
| 5685 | if (OtherCallersToSharedCalleeEdgeCount[CalleeCallerE->Caller] != |
| 5686 | NumCalleeClones) |
| 5687 | continue; |
| 5688 | // Make sure that each context along edge from candidate caller node |
| 5689 | // reaches an allocation also reached by this callee edge from Node. |
| 5690 | for (auto Id : CalleeCallerE->getContextIds()) { |
| 5691 | auto *Alloc = ContextIdToAllocationNode.lookup(Id); |
| 5692 | if (!Alloc) |
| 5693 | continue; |
| 5694 | // If not, simply reset the map entry to 0 so caller is ignored, and |
| 5695 | // reduce the count of candidate other caller nodes. |
| 5696 | if (!CurCalleeAllocNodes.contains(Alloc)) { |
| 5697 | OtherCallersToSharedCalleeEdgeCount[CalleeCallerE->Caller] = 0; |
| 5698 | PossibleOtherCallerNodes--; |
| 5699 | break; |
| 5700 | } |
| 5701 | } |
| 5702 | } |
| 5703 | } |
| 5704 | |
| 5705 | if (!PossibleOtherCallerNodes) |
| 5706 | return; |
| 5707 | |
| 5708 | // Build the set of other caller nodes that can use the same callee merge |
| 5709 | // node. |
| 5710 | for (auto &[OtherCaller, Count] : OtherCallersToSharedCalleeEdgeCount) { |
| 5711 | if (Count != NumCalleeClones) |
| 5712 | continue; |
| 5713 | OtherCallersToShareMerge.insert(OtherCaller); |
| 5714 | } |
| 5715 | } |
| 5716 | |
| 5717 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 5718 | const typename CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::ContextNode * |
| 5719 | CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::resolveImplicitSource( |
| 5720 | const ContextNode *Node, const ImplicitSource &Src) { |
| 5721 | if (!Src.Source->emptyContextIds()) |
| 5722 | return Src.Source; |
| 5723 | // Src.Source lost its contexts to another clone (e.g. its last caller was |
| 5724 | // moved to a new clone when specializing a function clone further). The |
| 5725 | // trimmed contexts stay together, so the clone carrying one of them is the |
| 5726 | // version the implicit caller requires. |
| 5727 | if (Node->hasContextId(Src.ContextId)) |
| 5728 | return Node; |
| 5729 | for (const ContextNode *C : Node->Clones) |
| 5730 | if (C->hasContextId(Src.ContextId)) |
| 5731 | return C; |
| 5732 | // Context ids only move between the clones of a node during function |
| 5733 | // assignment, so the id should still be on one of them. Return null, which |
| 5734 | // makes the call follow the original node, rather than an emptied clone. |
| 5735 | assert(false && "trimmed context id on no clone of its callsite node" ); |
| 5736 | return nullptr; |
| 5737 | } |
| 5738 | |
| 5739 | // Contexts whose call stacks were trimmed at a callsite node's frame are on |
| 5740 | // its callee edges but on none of its caller edges. When that node, or one |
| 5741 | // below it on the path of such a context, is cloned for the contexts of |
| 5742 | // another allocation reaching it through a caller edge, the trimmed contexts |
| 5743 | // stay with the node that was cloned, and the clone gets no edge to the callee |
| 5744 | // nodes the trimmed contexts flow through. The clone is therefore not a caller |
| 5745 | // of any clone of such a callee node, and would otherwise be assigned a |
| 5746 | // function clone of the callee without regard to it, possibly one calling the |
| 5747 | // wrong version of the callee for the trimmed contexts, which are agnostic to |
| 5748 | // the callers of the node they were trimmed at and apply to the clone equally. |
| 5749 | // |
| 5750 | // This is called once the clones of callsite Node have been assigned to |
| 5751 | // function clones. For each caller Y of a clone of Node whose edge carries |
| 5752 | // trimmed contexts, each clone K of Y's original node that has no edge to any |
| 5753 | // clone of Node, and which some of those trimmed contexts would reach (see |
| 5754 | // ImpliesContextIds), is an implicit caller of Node. For example, with a |
| 5755 | // context t trimmed at the frame of Y's callsite: |
| 5756 | // |
| 5757 | // (callers of Y) (callers of K) |
| 5758 | // | | |
| 5759 | // Y K clones of the same callsite node |
| 5760 | // | \ |
| 5761 | // t | \ K has no edge to any clone of Node |
| 5762 | // v v |
| 5763 | // Required (other callsites in Node's function) |
| 5764 | // (clone of Node called |
| 5765 | // by Y, assigned to |
| 5766 | // Y's function clone) |
| 5767 | // |
| 5768 | // Had t not been trimmed, it would also be on an edge from K, so K should get |
| 5769 | // the same version of Node's function as Y: |
| 5770 | // - If K was not yet assigned a function clone, assign it Y's, so that the |
| 5771 | // assignment of the callsites it does have edges to takes it into account |
| 5772 | // (creating further function clones derived from it if needed). |
| 5773 | // - If K was already assigned to a function clone without a clone of Node, |
| 5774 | // record that the Node call in that function clone should follow Required. |
| 5775 | // - Otherwise K's function clone calls a conflicting clone of Node, which is |
| 5776 | // left as is (and counted). |
| 5777 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 5778 | void CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::assignImplicitCallers( |
| 5779 | ContextNode *Node, const DenseSet<uint32_t> &TrimmedContextIds, |
| 5780 | DenseMap<ContextNode *, FuncInfo> &CallsiteToCalleeFuncCloneMap, |
| 5781 | const std::map<FuncInfo, ContextNode *> &FuncCloneToCurNodeCloneMap, |
| 5782 | ImplicitSourceMap &ImplicitSources) { |
| 5783 | auto HasEdgeToNodeFamily = [&](const ContextNode *N) { |
| 5784 | return llvm::any_of(N->CalleeEdges, |
| 5785 | [&](const std::shared_ptr<ContextEdge> &E) { |
| 5786 | return E->Callee->getOrigNode() == Node; |
| 5787 | }); |
| 5788 | }; |
| 5789 | // Returns true if any of the trimmed context ids Ids, which are on (the |
| 5790 | // clone) Y but not on K, a clone of the same callsite node, would reach K had |
| 5791 | // they not been trimmed: |
| 5792 | // - An id terminating at Y applies to all clones of Y's original node (see |
| 5793 | // the example above assignImplicitCallers). |
| 5794 | // - An id reaching Y from a caller Z applies to K only if it would equally |
| 5795 | // apply to some caller C of K that is a clone of Z's original node, i.e. |
| 5796 | // it is not on C (if it were, C explicitly directs it to a callee other |
| 5797 | // than K) and, recursively, it applies to C the way it applies to Z (e.g. |
| 5798 | // because it terminates at Z, so it applies to all clones of Z's original |
| 5799 | // node). The id would then flow from C to K just as it flows from Z to Y, |
| 5800 | // had it not been trimmed: |
| 5801 | // |
| 5802 | // ... ... t terminates at Z or one of its callers |
| 5803 | // | |
| 5804 | // Z C clones of the same callsite node |
| 5805 | // | : |
| 5806 | // t | : (t) t is not on C or K; it applies to them |
| 5807 | // v v because it was trimmed above |
| 5808 | // Y K clones of the same callsite node |
| 5809 | // | |
| 5810 | // t v |
| 5811 | // Required |
| 5812 | // |
| 5813 | // This is checked recursively on (C, Z) with the ids not on C. If C == Z, |
| 5814 | // Z explicitly directs the ids to Y rather than K, so they don't apply to |
| 5815 | // K. |
| 5816 | // |
| 5817 | // Visited records the ids already examined for each (K, Y) pair, which |
| 5818 | // guards against cycles from recursion while still examining ids first seen |
| 5819 | // for the pair on a reconverging caller chain. |
| 5820 | auto ImpliesContextIds = |
| 5821 | [&](ContextNode *K, ContextNode *Y, const DenseSet<uint32_t> &Ids, |
| 5822 | DenseMap<std::pair<ContextNode *, ContextNode *>, DenseSet<uint32_t>> |
| 5823 | &Visited, |
| 5824 | auto &&ImpliesContextIds) -> bool { |
| 5825 | assert(K->getOrigNode() == Y->getOrigNode()); |
| 5826 | DenseSet<uint32_t> &VisitedIds = Visited[{K, Y}]; |
| 5827 | DenseSet<uint32_t> NewIds; |
| 5828 | for (auto Id : Ids) |
| 5829 | if (VisitedIds.insert(V: Id).second) |
| 5830 | NewIds.insert(V: Id); |
| 5831 | if (NewIds.empty()) |
| 5832 | return false; |
| 5833 | // The ids found on each caller edge of Y, with that edge's caller Z, and |
| 5834 | // their union (a set rather than a count, as a recursive context can put |
| 5835 | // the same id on several caller edges). |
| 5836 | SmallVector<std::pair<ContextNode *, DenseSet<uint32_t>>> IdsFromCallers; |
| 5837 | DenseSet<uint32_t> IdsOnCallerEdges; |
| 5838 | for (auto &YEdge : Y->CallerEdges) { |
| 5839 | DenseSet<uint32_t> IdsFromZ = |
| 5840 | set_intersection(NewIds, YEdge->getContextIds()); |
| 5841 | if (IdsFromZ.empty()) |
| 5842 | continue; |
| 5843 | set_union(S1&: IdsOnCallerEdges, S2: IdsFromZ); |
| 5844 | IdsFromCallers.push_back({YEdge->Caller, std::move(IdsFromZ)}); |
| 5845 | } |
| 5846 | // Any ids on none of Y's caller edges terminate at Y, and so apply to all |
| 5847 | // clones of Y's original node, including K. |
| 5848 | if (IdsOnCallerEdges.size() < NewIds.size()) |
| 5849 | return true; |
| 5850 | // Otherwise check whether the ids reaching Y from a caller Z also apply to |
| 5851 | // K through a caller C of K that is a clone of Z's original node. |
| 5852 | for (auto &[Z, IdsFromZ] : IdsFromCallers) { |
| 5853 | for (auto &KEdge : K->CallerEdges) { |
| 5854 | ContextNode *C = KEdge->Caller; |
| 5855 | if (C == Z || C->getOrigNode() != Z->getOrigNode()) |
| 5856 | continue; |
| 5857 | // Only the ids not on C can apply to C implicitly (C directs the ones |
| 5858 | // it has elsewhere); recursively check whether they do. |
| 5859 | DenseSet<uint32_t> IdsNotOnC; |
| 5860 | for (auto Id : IdsFromZ) |
| 5861 | if (!C->hasContextId(Id)) |
| 5862 | IdsNotOnC.insert(Id); |
| 5863 | if (!IdsNotOnC.empty() && |
| 5864 | ImpliesContextIds(C, Z, IdsNotOnC, Visited, ImpliesContextIds)) |
| 5865 | return true; |
| 5866 | } |
| 5867 | } |
| 5868 | return false; |
| 5869 | }; |
| 5870 | // Returns true if the given clones of Node are interchangeable for an |
| 5871 | // implicit caller: the same clone, allocations with the same hint, or |
| 5872 | // callsites already assigned to call the same function clone. |
| 5873 | auto SameVersion = [&](const ContextNode *A, const ContextNode *B) { |
| 5874 | if (A == B) |
| 5875 | return true; |
| 5876 | if (Node->IsAllocation) |
| 5877 | return allocTypeToUse(A->AllocTypes) == allocTypeToUse(B->AllocTypes); |
| 5878 | auto AIt = CallsiteToCalleeFuncCloneMap.find(A); |
| 5879 | auto BIt = CallsiteToCalleeFuncCloneMap.find(B); |
| 5880 | return AIt != CallsiteToCalleeFuncCloneMap.end() && |
| 5881 | BIt != CallsiteToCalleeFuncCloneMap.end() && |
| 5882 | AIt->second == BIt->second; |
| 5883 | }; |
| 5884 | // The clones already identified as implicit callers of Node, so that each is |
| 5885 | // handled once (for the first caller edge whose trimmed contexts reach it). |
| 5886 | DenseSet<const ContextNode *> ImplicitCallers; |
| 5887 | // Handles the caller edge Edge of a clone of Node (or of Node itself): finds |
| 5888 | // the implicit callers among the clones of its caller Y's original node, and |
| 5889 | // assigns them, or records the clone of Node their calls must follow, as |
| 5890 | // described above. |
| 5891 | auto ProcessCaller = [&](const std::shared_ptr<ContextEdge> &Edge) { |
| 5892 | ContextNode *Y = Edge->Caller; |
| 5893 | ContextNode *Orig = Y->getOrigNode(); |
| 5894 | if (!Y->hasCall() || Orig->Clones.empty()) |
| 5895 | return; |
| 5896 | DenseSet<uint32_t> TrimmedIds = |
| 5897 | set_intersection(Edge->getContextIds(), TrimmedContextIds); |
| 5898 | if (TrimmedIds.empty()) |
| 5899 | return; |
| 5900 | auto YIt = CallsiteToCalleeFuncCloneMap.find(Y); |
| 5901 | if (YIt == CallsiteToCalleeFuncCloneMap.end()) |
| 5902 | return; |
| 5903 | // Copy, as recording assignments below may grow the map. |
| 5904 | const FuncInfo YFuncClone = YIt->second; |
| 5905 | // The clone of Node called by Y, which must be assigned to the same |
| 5906 | // function clone. |
| 5907 | auto RequiredIt = FuncCloneToCurNodeCloneMap.find(YFuncClone); |
| 5908 | assert(RequiredIt != FuncCloneToCurNodeCloneMap.end()); |
| 5909 | const ContextNode *Required = RequiredIt->second; |
| 5910 | // Record in ImplicitSources that the Node call in function clone F must |
| 5911 | // follow the callsite clone Src. This is consulted when the calls in |
| 5912 | // function clones without their own clone of Node are collected (see |
| 5913 | // UnassignedCallClones in assignFunctions), and makes the requirement |
| 5914 | // survive if Src is later moved to another function clone derived from F |
| 5915 | // (when one of its callers conflicts at another callsite), after which the |
| 5916 | // call in F would otherwise follow the function clone F was derived from. |
| 5917 | auto RecordImplicitSource = [&](const FuncInfo &F, const ContextNode *Src) { |
| 5918 | // The calls in the original function (clone 0) are not collected there, |
| 5919 | // so nothing can be recorded for it. |
| 5920 | if (F.cloneNo() != 0) |
| 5921 | ImplicitSources.try_emplace({Node, F.cloneNo()}, |
| 5922 | ImplicitSource{Src, *TrimmedIds.begin()}); |
| 5923 | }; |
| 5924 | // Check each clone of Y's original node (and the original itself) without |
| 5925 | // an edge to any clone of Node. |
| 5926 | for (ContextNode *K : llvm::concat<ContextNode *const>( |
| 5927 | Orig->Clones, ArrayRef<ContextNode *>(Orig))) { |
| 5928 | if (K == Y || !K->hasCall() || K->emptyContextIds() || |
| 5929 | HasEdgeToNodeFamily(K) || ImplicitCallers.contains(K)) |
| 5930 | continue; |
| 5931 | DenseMap<std::pair<ContextNode *, ContextNode *>, DenseSet<uint32_t>> |
| 5932 | Visited; |
| 5933 | if (!ImpliesContextIds(K, Y, TrimmedIds, Visited, ImpliesContextIds)) |
| 5934 | continue; |
| 5935 | // K is an implicit caller of Node. |
| 5936 | ImplicitCallers.insert(K); |
| 5937 | auto KIt = CallsiteToCalleeFuncCloneMap.find(K); |
| 5938 | if (KIt == CallsiteToCalleeFuncCloneMap.end()) { |
| 5939 | // Not yet assigned: give it Y's function clone, in which the Node call |
| 5940 | // then follows Required. |
| 5941 | CallsiteToCalleeFuncCloneMap[K] = YFuncClone; |
| 5942 | RecordImplicitSource(YFuncClone, Required); |
| 5943 | NumImplicitCallersAssigned++; |
| 5944 | continue; |
| 5945 | } |
| 5946 | const FuncInfo KFuncClone = KIt->second; |
| 5947 | if (KFuncClone == YFuncClone) { |
| 5948 | // Already in Y's function clone. |
| 5949 | RecordImplicitSource(YFuncClone, Required); |
| 5950 | continue; |
| 5951 | } |
| 5952 | unsigned KCloneNo = KFuncClone.cloneNo(); |
| 5953 | auto KNodeCloneIt = FuncCloneToCurNodeCloneMap.find(KFuncClone); |
| 5954 | if (KNodeCloneIt == FuncCloneToCurNodeCloneMap.end()) { |
| 5955 | // K's function clone has no clone of Node. |
| 5956 | if (KCloneNo == 0) { |
| 5957 | // The call in the original function keeps calling the original |
| 5958 | // callee function, or stays unhinted for an allocation. That |
| 5959 | // satisfies the requirement only if Required is that version. |
| 5960 | if (Node->IsAllocation) { |
| 5961 | if (allocTypeToUse(Required->AllocTypes) == AllocationType::NotCold) |
| 5962 | continue; |
| 5963 | } else if (auto It = CallsiteToCalleeFuncCloneMap.find(Required); |
| 5964 | It != CallsiteToCalleeFuncCloneMap.end() && |
| 5965 | It->second.cloneNo() == 0) |
| 5966 | continue; |
| 5967 | } else { |
| 5968 | // Record that the Node call in K's function clone must follow |
| 5969 | // Required, unless an earlier implicit caller already recorded a |
| 5970 | // conflicting clone. |
| 5971 | auto [It, Inserted] = ImplicitSources.try_emplace( |
| 5972 | {Node, KCloneNo}, ImplicitSource{Required, *TrimmedIds.begin()}); |
| 5973 | if (Inserted) |
| 5974 | NumImplicitCallersSourced++; |
| 5975 | if (SameVersion(It->second.Source, Required)) |
| 5976 | continue; |
| 5977 | } |
| 5978 | } else if (SameVersion(KNodeCloneIt->second, Required)) { |
| 5979 | // K's function clone has a clone of Node interchangeable with |
| 5980 | // Required. |
| 5981 | RecordImplicitSource(KFuncClone, KNodeCloneIt->second); |
| 5982 | continue; |
| 5983 | } |
| 5984 | // K's function clone calls a conflicting clone of Node. |
| 5985 | NumImplicitCallersUnresolved++; |
| 5986 | } |
| 5987 | }; |
| 5988 | // Process the caller edges of Node itself and of each of its clones: any of |
| 5989 | // them may carry trimmed contexts (a trimmed context stays with the clone |
| 5990 | // that was cloned, which may be Node or a clone made earlier). |
| 5991 | for (auto &Edge : Node->CallerEdges) |
| 5992 | ProcessCaller(Edge); |
| 5993 | for (auto *Clone : Node->Clones) |
| 5994 | for (auto &Edge : Clone->CallerEdges) |
| 5995 | ProcessCaller(Edge); |
| 5996 | } |
| 5997 | |
| 5998 | // This method assigns cloned callsites to functions, cloning the functions as |
| 5999 | // needed. The assignment is greedy and proceeds roughly as follows: |
| 6000 | // |
| 6001 | // For each function Func: |
| 6002 | // For each call with graph Node having clones: |
| 6003 | // Initialize ClonesWorklist to Node and its clones |
| 6004 | // Initialize NodeCloneCount to 0 |
| 6005 | // While ClonesWorklist is not empty: |
| 6006 | // Clone = pop front ClonesWorklist |
| 6007 | // NodeCloneCount++ |
| 6008 | // If Func has been cloned less than NodeCloneCount times: |
| 6009 | // If NodeCloneCount is 1: |
| 6010 | // Assign Clone to original Func |
| 6011 | // Continue |
| 6012 | // Create a new function clone |
| 6013 | // If other callers not assigned to call a function clone yet: |
| 6014 | // Assign them to call new function clone |
| 6015 | // Continue |
| 6016 | // Assign any other caller calling the cloned version to new clone |
| 6017 | // |
| 6018 | // For each caller of Clone: |
| 6019 | // If caller is assigned to call a specific function clone: |
| 6020 | // If we cannot assign Clone to that function clone: |
| 6021 | // Create new callsite Clone NewClone |
| 6022 | // Add NewClone to ClonesWorklist |
| 6023 | // Continue |
| 6024 | // Assign Clone to existing caller's called function clone |
| 6025 | // Else: |
| 6026 | // If Clone not already assigned to a function clone: |
| 6027 | // Assign to first function clone without assignment |
| 6028 | // Assign caller to selected function clone |
| 6029 | // For each call with graph Node having clones: |
| 6030 | // If number func clones > number call's callsite Node clones: |
| 6031 | // Record func CallInfo clones without Node clone in UnassignedCallClones |
| 6032 | // For callsite Nodes in DFS order from allocations: |
| 6033 | // If IsAllocation: |
| 6034 | // Update allocation with alloc type |
| 6035 | // Else: |
| 6036 | // For Call, all MatchingCalls, and associated UnnassignedCallClones: |
| 6037 | // Update call to call recorded callee clone |
| 6038 | // |
| 6039 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 6040 | bool CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::assignFunctions() { |
| 6041 | bool Changed = false; |
| 6042 | |
| 6043 | mergeClones(); |
| 6044 | |
| 6045 | // Context ids terminating at a non-allocation node which has or had callers |
| 6046 | // (i.e. is not a root), meaning their call stacks were trimmed there. Used |
| 6047 | // to identify implicit callers (see assignImplicitCallers). |
| 6048 | DenseSet<uint32_t> TrimmedContextIds; |
| 6049 | for (auto &NodePtr : NodeOwner) { |
| 6050 | ContextNode *N = NodePtr.get(); |
| 6051 | if (N->IsAllocation || N->CalleeEdges.empty() || |
| 6052 | (N->CallerEdges.empty() && N->Clones.empty() && !N->CloneOf)) |
| 6053 | continue; |
| 6054 | // A context id trimmed at N is on a callee edge but on no caller edge. |
| 6055 | // Compare distinct ids rather than edge sizes: a recursive context can put |
| 6056 | // the same id on several callee and caller edges of N. |
| 6057 | DenseSet<uint32_t> CallerIds; |
| 6058 | for (auto &E : N->CallerEdges) |
| 6059 | set_union(CallerIds, E->getContextIds()); |
| 6060 | for (auto &CalleeEdge : N->CalleeEdges) |
| 6061 | for (auto Id : CalleeEdge->getContextIds()) |
| 6062 | if (!CallerIds.contains(V: Id)) |
| 6063 | TrimmedContextIds.insert(Id); |
| 6064 | } |
| 6065 | NumTrimmedContextIds += TrimmedContextIds.size(); |
| 6066 | |
| 6067 | // Keep track of the assignment of nodes (callsites) to function clones they |
| 6068 | // call. |
| 6069 | DenseMap<ContextNode *, FuncInfo> CallsiteToCalleeFuncCloneMap; |
| 6070 | |
| 6071 | // Update caller node to call function version CalleeFunc, by recording the |
| 6072 | // assignment in CallsiteToCalleeFuncCloneMap. |
| 6073 | auto RecordCalleeFuncOfCallsite = [&](ContextNode *Caller, |
| 6074 | const FuncInfo &CalleeFunc) { |
| 6075 | assert(Caller->hasCall()); |
| 6076 | CallsiteToCalleeFuncCloneMap[Caller] = CalleeFunc; |
| 6077 | }; |
| 6078 | |
| 6079 | // Information for a single clone of this Func. |
| 6080 | struct FuncCloneInfo { |
| 6081 | // The function clone. |
| 6082 | FuncInfo FuncClone; |
| 6083 | // Remappings of each call of interest (from original uncloned call to the |
| 6084 | // corresponding cloned call in this function clone). |
| 6085 | DenseMap<CallInfo, CallInfo> CallMap; |
| 6086 | }; |
| 6087 | |
| 6088 | // Map to keep track of information needed to update calls in function clones |
| 6089 | // when their corresponding callsite node was not itself cloned for that |
| 6090 | // function clone. Because of call context trimming (i.e. we only keep as much |
| 6091 | // caller information as needed to distinguish hot vs cold), we may not have |
| 6092 | // caller edges coming to each callsite node from all possible function |
| 6093 | // callers. A function clone may get created for other callsites in the |
| 6094 | // function for which there are caller edges that were not trimmed. Any other |
| 6095 | // callsites in that function clone, which were not themselved cloned for |
| 6096 | // that function clone, should get updated the same way as the corresponding |
| 6097 | // callsite in the original function (which may call a clone of its callee). |
| 6098 | // |
| 6099 | // We build this map after completing function cloning for each function, so |
| 6100 | // that we can record the information from its call maps before they are |
| 6101 | // destructed. The map will be used as we update calls to update any still |
| 6102 | // unassigned call clones. Note that we may create new node clones as we clone |
| 6103 | // other functions, so later on we check which node clones were still not |
| 6104 | // created. To this end, the inner map is a map from function clone number to |
| 6105 | // the list of calls cloned for that function (can be more than one due to the |
| 6106 | // Node's MatchingCalls array). |
| 6107 | // |
| 6108 | // The alternative is creating new callsite clone nodes below as we clone the |
| 6109 | // function, but that is tricker to get right and likely more overhead. |
| 6110 | // |
| 6111 | // Each entry also records the callsite node whose callee assignment the |
| 6112 | // calls should follow (Source). By default that is the callsite clone (if |
| 6113 | // any) in the function clone the function clone was derived from, following |
| 6114 | // the chain of derivations to the original function, since a function clone |
| 6115 | // created to specialize another one should otherwise behave like it. It can |
| 6116 | // also be the callsite clone required by an implicit caller (see |
| 6117 | // assignImplicitCallers). A null Source means the original node. |
| 6118 | // |
| 6119 | // Inner map is a std::map so sorted by key (clone number), in order to get |
| 6120 | // ordered remarks in the full LTO case. |
| 6121 | struct UnassignedCallCloneInfo { |
| 6122 | SmallVector<CallInfo, 0> Calls; |
| 6123 | const ContextNode *Source = nullptr; |
| 6124 | }; |
| 6125 | DenseMap<const ContextNode *, std::map<unsigned, UnassignedCallCloneInfo>> |
| 6126 | UnassignedCallClones; |
| 6127 | |
| 6128 | // Walk all functions for which we saw calls with memprof metadata, and handle |
| 6129 | // cloning for each of its calls. |
| 6130 | for (auto &[Func, CallsWithMetadata] : FuncToCallsWithMetadata) { |
| 6131 | FuncInfo OrigFunc(Func); |
| 6132 | // Map from each clone number of OrigFunc to information about that function |
| 6133 | // clone (the function clone FuncInfo and call remappings). The index into |
| 6134 | // the vector is the clone number, as function clones are created and |
| 6135 | // numbered sequentially. |
| 6136 | std::vector<FuncCloneInfo> FuncCloneInfos; |
| 6137 | // For each function clone number, the clone number of the function clone |
| 6138 | // it was derived from (created to further specialize, when a caller |
| 6139 | // already assigned to it needed a different clone of some callsite), or 0 |
| 6140 | // for clones created from the original function. Used to decide which |
| 6141 | // callsite clone the calls not themselves cloned for a function clone |
| 6142 | // should follow (see UnassignedCallClones). |
| 6143 | std::vector<unsigned> FuncCloneParent; |
| 6144 | // Callsite clones required for calls in specific function clones by |
| 6145 | // implicit callers (see assignImplicitCallers). |
| 6146 | ImplicitSourceMap ImplicitSources; |
| 6147 | // Map from each original callsite node in this function to its original |
| 6148 | // call and matching calls (clone 0). The calls recorded on an original node |
| 6149 | // are updated to the calls in a function clone if that node is assigned to |
| 6150 | // a function clone, after which they can no longer be used to index the |
| 6151 | // function clone CallMaps (which are keyed by the original calls). No node |
| 6152 | // in this function has been assigned yet at this point. |
| 6153 | struct OrigCallsInfo { |
| 6154 | CallInfo Call; |
| 6155 | SmallVector<CallInfo, 0> MatchingCalls; |
| 6156 | }; |
| 6157 | DenseMap<const ContextNode *, OrigCallsInfo> OrigNodeToOrigCalls; |
| 6158 | for (auto &Call : CallsWithMetadata) { |
| 6159 | ContextNode *Node = getNodeForInst(C: Call); |
| 6160 | if (Node && Node->Call == Call) |
| 6161 | OrigNodeToOrigCalls[Node] = {Call, Node->MatchingCalls}; |
| 6162 | } |
| 6163 | for (auto &Call : CallsWithMetadata) { |
| 6164 | ContextNode *Node = getNodeForInst(C: Call); |
| 6165 | // Skip call if we do not have a node for it (all uses of its stack ids |
| 6166 | // were either on inlined chains or trimmed from the MIBs), or if we did |
| 6167 | // not create any clones for it. |
| 6168 | if (!Node || Node->Clones.empty()) |
| 6169 | continue; |
| 6170 | assert(Node->hasCall() && |
| 6171 | "Not having a call should have prevented cloning" ); |
| 6172 | |
| 6173 | // Track the assignment of function clones to clones of the current |
| 6174 | // callsite Node being handled. |
| 6175 | std::map<FuncInfo, ContextNode *> FuncCloneToCurNodeCloneMap; |
| 6176 | |
| 6177 | // Assign callsite version CallsiteClone to function version FuncClone, |
| 6178 | // and also assign (possibly cloned) Call to CallsiteClone. |
| 6179 | auto AssignCallsiteCloneToFuncClone = [&](const FuncInfo &FuncClone, |
| 6180 | CallInfo &Call, |
| 6181 | ContextNode *CallsiteClone, |
| 6182 | bool IsAlloc) { |
| 6183 | // Record the clone of callsite node assigned to this function clone. |
| 6184 | FuncCloneToCurNodeCloneMap[FuncClone] = CallsiteClone; |
| 6185 | |
| 6186 | assert(FuncCloneInfos.size() > FuncClone.cloneNo()); |
| 6187 | DenseMap<CallInfo, CallInfo> &CallMap = |
| 6188 | FuncCloneInfos[FuncClone.cloneNo()].CallMap; |
| 6189 | CallInfo CallClone(Call); |
| 6190 | if (auto It = CallMap.find(Call); It != CallMap.end()) |
| 6191 | CallClone = It->second; |
| 6192 | CallsiteClone->setCall(CallClone); |
| 6193 | // Need to do the same for all matching calls. Rebuild the list of |
| 6194 | // CallsiteClone from the original matching calls (it may have been |
| 6195 | // copied from a node whose calls were already remapped), using the |
| 6196 | // same mapping as for Call. |
| 6197 | assert(OrigNodeToOrigCalls.count(Node)); |
| 6198 | CallsiteClone->MatchingCalls.clear(); |
| 6199 | for (const CallInfo &OrigMatchingCall : |
| 6200 | OrigNodeToOrigCalls[Node].MatchingCalls) { |
| 6201 | CallInfo CallClone(OrigMatchingCall); |
| 6202 | if (auto It = CallMap.find(OrigMatchingCall); It != CallMap.end()) |
| 6203 | CallClone = It->second; |
| 6204 | CallsiteClone->MatchingCalls.push_back(CallClone); |
| 6205 | } |
| 6206 | }; |
| 6207 | |
| 6208 | // Invokes moveEdgeToNewCalleeClone which creates a new clone, and then |
| 6209 | // performs the necessary fixups (removing none type edges, and |
| 6210 | // importantly, propagating any function call assignment of the original |
| 6211 | // node to the new clone). |
| 6212 | auto MoveEdgeToNewCalleeCloneAndSetUp = |
| 6213 | [&](const std::shared_ptr<ContextEdge> &Edge) { |
| 6214 | ContextNode *OrigCallee = Edge->Callee; |
| 6215 | ContextNode *NewClone = moveEdgeToNewCalleeClone(Edge); |
| 6216 | removeNoneTypeCalleeEdges(Node: NewClone); |
| 6217 | assert(NewClone->AllocTypes != (uint8_t)AllocationType::None); |
| 6218 | // If the original Callee was already assigned to call a specific |
| 6219 | // function version, make sure its new clone is assigned to call |
| 6220 | // that same function clone. |
| 6221 | if (CallsiteToCalleeFuncCloneMap.count(OrigCallee)) |
| 6222 | RecordCalleeFuncOfCallsite( |
| 6223 | NewClone, CallsiteToCalleeFuncCloneMap[OrigCallee]); |
| 6224 | return NewClone; |
| 6225 | }; |
| 6226 | |
| 6227 | // Keep track of the clones of callsite Node that need to be assigned to |
| 6228 | // function clones. This list may be expanded in the loop body below if we |
| 6229 | // find additional cloning is required. |
| 6230 | std::deque<ContextNode *> ClonesWorklist; |
| 6231 | // Ignore original Node if we moved all of its contexts to clones. |
| 6232 | if (!Node->emptyContextIds()) |
| 6233 | ClonesWorklist.push_back(Node); |
| 6234 | llvm::append_range(ClonesWorklist, Node->Clones); |
| 6235 | |
| 6236 | // Now walk through all of the clones of this callsite Node that we need, |
| 6237 | // and determine the assignment to a corresponding clone of the current |
| 6238 | // function (creating new function clones as needed). |
| 6239 | unsigned NodeCloneCount = 0; |
| 6240 | while (!ClonesWorklist.empty()) { |
| 6241 | ContextNode *Clone = ClonesWorklist.front(); |
| 6242 | ClonesWorklist.pop_front(); |
| 6243 | NodeCloneCount++; |
| 6244 | if (VerifyNodes) |
| 6245 | checkNode<DerivedCCG, FuncTy, CallTy>(Clone); |
| 6246 | |
| 6247 | // Need to create a new function clone if we have more callsite clones |
| 6248 | // than existing function clones, which would have been assigned to an |
| 6249 | // earlier clone in the list (we assign callsite clones to function |
| 6250 | // clones greedily). |
| 6251 | if (FuncCloneInfos.size() < NodeCloneCount) { |
| 6252 | // If this is the first callsite copy, assign to original function. |
| 6253 | if (NodeCloneCount == 1) { |
| 6254 | // Since FuncCloneInfos is empty in this case, no clones have |
| 6255 | // been created for this function yet, and no callers should have |
| 6256 | // been assigned a function clone for this callee node yet. |
| 6257 | assert(llvm::none_of( |
| 6258 | Clone->CallerEdges, [&](const std::shared_ptr<ContextEdge> &E) { |
| 6259 | return CallsiteToCalleeFuncCloneMap.count(E->Caller); |
| 6260 | })); |
| 6261 | // Initialize with empty call map, assign Clone to original function |
| 6262 | // and its callers, and skip to the next clone. |
| 6263 | FuncCloneInfos.push_back( |
| 6264 | {OrigFunc, DenseMap<CallInfo, CallInfo>()}); |
| 6265 | FuncCloneParent.push_back(x: 0); |
| 6266 | AssignCallsiteCloneToFuncClone( |
| 6267 | OrigFunc, Call, Clone, |
| 6268 | AllocationCallToContextNodeMap.count(Call)); |
| 6269 | for (auto &CE : Clone->CallerEdges) { |
| 6270 | // Ignore any caller that does not have a recorded callsite Call. |
| 6271 | if (!CE->Caller->hasCall()) |
| 6272 | continue; |
| 6273 | RecordCalleeFuncOfCallsite(CE->Caller, OrigFunc); |
| 6274 | } |
| 6275 | continue; |
| 6276 | } |
| 6277 | |
| 6278 | // First locate which copy of OrigFunc to clone again. If a caller |
| 6279 | // of this callsite clone was already assigned to call a particular |
| 6280 | // function clone, we need to redirect all of those callers to the |
| 6281 | // new function clone, and update their other callees within this |
| 6282 | // function. |
| 6283 | FuncInfo PreviousAssignedFuncClone; |
| 6284 | auto EI = llvm::find_if( |
| 6285 | Clone->CallerEdges, [&](const std::shared_ptr<ContextEdge> &E) { |
| 6286 | return CallsiteToCalleeFuncCloneMap.count(E->Caller); |
| 6287 | }); |
| 6288 | bool CallerAssignedToCloneOfFunc = false; |
| 6289 | if (EI != Clone->CallerEdges.end()) { |
| 6290 | const std::shared_ptr<ContextEdge> &Edge = *EI; |
| 6291 | PreviousAssignedFuncClone = |
| 6292 | CallsiteToCalleeFuncCloneMap[Edge->Caller]; |
| 6293 | CallerAssignedToCloneOfFunc = true; |
| 6294 | } |
| 6295 | |
| 6296 | // Clone function and save it along with the CallInfo map created |
| 6297 | // during cloning in the FuncCloneInfos. |
| 6298 | DenseMap<CallInfo, CallInfo> NewCallMap; |
| 6299 | unsigned CloneNo = FuncCloneInfos.size(); |
| 6300 | assert(CloneNo > 0 && "Clone 0 is the original function, which " |
| 6301 | "should already exist in the map" ); |
| 6302 | FuncInfo NewFuncClone = cloneFunctionForCallsite( |
| 6303 | Func&: OrigFunc, Call, CallMap&: NewCallMap, CallsWithMetadataInFunc&: CallsWithMetadata, CloneNo); |
| 6304 | FuncCloneInfos.push_back({NewFuncClone, std::move(NewCallMap)}); |
| 6305 | // Record the function clone it is derived from (if any), whose |
| 6306 | // callers it will take over below. |
| 6307 | FuncCloneParent.push_back(CallerAssignedToCloneOfFunc |
| 6308 | ? PreviousAssignedFuncClone.cloneNo() |
| 6309 | : 0); |
| 6310 | FunctionClonesAnalysis++; |
| 6311 | Changed = true; |
| 6312 | |
| 6313 | // If no caller callsites were already assigned to a clone of this |
| 6314 | // function, we can simply assign this clone to the new func clone |
| 6315 | // and update all callers to it, then skip to the next clone. |
| 6316 | if (!CallerAssignedToCloneOfFunc) { |
| 6317 | AssignCallsiteCloneToFuncClone( |
| 6318 | NewFuncClone, Call, Clone, |
| 6319 | AllocationCallToContextNodeMap.count(Call)); |
| 6320 | for (auto &CE : Clone->CallerEdges) { |
| 6321 | // Ignore any caller that does not have a recorded callsite Call. |
| 6322 | if (!CE->Caller->hasCall()) |
| 6323 | continue; |
| 6324 | RecordCalleeFuncOfCallsite(CE->Caller, NewFuncClone); |
| 6325 | } |
| 6326 | continue; |
| 6327 | } |
| 6328 | |
| 6329 | // We may need to do additional node cloning in this case. |
| 6330 | // Reset the CallsiteToCalleeFuncCloneMap entry for any callers |
| 6331 | // that were previously assigned to call PreviousAssignedFuncClone, |
| 6332 | // to record that they now call NewFuncClone. |
| 6333 | // The none type edge removal may remove some of this Clone's caller |
| 6334 | // edges, if it is reached via another of its caller's callees. |
| 6335 | // Iterate over a copy and skip any that were removed. |
| 6336 | auto CallerEdges = Clone->CallerEdges; |
| 6337 | for (auto CE : CallerEdges) { |
| 6338 | // Skip any that have been removed on an earlier iteration. |
| 6339 | if (CE->isRemoved()) { |
| 6340 | assert(!is_contained(Clone->CallerEdges, CE)); |
| 6341 | continue; |
| 6342 | } |
| 6343 | assert(CE); |
| 6344 | // Ignore any caller that does not have a recorded callsite Call. |
| 6345 | if (!CE->Caller->hasCall()) |
| 6346 | continue; |
| 6347 | |
| 6348 | if (!CallsiteToCalleeFuncCloneMap.count(CE->Caller) || |
| 6349 | // We subsequently fall through to later handling that |
| 6350 | // will perform any additional cloning required for |
| 6351 | // callers that were calling other function clones. |
| 6352 | CallsiteToCalleeFuncCloneMap[CE->Caller] != |
| 6353 | PreviousAssignedFuncClone) |
| 6354 | continue; |
| 6355 | |
| 6356 | RecordCalleeFuncOfCallsite(CE->Caller, NewFuncClone); |
| 6357 | |
| 6358 | // If we are cloning a function that was already assigned to some |
| 6359 | // callers, then essentially we are creating new callsite clones |
| 6360 | // of the other callsites in that function that are reached by those |
| 6361 | // callers. Clone the other callees of the current callsite's caller |
| 6362 | // that were already assigned to PreviousAssignedFuncClone |
| 6363 | // accordingly. This is important since we subsequently update the |
| 6364 | // calls from the nodes in the graph and their assignments to callee |
| 6365 | // functions recorded in CallsiteToCalleeFuncCloneMap. |
| 6366 | // The none type edge removal may remove some of this caller's |
| 6367 | // callee edges, if it is reached via another of its callees. |
| 6368 | // Iterate over a copy and skip any that were removed. |
| 6369 | auto CalleeEdges = CE->Caller->CalleeEdges; |
| 6370 | for (auto CalleeEdge : CalleeEdges) { |
| 6371 | // Skip any that have been removed on an earlier iteration when |
| 6372 | // cleaning up newly None type callee edges. |
| 6373 | if (CalleeEdge->isRemoved()) { |
| 6374 | assert(!is_contained(CE->Caller->CalleeEdges, CalleeEdge)); |
| 6375 | continue; |
| 6376 | } |
| 6377 | assert(CalleeEdge); |
| 6378 | ContextNode *Callee = CalleeEdge->Callee; |
| 6379 | // Skip the current callsite, we are looking for other |
| 6380 | // callsites Caller calls, as well as any that does not have a |
| 6381 | // recorded callsite Call. |
| 6382 | if (Callee == Clone || !Callee->hasCall()) |
| 6383 | continue; |
| 6384 | // Skip direct recursive calls. We don't need/want to clone the |
| 6385 | // caller node again, and this loop will not behave as expected if |
| 6386 | // we tried. |
| 6387 | if (Callee == CalleeEdge->Caller) |
| 6388 | continue; |
| 6389 | ContextNode *NewClone = |
| 6390 | MoveEdgeToNewCalleeCloneAndSetUp(CalleeEdge); |
| 6391 | // Moving the edge may have resulted in some none type |
| 6392 | // callee edges on the original Callee. |
| 6393 | removeNoneTypeCalleeEdges(Node: Callee); |
| 6394 | // Update NewClone with the new Call clone of this callsite's Call |
| 6395 | // created for the new function clone created earlier. |
| 6396 | // Recall that we have already ensured when building the graph |
| 6397 | // that each caller can only call callsites within the same |
| 6398 | // function, so we are guaranteed that Callee Call is in the |
| 6399 | // current OrigFunc. |
| 6400 | // CallMap is set up as indexed by original calls at clone 0. Note |
| 6401 | // that we cannot simply use the calls recorded on Callee's |
| 6402 | // original node, as those are updated to the calls in a function |
| 6403 | // clone if the original node was itself assigned to a function |
| 6404 | // clone. |
| 6405 | assert(OrigNodeToOrigCalls.count(Callee->getOrigNode())); |
| 6406 | auto &OrigCalls = OrigNodeToOrigCalls[Callee->getOrigNode()]; |
| 6407 | DenseMap<CallInfo, CallInfo> &CallMap = |
| 6408 | FuncCloneInfos[NewFuncClone.cloneNo()].CallMap; |
| 6409 | assert(CallMap.count(OrigCalls.Call)); |
| 6410 | CallInfo NewCall(CallMap[OrigCalls.Call]); |
| 6411 | assert(NewCall); |
| 6412 | NewClone->setCall(NewCall); |
| 6413 | // Need to do the same for all matching calls. |
| 6414 | NewClone->MatchingCalls.clear(); |
| 6415 | for (auto &OrigMatchingCall : OrigCalls.MatchingCalls) { |
| 6416 | assert(CallMap.count(OrigMatchingCall)); |
| 6417 | CallInfo NewCall(CallMap[OrigMatchingCall]); |
| 6418 | assert(NewCall); |
| 6419 | NewClone->MatchingCalls.push_back(NewCall); |
| 6420 | } |
| 6421 | } |
| 6422 | } |
| 6423 | // Fall through to handling below to perform the recording of the |
| 6424 | // function for this callsite clone. This enables handling of cases |
| 6425 | // where the callers were assigned to different clones of a function. |
| 6426 | } |
| 6427 | |
| 6428 | auto FindFirstAvailFuncClone = [&]() { |
| 6429 | // Find first function in FuncCloneInfos without an assigned |
| 6430 | // clone of this callsite Node. We should always have one |
| 6431 | // available at this point due to the earlier cloning when the |
| 6432 | // FuncCloneInfos size was smaller than the clone number. |
| 6433 | for (auto &CF : FuncCloneInfos) { |
| 6434 | if (!FuncCloneToCurNodeCloneMap.count(CF.FuncClone)) |
| 6435 | return CF.FuncClone; |
| 6436 | } |
| 6437 | llvm_unreachable( |
| 6438 | "Expected an available func clone for this callsite clone" ); |
| 6439 | }; |
| 6440 | |
| 6441 | // See if we can use existing function clone. Walk through |
| 6442 | // all caller edges to see if any have already been assigned to |
| 6443 | // a clone of this callsite's function. If we can use it, do so. If not, |
| 6444 | // because that function clone is already assigned to a different clone |
| 6445 | // of this callsite, then we need to clone again. |
| 6446 | // Basically, this checking is needed to handle the case where different |
| 6447 | // caller functions/callsites may need versions of this function |
| 6448 | // containing different mixes of callsite clones across the different |
| 6449 | // callsites within the function. If that happens, we need to create |
| 6450 | // additional function clones to handle the various combinations. |
| 6451 | // |
| 6452 | // Keep track of any new clones of this callsite created by the |
| 6453 | // following loop, as well as any existing clone that we decided to |
| 6454 | // assign this clone to. |
| 6455 | std::map<FuncInfo, ContextNode *> FuncCloneToNewCallsiteCloneMap; |
| 6456 | FuncInfo FuncCloneAssignedToCurCallsiteClone; |
| 6457 | // Iterate over a copy of Clone's caller edges, since we may need to |
| 6458 | // remove edges in the moveEdgeTo* methods, and this simplifies the |
| 6459 | // handling and makes it less error-prone. |
| 6460 | auto CloneCallerEdges = Clone->CallerEdges; |
| 6461 | for (auto &Edge : CloneCallerEdges) { |
| 6462 | // Skip removed edges (due to direct recursive edges updated when |
| 6463 | // updating callee edges when moving an edge and subsequently |
| 6464 | // removed by call to removeNoneTypeCalleeEdges on the Clone). |
| 6465 | if (Edge->isRemoved()) |
| 6466 | continue; |
| 6467 | // Ignore any caller that does not have a recorded callsite Call. |
| 6468 | if (!Edge->Caller->hasCall()) |
| 6469 | continue; |
| 6470 | // If this caller already assigned to call a version of OrigFunc, need |
| 6471 | // to ensure we can assign this callsite clone to that function clone. |
| 6472 | if (CallsiteToCalleeFuncCloneMap.count(Edge->Caller)) { |
| 6473 | FuncInfo FuncCloneCalledByCaller = |
| 6474 | CallsiteToCalleeFuncCloneMap[Edge->Caller]; |
| 6475 | // First we need to confirm that this function clone is available |
| 6476 | // for use by this callsite node clone. |
| 6477 | // |
| 6478 | // While FuncCloneToCurNodeCloneMap is built only for this Node and |
| 6479 | // its callsite clones, one of those callsite clones X could have |
| 6480 | // been assigned to the same function clone called by Edge's caller |
| 6481 | // - if Edge's caller calls another callsite within Node's original |
| 6482 | // function, and that callsite has another caller reaching clone X. |
| 6483 | // We need to clone Node again in this case. |
| 6484 | if ((FuncCloneToCurNodeCloneMap.count(FuncCloneCalledByCaller) && |
| 6485 | FuncCloneToCurNodeCloneMap[FuncCloneCalledByCaller] != |
| 6486 | Clone) || |
| 6487 | // Detect when we have multiple callers of this callsite that |
| 6488 | // have already been assigned to specific, and different, clones |
| 6489 | // of OrigFunc (due to other unrelated callsites in Func they |
| 6490 | // reach via call contexts). Is this Clone of callsite Node |
| 6491 | // assigned to a different clone of OrigFunc? If so, clone Node |
| 6492 | // again. |
| 6493 | (FuncCloneAssignedToCurCallsiteClone && |
| 6494 | FuncCloneAssignedToCurCallsiteClone != |
| 6495 | FuncCloneCalledByCaller)) { |
| 6496 | // We need to use a different newly created callsite clone, in |
| 6497 | // order to assign it to another new function clone on a |
| 6498 | // subsequent iteration over the Clones array (adjusted below). |
| 6499 | // Note we specifically do not reset the |
| 6500 | // CallsiteToCalleeFuncCloneMap entry for this caller, so that |
| 6501 | // when this new clone is processed later we know which version of |
| 6502 | // the function to copy (so that other callsite clones we have |
| 6503 | // assigned to that function clone are properly cloned over). See |
| 6504 | // comments in the function cloning handling earlier. |
| 6505 | |
| 6506 | // Check if we already have cloned this callsite again while |
| 6507 | // walking through caller edges, for a caller calling the same |
| 6508 | // function clone. If so, we can move this edge to that new clone |
| 6509 | // rather than creating yet another new clone. |
| 6510 | if (FuncCloneToNewCallsiteCloneMap.count( |
| 6511 | FuncCloneCalledByCaller)) { |
| 6512 | ContextNode *NewClone = |
| 6513 | FuncCloneToNewCallsiteCloneMap[FuncCloneCalledByCaller]; |
| 6514 | moveEdgeToExistingCalleeClone(Edge, NewCallee: NewClone); |
| 6515 | // Cleanup any none type edges cloned over. |
| 6516 | removeNoneTypeCalleeEdges(Node: NewClone); |
| 6517 | } else { |
| 6518 | // Create a new callsite clone. |
| 6519 | ContextNode *NewClone = MoveEdgeToNewCalleeCloneAndSetUp(Edge); |
| 6520 | FuncCloneToNewCallsiteCloneMap[FuncCloneCalledByCaller] = |
| 6521 | NewClone; |
| 6522 | // Add to list of clones and process later. |
| 6523 | ClonesWorklist.push_back(NewClone); |
| 6524 | } |
| 6525 | // Moving the caller edge may have resulted in some none type |
| 6526 | // callee edges. |
| 6527 | removeNoneTypeCalleeEdges(Node: Clone); |
| 6528 | // We will handle the newly created callsite clone in a subsequent |
| 6529 | // iteration over this Node's Clones. |
| 6530 | continue; |
| 6531 | } |
| 6532 | |
| 6533 | // Otherwise, we can use the function clone already assigned to this |
| 6534 | // caller. |
| 6535 | if (!FuncCloneAssignedToCurCallsiteClone) { |
| 6536 | FuncCloneAssignedToCurCallsiteClone = FuncCloneCalledByCaller; |
| 6537 | // Assign Clone to FuncCloneCalledByCaller |
| 6538 | AssignCallsiteCloneToFuncClone( |
| 6539 | FuncCloneCalledByCaller, Call, Clone, |
| 6540 | AllocationCallToContextNodeMap.count(Call)); |
| 6541 | } else |
| 6542 | // Don't need to do anything - callsite is already calling this |
| 6543 | // function clone. |
| 6544 | assert(FuncCloneAssignedToCurCallsiteClone == |
| 6545 | FuncCloneCalledByCaller); |
| 6546 | |
| 6547 | } else { |
| 6548 | // We have not already assigned this caller to a version of |
| 6549 | // OrigFunc. Do the assignment now. |
| 6550 | |
| 6551 | // First check if we have already assigned this callsite clone to a |
| 6552 | // clone of OrigFunc for another caller during this iteration over |
| 6553 | // its caller edges. |
| 6554 | if (!FuncCloneAssignedToCurCallsiteClone) { |
| 6555 | FuncCloneAssignedToCurCallsiteClone = FindFirstAvailFuncClone(); |
| 6556 | assert(FuncCloneAssignedToCurCallsiteClone); |
| 6557 | // Assign Clone to FuncCloneAssignedToCurCallsiteClone |
| 6558 | AssignCallsiteCloneToFuncClone( |
| 6559 | FuncCloneAssignedToCurCallsiteClone, Call, Clone, |
| 6560 | AllocationCallToContextNodeMap.count(Call)); |
| 6561 | } else |
| 6562 | assert(FuncCloneToCurNodeCloneMap |
| 6563 | [FuncCloneAssignedToCurCallsiteClone] == Clone); |
| 6564 | // Update callers to record function version called. |
| 6565 | RecordCalleeFuncOfCallsite(Edge->Caller, |
| 6566 | FuncCloneAssignedToCurCallsiteClone); |
| 6567 | } |
| 6568 | } |
| 6569 | // If we didn't assign a function clone to this callsite clone yet, e.g. |
| 6570 | // none of its callers has a non-null call, do the assignment here. |
| 6571 | // We want to ensure that every callsite clone is assigned to some |
| 6572 | // function clone, so that the call updates below work as expected. |
| 6573 | // In particular if this is the original callsite, we want to ensure it |
| 6574 | // is assigned to the original function, otherwise the original function |
| 6575 | // will appear available for assignment to other callsite clones, |
| 6576 | // leading to unintended effects. For one, the unknown and not updated |
| 6577 | // callers will call into cloned paths leading to the wrong hints, |
| 6578 | // because they still call the original function (clone 0). Also, |
| 6579 | // because all callsites start out as being clone 0 by default, we can't |
| 6580 | // easily distinguish between callsites explicitly assigned to clone 0 |
| 6581 | // vs those never assigned, which can lead to multiple updates of the |
| 6582 | // calls when invoking updateCall below, with mismatched clone values. |
| 6583 | // TODO: Add a flag to the callsite nodes or some other mechanism to |
| 6584 | // better distinguish and identify callsite clones that are not getting |
| 6585 | // assigned to function clones as expected. |
| 6586 | if (!FuncCloneAssignedToCurCallsiteClone) { |
| 6587 | FuncCloneAssignedToCurCallsiteClone = FindFirstAvailFuncClone(); |
| 6588 | assert(FuncCloneAssignedToCurCallsiteClone && |
| 6589 | "No available func clone for this callsite clone" ); |
| 6590 | AssignCallsiteCloneToFuncClone( |
| 6591 | FuncCloneAssignedToCurCallsiteClone, Call, Clone, |
| 6592 | /*IsAlloc=*/AllocationCallToContextNodeMap.contains(Call)); |
| 6593 | } |
| 6594 | } |
| 6595 | assignImplicitCallers(Node, TrimmedContextIds, |
| 6596 | CallsiteToCalleeFuncCloneMap, |
| 6597 | FuncCloneToCurNodeCloneMap, ImplicitSources); |
| 6598 | if (VerifyCCG) { |
| 6599 | checkNode<DerivedCCG, FuncTy, CallTy>(Node); |
| 6600 | for (const auto &PE : Node->CalleeEdges) |
| 6601 | checkNode<DerivedCCG, FuncTy, CallTy>(PE->Callee); |
| 6602 | for (const auto &CE : Node->CallerEdges) |
| 6603 | checkNode<DerivedCCG, FuncTy, CallTy>(CE->Caller); |
| 6604 | for (auto *Clone : Node->Clones) { |
| 6605 | checkNode<DerivedCCG, FuncTy, CallTy>(Clone); |
| 6606 | for (const auto &PE : Clone->CalleeEdges) |
| 6607 | checkNode<DerivedCCG, FuncTy, CallTy>(PE->Callee); |
| 6608 | for (const auto &CE : Clone->CallerEdges) |
| 6609 | checkNode<DerivedCCG, FuncTy, CallTy>(CE->Caller); |
| 6610 | } |
| 6611 | } |
| 6612 | } |
| 6613 | |
| 6614 | if (MemProfReportHintedSizesDetail) |
| 6615 | FuncToNumClones[Func] = std::max<unsigned>(1, FuncCloneInfos.size()); |
| 6616 | if (FuncCloneInfos.size() < 2) |
| 6617 | continue; |
| 6618 | |
| 6619 | // In this case there is more than just the original function copy. |
| 6620 | // Record call clones of any callsite nodes in the function that did not |
| 6621 | // themselves get cloned for all of the function clones. |
| 6622 | for (auto &Call : CallsWithMetadata) { |
| 6623 | ContextNode *Node = getNodeForInst(C: Call); |
| 6624 | if (!Node || !Node->hasCall() || Node->emptyContextIds()) |
| 6625 | continue; |
| 6626 | // First collect the function clones that have a clone of this callsite |
| 6627 | // node (or the node itself). They may not be sequential due to empty |
| 6628 | // clones e.g. |
| 6629 | DenseMap<unsigned, const ContextNode *> FuncCloneToNodeClone; |
| 6630 | FuncCloneToNodeClone[Node->Call.cloneNo()] = Node; |
| 6631 | // A clone left without context ids (its callers were all moved to other |
| 6632 | // clones while assigning function clones) has no edges and is not |
| 6633 | // reached by the call updates below, so the call in its function clone |
| 6634 | // keeps the original callee. If an implicit caller recorded the callsite |
| 6635 | // clone that call must follow, treat the function clone as having no |
| 6636 | // clone of this node, so that the call follows the recorded source like |
| 6637 | // any other call not cloned for its function clone. |
| 6638 | for (auto *C : Node->Clones) |
| 6639 | if (!C->emptyContextIds() || |
| 6640 | !ImplicitSources.contains({Node, C->Call.cloneNo()})) |
| 6641 | FuncCloneToNodeClone[C->Call.cloneNo()] = C; |
| 6642 | // If Node has enough clones already to cover all function clones, we can |
| 6643 | // skip it. |
| 6644 | if (FuncCloneToNodeClone.size() >= FuncCloneInfos.size()) |
| 6645 | continue; |
| 6646 | // Now check all the function clones. |
| 6647 | for (auto &FC : FuncCloneInfos) { |
| 6648 | unsigned I = FC.FuncClone.cloneNo(); |
| 6649 | // Skip the first clone which got the original call. |
| 6650 | // Also skip any other clones created for this Node. |
| 6651 | if (I == 0 || FuncCloneToNodeClone.contains(I)) { |
| 6652 | continue; |
| 6653 | } |
| 6654 | // Record the call clones created for this callsite in this function |
| 6655 | // clone, and the callsite node whose callee they should follow: the |
| 6656 | // one required by an implicit caller, if any, else the clone of this |
| 6657 | // callsite node in the nearest function clone in the chain of function |
| 6658 | // clones this one was derived from (null for the original node). |
| 6659 | auto &Info = UnassignedCallClones[Node][I]; |
| 6660 | auto &CallVector = Info.Calls; |
| 6661 | for (unsigned P = I;; P = FuncCloneParent[P]) { |
| 6662 | if (auto It = ImplicitSources.find({Node, P}); |
| 6663 | It != ImplicitSources.end()) { |
| 6664 | Info.Source = resolveImplicitSource(Node, Src: It->second); |
| 6665 | break; |
| 6666 | } |
| 6667 | if (auto It = FuncCloneToNodeClone.find(P); |
| 6668 | It != FuncCloneToNodeClone.end()) { |
| 6669 | Info.Source = It->second; |
| 6670 | break; |
| 6671 | } |
| 6672 | if (P == 0) |
| 6673 | break; |
| 6674 | } |
| 6675 | DenseMap<CallInfo, CallInfo> &CallMap = FC.CallMap; |
| 6676 | if (auto It = CallMap.find(Call); It != CallMap.end()) { |
| 6677 | CallInfo CallClone = It->second; |
| 6678 | CallVector.push_back(CallClone); |
| 6679 | } else { |
| 6680 | // All but the original clone (skipped earlier) should have an entry |
| 6681 | // for all calls. |
| 6682 | assert(false && "Expected to find call in CallMap" ); |
| 6683 | } |
| 6684 | // Need to do the same for all matching calls. Use the original |
| 6685 | // matching calls, as those on Node may have been remapped to the calls |
| 6686 | // of a function clone if Node was assigned to one. |
| 6687 | assert(OrigNodeToOrigCalls.count(Node)); |
| 6688 | for (const CallInfo &OrigMatchingCall : |
| 6689 | OrigNodeToOrigCalls[Node].MatchingCalls) { |
| 6690 | if (auto It = CallMap.find(OrigMatchingCall); It != CallMap.end()) { |
| 6691 | CallInfo CallClone = It->second; |
| 6692 | CallVector.push_back(CallClone); |
| 6693 | } else { |
| 6694 | // All but the original clone (skipped earlier) should have an entry |
| 6695 | // for all calls. |
| 6696 | assert(false && "Expected to find call in CallMap" ); |
| 6697 | } |
| 6698 | } |
| 6699 | } |
| 6700 | } |
| 6701 | } |
| 6702 | |
| 6703 | uint8_t BothTypes = |
| 6704 | (uint8_t)AllocationType::Cold | (uint8_t)AllocationType::NotCold; |
| 6705 | |
| 6706 | auto UpdateCalls = [&](ContextNode *Node, |
| 6707 | DenseSet<const ContextNode *> &Visited, |
| 6708 | auto &&UpdateCalls) { |
| 6709 | auto Inserted = Visited.insert(Node); |
| 6710 | if (!Inserted.second) |
| 6711 | return; |
| 6712 | |
| 6713 | for (auto *Clone : Node->Clones) |
| 6714 | UpdateCalls(Clone, Visited, UpdateCalls); |
| 6715 | |
| 6716 | for (auto &Edge : Node->CallerEdges) |
| 6717 | UpdateCalls(Edge->Caller, Visited, UpdateCalls); |
| 6718 | |
| 6719 | // Skip if either no call to update, or if we ended up with no context ids |
| 6720 | // (we moved all edges onto other clones). |
| 6721 | if (!Node->hasCall() || Node->emptyContextIds()) |
| 6722 | return; |
| 6723 | |
| 6724 | // Returns the recorded call clones in function clones which don't have a |
| 6725 | // clone of this callsite node, along with the node whose decision each |
| 6726 | // should follow (the Source recorded earlier, falling back to Node). |
| 6727 | auto GetUnassignedCallClones = [&](ContextNode *Node) { |
| 6728 | SmallVector<std::pair<const ContextNode *, MutableArrayRef<CallInfo>>> |
| 6729 | Unassigned; |
| 6730 | auto UnassignedIt = UnassignedCallClones.find(Node); |
| 6731 | if (UnassignedIt == UnassignedCallClones.end()) |
| 6732 | return Unassigned; |
| 6733 | DenseSet<unsigned> NodeCallClones; |
| 6734 | for (auto *C : Node->Clones) |
| 6735 | if (!C->emptyContextIds()) |
| 6736 | NodeCallClones.insert(C->Call.cloneNo()); |
| 6737 | for (auto &[CloneNo, Info] : UnassignedIt->second) { |
| 6738 | // Should start at 1 as we never create an entry for original node. |
| 6739 | assert(CloneNo > 0); |
| 6740 | // If we subsequently created a clone, skip this one. |
| 6741 | if (NodeCallClones.contains(V: CloneNo)) |
| 6742 | continue; |
| 6743 | Unassigned.push_back({Info.Source ? Info.Source : Node, Info.Calls}); |
| 6744 | } |
| 6745 | return Unassigned; |
| 6746 | }; |
| 6747 | |
| 6748 | if (Node->IsAllocation) { |
| 6749 | auto GetAllocTypeToUse = [&](const ContextNode *Node) { |
| 6750 | auto AT = allocTypeToUse(Node->AllocTypes); |
| 6751 | // If the allocation type is still ambiguous after cloning, record |
| 6752 | // that as the allocation type (the attribute stays "ambiguous") rather |
| 6753 | // than falling back to the default not cold, unless ambiguous |
| 6754 | // attributes are disabled. |
| 6755 | if (Node->AllocTypes == BothTypes && MemProfUseAmbiguousAttributes) |
| 6756 | AT = (AllocationType)BothTypes; |
| 6757 | // If the allocation type is ambiguous, and more aggressive hinting |
| 6758 | // has been enabled via the MinClonedColdBytePercent flag, see if this |
| 6759 | // allocation should be hinted cold anyway because its fraction cold |
| 6760 | // bytes allocated is at least the given threshold. |
| 6761 | if (Node->AllocTypes == BothTypes && MinClonedColdBytePercent < 100 && |
| 6762 | !ContextIdToContextSizeInfos.empty()) { |
| 6763 | uint64_t TotalCold = 0; |
| 6764 | uint64_t Total = 0; |
| 6765 | for (auto Id : Node->getContextIds()) { |
| 6766 | auto TypeI = ContextIdToAllocationType.find(Id); |
| 6767 | assert(TypeI != ContextIdToAllocationType.end()); |
| 6768 | auto CSI = ContextIdToContextSizeInfos.find(Id); |
| 6769 | if (CSI != ContextIdToContextSizeInfos.end()) { |
| 6770 | for (auto &Info : CSI->second) { |
| 6771 | Total += Info.TotalSize; |
| 6772 | if (TypeI->second == AllocationType::Cold) |
| 6773 | TotalCold += Info.TotalSize; |
| 6774 | } |
| 6775 | } |
| 6776 | } |
| 6777 | if (TotalCold * 100 >= Total * MinClonedColdBytePercent) |
| 6778 | AT = AllocationType::Cold; |
| 6779 | } |
| 6780 | return AT; |
| 6781 | }; |
| 6782 | updateAllocationCall(Call&: Node->Call, AllocType: GetAllocTypeToUse(Node)); |
| 6783 | if (MemProfReportHintedSizesDetail) |
| 6784 | OrigAllocAndCloneToNode[{Node->getOrigNode(), Node->Call.cloneNo()}] = |
| 6785 | Node; |
| 6786 | assert(Node->MatchingCalls.empty()); |
| 6787 | // Update the allocation calls in function clones without a clone of this |
| 6788 | // node the same way as the node they should follow. |
| 6789 | for (auto &[Source, Calls] : GetUnassignedCallClones(Node)) { |
| 6790 | auto AT = GetAllocTypeToUse(Source); |
| 6791 | for (auto &Call : Calls) { |
| 6792 | updateAllocationCall(Call, AllocType: AT); |
| 6793 | // For the report, such an allocation clone follows Source's hint. |
| 6794 | if (MemProfReportHintedSizesDetail) |
| 6795 | OrigAllocAndCloneToNode[{Node->getOrigNode(), Call.cloneNo()}] = |
| 6796 | Source; |
| 6797 | } |
| 6798 | } |
| 6799 | return; |
| 6800 | } |
| 6801 | |
| 6802 | auto CalleeFuncIt = CallsiteToCalleeFuncCloneMap.find(Node); |
| 6803 | if (CalleeFuncIt != CallsiteToCalleeFuncCloneMap.end()) { |
| 6804 | auto &CalleeFunc = CalleeFuncIt->second; |
| 6805 | updateCall(CallerCall&: Node->Call, CalleeFunc); |
| 6806 | if (MemProfReportHintedSizesDetail) |
| 6807 | OrigNodeAndCloneToCalleeClone[{ |
| 6808 | Node->getOrigNode(), Node->Call.cloneNo()}] = CalleeFunc.cloneNo(); |
| 6809 | // Update all the matching calls as well. |
| 6810 | for (auto &Call : Node->MatchingCalls) |
| 6811 | updateCall(CallerCall&: Call, CalleeFunc); |
| 6812 | } |
| 6813 | |
| 6814 | // Now update all calls recorded earlier that are still in function clones |
| 6815 | // which don't have a clone of this callsite node, the same way as the |
| 6816 | // node they should follow. |
| 6817 | for (auto &[Source, Calls] : GetUnassignedCallClones(Node)) { |
| 6818 | auto It = CallsiteToCalleeFuncCloneMap.find(Source); |
| 6819 | // Fall back to the original Node's callee function. |
| 6820 | if (It == CallsiteToCalleeFuncCloneMap.end()) |
| 6821 | It = CalleeFuncIt; |
| 6822 | if (It == CallsiteToCalleeFuncCloneMap.end()) |
| 6823 | continue; |
| 6824 | for (auto &Call : Calls) { |
| 6825 | updateCall(CallerCall&: Call, CalleeFunc: It->second); |
| 6826 | if (MemProfReportHintedSizesDetail) |
| 6827 | OrigNodeAndCloneToCalleeClone[{Node->getOrigNode(), Call.cloneNo()}] = |
| 6828 | It->second.cloneNo(); |
| 6829 | } |
| 6830 | } |
| 6831 | }; |
| 6832 | |
| 6833 | // Performs DFS traversal starting from allocation nodes to update calls to |
| 6834 | // reflect cloning decisions recorded earlier. For regular LTO this will |
| 6835 | // update the actual calls in the IR to call the appropriate function clone |
| 6836 | // (and add attributes to allocation calls), whereas for ThinLTO the decisions |
| 6837 | // are recorded in the summary entries. |
| 6838 | DenseSet<const ContextNode *> Visited; |
| 6839 | for (auto &Entry : AllocationCallToContextNodeMap) |
| 6840 | UpdateCalls(Entry.second, Visited, UpdateCalls); |
| 6841 | |
| 6842 | return Changed; |
| 6843 | } |
| 6844 | |
| 6845 | // Compute a SHA1 hash of the callsite and alloc version information of clone I |
| 6846 | // in the summary, to use in detection of duplicate clones. |
| 6847 | uint64_t ComputeHash(const FunctionSummary *FS, unsigned I) { |
| 6848 | SHA1 Hasher; |
| 6849 | // Add a record for the callsite or alloc at the given index in the summary, |
| 6850 | // with the given version. The index must be included, since two clones that |
| 6851 | // assign the same versions to different callsites or allocs are not |
| 6852 | // duplicates. |
| 6853 | auto AddRecord = [&](char Kind, uint32_t Index, uint32_t Version) { |
| 6854 | uint8_t Data[1 + 2 * sizeof(uint32_t)]; |
| 6855 | Data[0] = Kind; |
| 6856 | support::endian::write32le(P: &Data[1], V: Index); |
| 6857 | support::endian::write32le(P: &Data[1 + sizeof(uint32_t)], V: Version); |
| 6858 | Hasher.update(Data); |
| 6859 | }; |
| 6860 | // Update hash with any callsites that call non-default (non-zero) callee |
| 6861 | // versions. |
| 6862 | for (auto [Index, SN] : enumerate(First: FS->callsites())) { |
| 6863 | // In theory all callsites and allocs in this function should have the same |
| 6864 | // number of clone entries, but handle any discrepancies gracefully below |
| 6865 | // for NDEBUG builds. |
| 6866 | assert( |
| 6867 | SN.Clones.size() > I && |
| 6868 | "Callsite summary has fewer entries than other summaries in function" ); |
| 6869 | if (SN.Clones.size() <= I || !SN.Clones[I]) |
| 6870 | continue; |
| 6871 | AddRecord('c', Index, SN.Clones[I]); |
| 6872 | } |
| 6873 | // Update hash with any allocs that have non-default (non-None) hints. |
| 6874 | for (auto [Index, AN] : enumerate(First: FS->allocs())) { |
| 6875 | // In theory all callsites and allocs in this function should have the same |
| 6876 | // number of clone entries, but handle any discrepancies gracefully below |
| 6877 | // for NDEBUG builds. |
| 6878 | assert(AN.Versions.size() > I && |
| 6879 | "Alloc summary has fewer entries than other summaries in function" ); |
| 6880 | if (AN.Versions.size() <= I || |
| 6881 | (AllocationType)AN.Versions[I] == AllocationType::None) |
| 6882 | continue; |
| 6883 | AddRecord('a', Index, AN.Versions[I]); |
| 6884 | } |
| 6885 | return support::endian::read64le(P: Hasher.result().data()); |
| 6886 | } |
| 6887 | |
| 6888 | #ifndef NDEBUG |
| 6889 | // Check that clones I and J of FS have identical callsite and alloc versions. |
| 6890 | static bool clonesHaveSameVersions(const FunctionSummary *FS, unsigned I, |
| 6891 | unsigned J) { |
| 6892 | for (auto &SN : FS->callsites()) { |
| 6893 | assert(SN.Clones.size() > I && SN.Clones.size() > J && |
| 6894 | "Callsite summary has fewer entries than other summaries in " |
| 6895 | "function" ); |
| 6896 | if (SN.Clones[I] != SN.Clones[J]) |
| 6897 | return false; |
| 6898 | } |
| 6899 | for (auto &AN : FS->allocs()) { |
| 6900 | assert(AN.Versions.size() > I && AN.Versions.size() > J && |
| 6901 | "Alloc summary has fewer entries than other summaries in function" ); |
| 6902 | if (AN.Versions[I] != AN.Versions[J]) |
| 6903 | return false; |
| 6904 | } |
| 6905 | return true; |
| 6906 | } |
| 6907 | #endif |
| 6908 | |
| 6909 | static SmallVector<std::unique_ptr<ValueToValueMapTy>, 4> ( |
| 6910 | Function &F, unsigned NumClones, Module &M, OptimizationRemarkEmitter &ORE, |
| 6911 | std::map<const Function *, SmallPtrSet<const GlobalAlias *, 1>> |
| 6912 | &FuncToAliasMap, |
| 6913 | FunctionSummary *FS) { |
| 6914 | auto TakeDeclNameAndReplace = [](GlobalValue *DeclGV, GlobalValue *NewGV) { |
| 6915 | // We might have created this when adjusting callsite in another |
| 6916 | // function. It should be a declaration. |
| 6917 | assert(DeclGV->isDeclaration()); |
| 6918 | NewGV->takeName(V: DeclGV); |
| 6919 | DeclGV->replaceAllUsesWith(V: NewGV); |
| 6920 | DeclGV->eraseFromParent(); |
| 6921 | }; |
| 6922 | |
| 6923 | // Handle aliases to this function, and create analogous alias clones to the |
| 6924 | // provided clone of this function. |
| 6925 | auto CloneFuncAliases = [&](Function *NewF, unsigned I) { |
| 6926 | if (!FuncToAliasMap.count(x: &F)) |
| 6927 | return; |
| 6928 | for (auto *A : FuncToAliasMap[&F]) { |
| 6929 | std::string AliasName = getMemProfFuncName(Base: A->getName(), CloneNo: I); |
| 6930 | auto *PrevA = M.getNamedAlias(Name: AliasName); |
| 6931 | auto *NewA = GlobalAlias::create(Ty: A->getValueType(), |
| 6932 | AddressSpace: A->getType()->getPointerAddressSpace(), |
| 6933 | Linkage: A->getLinkage(), Name: AliasName, Aliasee: NewF); |
| 6934 | NewA->copyAttributesFrom(Src: A); |
| 6935 | if (PrevA) |
| 6936 | TakeDeclNameAndReplace(PrevA, NewA); |
| 6937 | } |
| 6938 | }; |
| 6939 | |
| 6940 | // The first "clone" is the original copy, we should only call this if we |
| 6941 | // needed to create new clones. |
| 6942 | assert(NumClones > 1); |
| 6943 | SmallVector<std::unique_ptr<ValueToValueMapTy>, 4> VMaps; |
| 6944 | VMaps.reserve(N: NumClones - 1); |
| 6945 | FunctionsClonedThinBackend++; |
| 6946 | |
| 6947 | // Map of hash of callsite/alloc versions to the instantiated function clone |
| 6948 | // (possibly the original) implementing those calls and its clone number. Used |
| 6949 | // to avoid instantiating duplicate function clones. |
| 6950 | // FIXME: Ideally the thin link would not generate such duplicate clones to |
| 6951 | // start with, but right now it happens due to phase ordering in the function |
| 6952 | // assignment and possible new clones that produces. We simply make each |
| 6953 | // duplicate an alias to the matching instantiated clone recorded in the map |
| 6954 | // (except for available_externally which are made declarations as they would |
| 6955 | // be aliases in the prevailing module, and available_externally aliases are |
| 6956 | // not well supported right now). |
| 6957 | DenseMap<uint64_t, std::pair<Function *, unsigned>> HashToFunc; |
| 6958 | |
| 6959 | // Save the hash of the original function version. |
| 6960 | HashToFunc[ComputeHash(FS, I: 0)] = {&F, 0}; |
| 6961 | |
| 6962 | for (unsigned I = 1; I < NumClones; I++) { |
| 6963 | VMaps.emplace_back(Args: std::make_unique<ValueToValueMapTy>()); |
| 6964 | std::string Name = getMemProfFuncName(Base: F.getName(), CloneNo: I); |
| 6965 | auto Hash = ComputeHash(FS, I); |
| 6966 | // If this clone would duplicate a previously seen clone, don't generate the |
| 6967 | // duplicate clone body, just make an alias to satisfy any (potentially |
| 6968 | // cross-module) references. |
| 6969 | if (auto It = HashToFunc.find(Val: Hash); It != HashToFunc.end()) { |
| 6970 | FunctionCloneDuplicatesThinBackend++; |
| 6971 | auto [Func, OrigI] = It->second; |
| 6972 | assert(clonesHaveSameVersions(FS, I, OrigI) && |
| 6973 | "Duplicate clone hash collision" ); |
| 6974 | if (Func->hasAvailableExternallyLinkage()) { |
| 6975 | // Skip these as EliminateAvailableExternallyPass does not handle |
| 6976 | // available_externally aliases correctly and we end up with an |
| 6977 | // available_externally alias to a declaration. Just create a |
| 6978 | // declaration for now as we know we will have a definition in another |
| 6979 | // module. |
| 6980 | auto Decl = M.getOrInsertFunction(Name, T: Func->getFunctionType()); |
| 6981 | ORE.emit(OptDiag: OptimizationRemark(DEBUG_TYPE, "MemprofClone" , &F) |
| 6982 | << "created clone decl " << ore::NV("Decl" , Decl.getCallee())); |
| 6983 | continue; |
| 6984 | } |
| 6985 | auto *PrevF = M.getFunction(Name); |
| 6986 | auto *Alias = GlobalAlias::create(Name, Aliasee: Func); |
| 6987 | if (PrevF) |
| 6988 | TakeDeclNameAndReplace(PrevF, Alias); |
| 6989 | ORE.emit(OptDiag: OptimizationRemark(DEBUG_TYPE, "MemprofClone" , &F) |
| 6990 | << "created clone alias " << ore::NV("Alias" , Alias)); |
| 6991 | |
| 6992 | // Now handle aliases to this function, and clone those as well. |
| 6993 | CloneFuncAliases(Func, I); |
| 6994 | continue; |
| 6995 | } |
| 6996 | auto *NewF = CloneFunction(F: &F, VMap&: *VMaps.back()); |
| 6997 | HashToFunc[Hash] = {NewF, I}; |
| 6998 | FunctionClonesThinBackend++; |
| 6999 | // Strip !kcfi_type (see comment in cloneFunctionForCallsite). |
| 7000 | NewF->setMetadata(KindID: LLVMContext::MD_kcfi_type, Node: nullptr); |
| 7001 | // Strip memprof and callsite metadata from clone as they are no longer |
| 7002 | // needed. |
| 7003 | for (auto &BB : *NewF) { |
| 7004 | for (auto &Inst : BB) { |
| 7005 | Inst.setMetadata(KindID: LLVMContext::MD_memprof, Node: nullptr); |
| 7006 | Inst.setMetadata(KindID: LLVMContext::MD_callsite, Node: nullptr); |
| 7007 | } |
| 7008 | } |
| 7009 | auto *PrevF = M.getFunction(Name); |
| 7010 | if (PrevF) |
| 7011 | TakeDeclNameAndReplace(PrevF, NewF); |
| 7012 | else |
| 7013 | NewF->setName(Name); |
| 7014 | updateSubprogramLinkageName(NewFunc: NewF, Name); |
| 7015 | ORE.emit(OptDiag: OptimizationRemark(DEBUG_TYPE, "MemprofClone" , &F) |
| 7016 | << "created clone " << ore::NV("NewFunction" , NewF)); |
| 7017 | |
| 7018 | // Now handle aliases to this function, and clone those as well. |
| 7019 | CloneFuncAliases(NewF, I); |
| 7020 | } |
| 7021 | return VMaps; |
| 7022 | } |
| 7023 | |
| 7024 | // If F is an imported copy of an alias (materialized as a Function by |
| 7025 | // replaceAliasWithAliasee during FunctionImport while preserving the aliasee's |
| 7026 | // !guid metadata for body-level passes), return the original alias GUID |
| 7027 | // recorded in its thinlto_alias_guid metadata. |
| 7028 | static std::optional<GlobalValue::GUID> |
| 7029 | getImportedAliasGUID(const Function &F) { |
| 7030 | if (auto *MD = F.getMetadata(Kind: "thinlto_alias_guid" )) |
| 7031 | return mdconst::extract<ConstantInt>(MD: MD->getOperand(I: 0))->getZExtValue(); |
| 7032 | return std::nullopt; |
| 7033 | } |
| 7034 | |
| 7035 | // Locate the summary for GV. This is complicated by the fact that it might |
| 7036 | // have been internalized, promoted, or imported as a copy of an alias. |
| 7037 | static ValueInfo findValueInfoForGV(const GlobalValue &GV, const Module &M, |
| 7038 | const ModuleSummaryIndex *ImportSummary, |
| 7039 | const Function *CallingFunc = nullptr) { |
| 7040 | if (const auto *F = dyn_cast<Function>(Val: &GV)) |
| 7041 | if (auto AliasGUID = getImportedAliasGUID(F: *F)) { |
| 7042 | ValueInfo AliasVI = ImportSummary->getValueInfo(GUID: *AliasGUID); |
| 7043 | assert(AliasVI && "Expected ValueInfo for imported alias GUID" ); |
| 7044 | return AliasVI; |
| 7045 | } |
| 7046 | |
| 7047 | // FIXME: Ideally we would retain the original GUID in some fashion on the |
| 7048 | // function (e.g. as metadata), but for now do our best to locate the |
| 7049 | // summary without that information. |
| 7050 | ValueInfo TheFnVI = ImportSummary->getValueInfo(GUID: GV.getGUID()); |
| 7051 | if (!TheFnVI) |
| 7052 | // See if GV was internalized, by checking index directly with |
| 7053 | // original name (this avoids the name adjustment done by getGUID() for |
| 7054 | // internal symbols). |
| 7055 | TheFnVI = ImportSummary->getValueInfo( |
| 7056 | GUID: GlobalValue::getGUIDAssumingExternalLinkage(GlobalName: GV.getName())); |
| 7057 | if (TheFnVI) |
| 7058 | return TheFnVI; |
| 7059 | // Now query with the original name before any promotion was performed. |
| 7060 | StringRef OrigName = |
| 7061 | ModuleSummaryIndex::getOriginalNameBeforePromote(Name: GV.getName()); |
| 7062 | // When this pass is enabled, we always add thinlto_src_file provenance |
| 7063 | // metadata to imported function definitions, which allows us to recreate the |
| 7064 | // original internal symbol's GUID. |
| 7065 | const auto *GO = dyn_cast<GlobalObject>(Val: &GV); |
| 7066 | MDNode *SrcFileMD = GO ? GO->getMetadata(Kind: "thinlto_src_file" ) : nullptr; |
| 7067 | // If this is a call to an imported/promoted local for which we didn't import |
| 7068 | // the definition, the metadata will not exist on the declaration. However, |
| 7069 | // since we are doing this early, before any inlining in the LTO backend, we |
| 7070 | // can simply look at the metadata on the calling function which must have |
| 7071 | // been from the same module if GV was an internal symbol originally. |
| 7072 | if (!SrcFileMD && GV.isDeclaration()) { |
| 7073 | // We would only call this for a declaration for a direct callsite, in which |
| 7074 | // case the caller would have provided the calling function pointer. |
| 7075 | assert(CallingFunc); |
| 7076 | SrcFileMD = CallingFunc->getMetadata(Kind: "thinlto_src_file" ); |
| 7077 | // If this is a promoted local (OrigName != GV.getName()), since this is a |
| 7078 | // declaration, it must be imported from a different module and therefore we |
| 7079 | // should always find the metadata on its calling function. Any call to a |
| 7080 | // promoted local that came from this module should still be a definition. |
| 7081 | assert(SrcFileMD || OrigName == GV.getName()); |
| 7082 | } |
| 7083 | StringRef SrcFile = M.getSourceFileName(); |
| 7084 | if (SrcFileMD) |
| 7085 | SrcFile = dyn_cast<MDString>(Val: SrcFileMD->getOperand(I: 0))->getString(); |
| 7086 | std::string OrigId = GlobalValue::getGlobalIdentifier( |
| 7087 | Name: OrigName, Linkage: GlobalValue::InternalLinkage, FileName: SrcFile); |
| 7088 | TheFnVI = ImportSummary->getValueInfo( |
| 7089 | GUID: GlobalValue::getGUIDAssumingExternalLinkage(GlobalName: OrigId)); |
| 7090 | // Internal func in original module may have gotten a numbered suffix if we |
| 7091 | // imported an external function with the same name. This happens |
| 7092 | // automatically during IR linking for naming conflicts. It would have to |
| 7093 | // still be internal in that case (otherwise it would have been renamed on |
| 7094 | // promotion in which case we wouldn't have a naming conflict). |
| 7095 | if (!TheFnVI && OrigName == GV.getName() && GV.hasLocalLinkage() && |
| 7096 | GV.getName().contains(C: '.')) { |
| 7097 | OrigName = GV.getName().rsplit(Separator: '.').first; |
| 7098 | OrigId = GlobalValue::getGlobalIdentifier( |
| 7099 | Name: OrigName, Linkage: GlobalValue::InternalLinkage, FileName: SrcFile); |
| 7100 | TheFnVI = ImportSummary->getValueInfo( |
| 7101 | GUID: GlobalValue::getGUIDAssumingExternalLinkage(GlobalName: OrigId)); |
| 7102 | } |
| 7103 | // The only way we may not have a VI is if this is a declaration created for |
| 7104 | // an imported reference. For distributed ThinLTO we may not have a VI for |
| 7105 | // such declarations in the distributed summary. |
| 7106 | assert(TheFnVI || GV.isDeclaration()); |
| 7107 | return TheFnVI; |
| 7108 | } |
| 7109 | |
| 7110 | // Find the CallsiteInfo record synthesized by the thin link for profiled target |
| 7111 | // CandidateGUID of an indirect tail call, in MapTailCallCalleeVIToCallsite |
| 7112 | // (which maps each synthesized record's callee to the record). Returns null if |
| 7113 | // there is none. OrigNameGUIDToTailCallCallsite caches a map from the original |
| 7114 | // name GUID of each record's callee to the record, and is built on first use |
| 7115 | // (along with AliasVIToGV, which caches alias ValueInfos in M). |
| 7116 | static const CallsiteInfo *findIndirectTailCallCallsite( |
| 7117 | uint64_t CandidateGUID, |
| 7118 | const DenseMap<ValueInfo, CallsiteInfo> &MapTailCallCalleeVIToCallsite, |
| 7119 | std::optional<DenseMap<uint64_t, const CallsiteInfo *>> |
| 7120 | &OrigNameGUIDToTailCallCallsite, |
| 7121 | std::optional<DenseMap<ValueInfo, const GlobalAlias *>> &AliasVIToGV, |
| 7122 | const Module &M, const ModuleSummaryIndex *ImportSummary, |
| 7123 | InstrProfSymtab &Symtab) { |
| 7124 | auto End = MapTailCallCalleeVIToCallsite.end(); |
| 7125 | // The callee of the synthesized record is normally the VP metadata GUID, so |
| 7126 | // check that first. |
| 7127 | if (ValueInfo VI = ImportSummary->getValueInfo(GUID: CandidateGUID)) { |
| 7128 | auto It = MapTailCallCalleeVIToCallsite.find(Val: VI); |
| 7129 | if (It != End) |
| 7130 | return &It->second; |
| 7131 | } |
| 7132 | // For local targets the VP metadata may contain the GUID of the original |
| 7133 | // name without the source file prefix, which the thin link translates to the |
| 7134 | // actual GUID of the local function (see updateValueInfoForIndirectCalls). |
| 7135 | if (auto LocalGUID = ImportSummary->getGUIDFromOriginalID(OriginalID: CandidateGUID)) { |
| 7136 | if (ValueInfo LocalVI = ImportSummary->getValueInfo(GUID: LocalGUID)) { |
| 7137 | auto It = MapTailCallCalleeVIToCallsite.find(Val: LocalVI); |
| 7138 | if (It != End) |
| 7139 | return &It->second; |
| 7140 | } |
| 7141 | } |
| 7142 | // The index for a distributed ThinLTO backend does not record the original |
| 7143 | // names of local functions, in which case the above lookup via |
| 7144 | // getGUIDFromOriginalID fails. Instead, compute the original name GUID of |
| 7145 | // each record's callee from its function or alias in this module, using its |
| 7146 | // name before any promotion. |
| 7147 | if (!OrigNameGUIDToTailCallCallsite) { |
| 7148 | // Construct the (initially empty) map so that it is marked as built, even |
| 7149 | // if no entries are added below, to avoid rebuilding it on later calls. |
| 7150 | OrigNameGUIDToTailCallCallsite.emplace(); |
| 7151 | for (const auto &[CalleeVI, Callsite] : MapTailCallCalleeVIToCallsite) { |
| 7152 | const GlobalValue *Callee = Symtab.getFunction(FuncMD5Hash: CalleeVI.getGUID()); |
| 7153 | if (!Callee) { |
| 7154 | // InstrProfSymtab only indexes Functions, not GlobalAliases (imported |
| 7155 | // alias copies are materialized as Functions with thinlto_alias_guid |
| 7156 | // metadata and are found via Symtab above). Lazily build a module-wide |
| 7157 | // map for same-module function GlobalAliases once per module so we can |
| 7158 | // recover the alias's pre-promotion name. |
| 7159 | if (!AliasVIToGV) { |
| 7160 | AliasVIToGV.emplace(); |
| 7161 | for (const GlobalAlias &A : M.aliases()) { |
| 7162 | if (isMemProfClone(GV: A) || |
| 7163 | !isa_and_nonnull<Function>(Val: A.getAliaseeObject())) |
| 7164 | continue; |
| 7165 | if (ValueInfo VI = findValueInfoForGV(GV: A, M, ImportSummary)) |
| 7166 | AliasVIToGV->try_emplace(Key: VI, Args: &A); |
| 7167 | } |
| 7168 | } |
| 7169 | Callee = AliasVIToGV->lookup(Val: CalleeVI); |
| 7170 | if (!Callee) |
| 7171 | continue; |
| 7172 | } |
| 7173 | StringRef OrigName = |
| 7174 | ModuleSummaryIndex::getOriginalNameBeforePromote(Name: Callee->getName()); |
| 7175 | auto [It, Inserted] = OrigNameGUIDToTailCallCallsite->try_emplace( |
| 7176 | Key: GlobalValue::getGUIDAssumingExternalLinkage(GlobalName: OrigName), Args: &Callsite); |
| 7177 | // If multiple callees have the same original name, we can't tell which |
| 7178 | // corresponds to the profiled target. |
| 7179 | if (!Inserted && It->second != &Callsite) |
| 7180 | It->second = nullptr; |
| 7181 | } |
| 7182 | } |
| 7183 | return OrigNameGUIDToTailCallCallsite->lookup(Val: CandidateGUID); |
| 7184 | } |
| 7185 | |
| 7186 | bool MemProfContextDisambiguation::initializeIndirectCallPromotionInfo( |
| 7187 | Module &M) { |
| 7188 | ICallAnalysis = std::make_unique<ICallPromotionAnalysis>(); |
| 7189 | Symtab = std::make_unique<InstrProfSymtab>(); |
| 7190 | // Don't add canonical names, to avoid multiple functions to the symtab |
| 7191 | // when they both have the same root name with "." suffixes stripped. |
| 7192 | // If we pick the wrong one then this could lead to incorrect ICP and calling |
| 7193 | // a memprof clone that we don't actually create (resulting in linker unsats). |
| 7194 | // What this means is that the GUID of the function *must* match that in the |
| 7195 | // VP metadata to allow promotion. In practice this should not be a |
| 7196 | // limitation, since local functions have a GUID from before they were |
| 7197 | // promoted and global function names shouldn't need any special handling |
| 7198 | // (they should not get the ".llvm.*" suffix that the canonicalization |
| 7199 | // handling is attempting to strip). |
| 7200 | // Key imported copies of aliases by the alias's GUID rather than the |
| 7201 | // aliasee's GUID retained in their !guid metadata, since we rely on the |
| 7202 | // looked up function's name (to compute its original name GUID and the names |
| 7203 | // of its clones), and so that they don't shadow the aliasee's own function |
| 7204 | // when both are in the module. See AliaseeGUIDToImportedAlias for how |
| 7205 | // performICP handles profiled targets that are the aliasee of an imported |
| 7206 | // alias copy when the aliasee itself is not in the module. |
| 7207 | AliaseeGUIDToImportedAlias.reset(); |
| 7208 | if (Error E = Symtab->create(M, /*InLTO=*/true, /*AddCanonical=*/false, |
| 7209 | /*KeyImportedAliasesByAliasGUID=*/true)) { |
| 7210 | std::string SymtabFailure = toString(E: std::move(E)); |
| 7211 | M.getContext().emitError(ErrorStr: "Failed to create symtab: " + SymtabFailure); |
| 7212 | return false; |
| 7213 | } |
| 7214 | return true; |
| 7215 | } |
| 7216 | |
| 7217 | #ifndef NDEBUG |
| 7218 | // Sanity check that the MIB stack ids match between the summary and |
| 7219 | // instruction metadata. |
| 7220 | static void checkAllocContextIds( |
| 7221 | const AllocInfo &AllocNode, const MDNode *MemProfMD, |
| 7222 | const CallStack<MDNode, MDNode::op_iterator> &CallsiteContext, |
| 7223 | const ModuleSummaryIndex *ImportSummary) { |
| 7224 | auto MIBIter = AllocNode.MIBs.begin(); |
| 7225 | for (auto &MDOp : MemProfMD->operands()) { |
| 7226 | assert(MIBIter != AllocNode.MIBs.end()); |
| 7227 | // The combined index may have been written without the MIB contexts (see |
| 7228 | // -combined-index-memprof-context, which defaults to false in NDEBUG |
| 7229 | // builds), e.g. if the thin link was performed by a release build while |
| 7230 | // this backend has assertions enabled. Since the MIB context in the summary |
| 7231 | // can never be empty after trimming the allocation's callsite context, an |
| 7232 | // empty list means the context was not serialized, and there is nothing to |
| 7233 | // check. |
| 7234 | if (MIBIter->StackIdIndices.empty()) { |
| 7235 | MIBIter++; |
| 7236 | continue; |
| 7237 | } |
| 7238 | auto StackIdIndexIter = MIBIter->StackIdIndices.begin(); |
| 7239 | auto *MIBMD = cast<const MDNode>(MDOp); |
| 7240 | MDNode *StackMDNode = getMIBStackNode(MIBMD); |
| 7241 | assert(StackMDNode); |
| 7242 | CallStack<MDNode, MDNode::op_iterator> StackContext(StackMDNode); |
| 7243 | auto ContextIterBegin = |
| 7244 | StackContext.beginAfterSharedPrefix(CallsiteContext); |
| 7245 | // Skip the checking on the first iteration. |
| 7246 | uint64_t LastStackContextId = |
| 7247 | (ContextIterBegin != StackContext.end() && *ContextIterBegin == 0) ? 1 |
| 7248 | : 0; |
| 7249 | for (auto ContextIter = ContextIterBegin; ContextIter != StackContext.end(); |
| 7250 | ++ContextIter) { |
| 7251 | // If this is a direct recursion, simply skip the duplicate |
| 7252 | // entries, to be consistent with how the summary ids were |
| 7253 | // generated during ModuleSummaryAnalysis. |
| 7254 | if (LastStackContextId == *ContextIter) |
| 7255 | continue; |
| 7256 | LastStackContextId = *ContextIter; |
| 7257 | assert(StackIdIndexIter != MIBIter->StackIdIndices.end()); |
| 7258 | assert(ImportSummary->getStackIdAtIndex(*StackIdIndexIter) == |
| 7259 | *ContextIter); |
| 7260 | StackIdIndexIter++; |
| 7261 | } |
| 7262 | MIBIter++; |
| 7263 | } |
| 7264 | } |
| 7265 | #endif |
| 7266 | |
| 7267 | bool MemProfContextDisambiguation::applyImport(Module &M) { |
| 7268 | assert(ImportSummary); |
| 7269 | bool Changed = false; |
| 7270 | |
| 7271 | // We also need to clone any aliases that reference cloned functions, because |
| 7272 | // the modified callsites may invoke via the alias. Keep track of the aliases |
| 7273 | // for each function. |
| 7274 | std::map<const Function *, SmallPtrSet<const GlobalAlias *, 1>> |
| 7275 | FuncToAliasMap; |
| 7276 | for (auto &A : M.aliases()) { |
| 7277 | auto *Aliasee = A.getAliaseeObject(); |
| 7278 | if (auto *F = dyn_cast<Function>(Val: Aliasee)) |
| 7279 | FuncToAliasMap[F].insert(Ptr: &A); |
| 7280 | } |
| 7281 | |
| 7282 | if (!initializeIndirectCallPromotionInfo(M)) |
| 7283 | return false; |
| 7284 | |
| 7285 | // Lazily populated map from alias ValueInfo to the corresponding same-module |
| 7286 | // GlobalAlias, used by findIndirectTailCallCallsite when a local alias target |
| 7287 | // is not in Symtab. |
| 7288 | std::optional<DenseMap<ValueInfo, const GlobalAlias *>> AliasVIToGV; |
| 7289 | |
| 7290 | for (auto &F : M) { |
| 7291 | if (F.isDeclaration() || isMemProfClone(GV: F)) |
| 7292 | continue; |
| 7293 | |
| 7294 | OptimizationRemarkEmitter ORE(&F); |
| 7295 | |
| 7296 | SmallVector<std::unique_ptr<ValueToValueMapTy>, 4> VMaps; |
| 7297 | bool ClonesCreated = false; |
| 7298 | unsigned NumClonesCreated = 0; |
| 7299 | auto CloneFuncIfNeeded = [&](unsigned NumClones, FunctionSummary *FS) { |
| 7300 | // We should at least have version 0 which is the original copy. |
| 7301 | assert(NumClones > 0); |
| 7302 | // If only one copy needed use original. |
| 7303 | if (NumClones == 1) |
| 7304 | return; |
| 7305 | // If we already performed cloning of this function, confirm that the |
| 7306 | // requested number of clones matches (the thin link should ensure the |
| 7307 | // number of clones for each constituent callsite is consistent within |
| 7308 | // each function), before returning. |
| 7309 | if (ClonesCreated) { |
| 7310 | assert(NumClonesCreated == NumClones); |
| 7311 | return; |
| 7312 | } |
| 7313 | VMaps = createFunctionClones(F, NumClones, M, ORE, FuncToAliasMap, FS); |
| 7314 | // The first "clone" is the original copy, which doesn't have a VMap. |
| 7315 | assert(VMaps.size() == NumClones - 1); |
| 7316 | Changed = true; |
| 7317 | ClonesCreated = true; |
| 7318 | NumClonesCreated = NumClones; |
| 7319 | }; |
| 7320 | |
| 7321 | auto CloneCallsite = [&](const CallsiteInfo &StackNode, CallBase *CB, |
| 7322 | Function *CalledFunction, FunctionSummary *FS) { |
| 7323 | // Perform cloning if not yet done. |
| 7324 | CloneFuncIfNeeded(/*NumClones=*/StackNode.Clones.size(), FS); |
| 7325 | |
| 7326 | assert(!isMemProfClone(*CalledFunction)); |
| 7327 | |
| 7328 | // Because we update the cloned calls by calling setCalledOperand (see |
| 7329 | // comment below), out of an abundance of caution make sure the called |
| 7330 | // function was actually the called operand (or its aliasee). We also |
| 7331 | // strip pointer casts when looking for calls (to match behavior during |
| 7332 | // summary generation), however, with opaque pointers in theory this |
| 7333 | // should not be an issue. Note we still clone the current function |
| 7334 | // (containing this call) above, as that could be needed for its callers. |
| 7335 | auto *GA = dyn_cast_or_null<GlobalAlias>(Val: CB->getCalledOperand()); |
| 7336 | if (CalledFunction != CB->getCalledOperand() && |
| 7337 | (!GA || CalledFunction != GA->getAliaseeObject())) { |
| 7338 | SkippedCallsCloning++; |
| 7339 | return; |
| 7340 | } |
| 7341 | // Update the calls per the summary info. |
| 7342 | // Save orig name since it gets updated in the first iteration |
| 7343 | // below. |
| 7344 | auto CalleeOrigName = CalledFunction->getName(); |
| 7345 | for (unsigned J = 0; J < StackNode.Clones.size(); J++) { |
| 7346 | // If the VMap is empty, this clone was a duplicate of another and was |
| 7347 | // created as an alias or a declaration. |
| 7348 | if (J > 0 && VMaps[J - 1]->empty()) |
| 7349 | continue; |
| 7350 | // Do nothing if this version calls the original version of its |
| 7351 | // callee. |
| 7352 | if (!StackNode.Clones[J]) |
| 7353 | continue; |
| 7354 | auto NewF = M.getOrInsertFunction( |
| 7355 | Name: getMemProfFuncName(Base: CalleeOrigName, CloneNo: StackNode.Clones[J]), |
| 7356 | T: CalledFunction->getFunctionType()); |
| 7357 | CallBase *CBClone; |
| 7358 | // Copy 0 is the original function. |
| 7359 | if (!J) |
| 7360 | CBClone = CB; |
| 7361 | else |
| 7362 | CBClone = cast<CallBase>(Val&: (*VMaps[J - 1])[CB]); |
| 7363 | // Set the called operand directly instead of calling setCalledFunction, |
| 7364 | // as the latter mutates the function type on the call. In rare cases |
| 7365 | // we may have a slightly different type on a callee function |
| 7366 | // declaration due to it being imported from a different module with |
| 7367 | // incomplete types. We really just want to change the name of the |
| 7368 | // function to the clone, and not make any type changes. |
| 7369 | CBClone->setCalledOperand(NewF.getCallee()); |
| 7370 | ORE.emit(OptDiag: OptimizationRemark(DEBUG_TYPE, "MemprofCall" , CBClone) |
| 7371 | << ore::NV("Call" , CBClone) << " in clone " |
| 7372 | << ore::NV("Caller" , CBClone->getFunction()) |
| 7373 | << " assigned to call function clone " |
| 7374 | << ore::NV("Callee" , NewF.getCallee())); |
| 7375 | } |
| 7376 | }; |
| 7377 | |
| 7378 | // Locate the summary for F. |
| 7379 | ValueInfo TheFnVI = findValueInfoForGV(GV: F, M, ImportSummary); |
| 7380 | // If not found, this could be an imported local (see comment in |
| 7381 | // findValueInfoForGV). Skip for now as it will be cloned in its original |
| 7382 | // module (where it would have been promoted to global scope so should |
| 7383 | // satisfy any reference in this module). |
| 7384 | if (!TheFnVI) |
| 7385 | continue; |
| 7386 | |
| 7387 | auto *GVSummary = |
| 7388 | ImportSummary->findSummaryInModule(VI: TheFnVI, ModuleId: M.getModuleIdentifier()); |
| 7389 | if (!GVSummary) { |
| 7390 | // Must have been imported, use the summary which matches the definition。 |
| 7391 | // (might be multiple if this was a linkonce_odr). |
| 7392 | auto SrcModuleMD = F.getMetadata(Kind: "thinlto_src_module" ); |
| 7393 | assert(SrcModuleMD && |
| 7394 | "enable-import-metadata is needed to emit thinlto_src_module" ); |
| 7395 | StringRef SrcModule = |
| 7396 | dyn_cast<MDString>(Val: SrcModuleMD->getOperand(I: 0))->getString(); |
| 7397 | for (auto &GVS : TheFnVI.getSummaryList()) { |
| 7398 | if (GVS->modulePath() == SrcModule) { |
| 7399 | GVSummary = GVS.get(); |
| 7400 | break; |
| 7401 | } |
| 7402 | } |
| 7403 | assert(GVSummary && GVSummary->modulePath() == SrcModule); |
| 7404 | } |
| 7405 | |
| 7406 | // If this was an imported alias skip it as we won't have the function |
| 7407 | // summary, and it should be cloned in the original module. |
| 7408 | if (isa<AliasSummary>(Val: GVSummary)) |
| 7409 | continue; |
| 7410 | |
| 7411 | auto *FS = cast<FunctionSummary>(Val: GVSummary->getBaseObject()); |
| 7412 | |
| 7413 | if (FS->allocs().empty() && FS->callsites().empty()) |
| 7414 | continue; |
| 7415 | |
| 7416 | auto SI = FS->callsites().begin(); |
| 7417 | auto AI = FS->allocs().begin(); |
| 7418 | |
| 7419 | // To handle callsite infos synthesized for tail calls which have missing |
| 7420 | // frames in the profiled context, map callee VI to the synthesized callsite |
| 7421 | // info. |
| 7422 | DenseMap<ValueInfo, CallsiteInfo> MapTailCallCalleeVIToCallsite; |
| 7423 | // Iterate the callsites for this function in reverse, since we place all |
| 7424 | // those synthesized for tail calls at the end. |
| 7425 | for (auto CallsiteIt = FS->callsites().rbegin(); |
| 7426 | CallsiteIt != FS->callsites().rend(); CallsiteIt++) { |
| 7427 | auto &Callsite = *CallsiteIt; |
| 7428 | // Stop as soon as we see a non-synthesized callsite info (see comment |
| 7429 | // above loop). All the entries added for discovered tail calls have empty |
| 7430 | // stack ids. |
| 7431 | if (!Callsite.StackIdIndices.empty()) |
| 7432 | break; |
| 7433 | MapTailCallCalleeVIToCallsite.insert(KV: {Callsite.Callee, Callsite}); |
| 7434 | } |
| 7435 | // Maps the original name GUID of each synthesized callsite info's callee to |
| 7436 | // the callsite info, built on first use by findIndirectTailCallCallsite. |
| 7437 | std::optional<DenseMap<uint64_t, const CallsiteInfo *>> |
| 7438 | OrigNameGUIDToTailCallCallsite; |
| 7439 | |
| 7440 | // Keeps track of needed ICP for the function. |
| 7441 | SmallVector<ICallAnalysisData> ICallAnalysisInfo; |
| 7442 | |
| 7443 | // Assume for now that the instructions are in the exact same order |
| 7444 | // as when the summary was created, but confirm this is correct by |
| 7445 | // matching the stack ids. |
| 7446 | for (auto &BB : F) { |
| 7447 | for (auto &I : BB) { |
| 7448 | auto *CB = dyn_cast<CallBase>(Val: &I); |
| 7449 | // Same handling as when creating module summary. |
| 7450 | if (!mayHaveMemprofSummary(CB)) |
| 7451 | continue; |
| 7452 | |
| 7453 | auto *CalledValue = CB->getCalledOperand(); |
| 7454 | auto *CalledFunction = CB->getCalledFunction(); |
| 7455 | if (CalledValue && !CalledFunction) { |
| 7456 | CalledValue = CalledValue->stripPointerCasts(); |
| 7457 | // Stripping pointer casts can reveal a called function. |
| 7458 | CalledFunction = dyn_cast<Function>(Val: CalledValue); |
| 7459 | } |
| 7460 | // Check if this is an alias to a function. If so, get the |
| 7461 | // called aliasee for the checks below. |
| 7462 | if (auto *GA = dyn_cast<GlobalAlias>(Val: CalledValue)) { |
| 7463 | assert(!CalledFunction && |
| 7464 | "Expected null called function in callsite for alias" ); |
| 7465 | CalledFunction = dyn_cast<Function>(Val: GA->getAliaseeObject()); |
| 7466 | } |
| 7467 | |
| 7468 | CallStack<MDNode, MDNode::op_iterator> CallsiteContext( |
| 7469 | I.getMetadata(KindID: LLVMContext::MD_callsite)); |
| 7470 | auto *MemProfMD = I.getMetadata(KindID: LLVMContext::MD_memprof); |
| 7471 | |
| 7472 | // Include allocs that were already assigned a memprof function |
| 7473 | // attribute in the statistics. Only do this for those that do not have |
| 7474 | // memprof metadata, since we add an "ambiguous" memprof attribute by |
| 7475 | // default. |
| 7476 | if (CB->getAttributes().hasFnAttr(Kind: "memprof" ) && !MemProfMD) { |
| 7477 | CB->getAttributes().getFnAttr(Kind: "memprof" ).getValueAsString() == "cold" |
| 7478 | ? AllocTypeColdThinBackend++ |
| 7479 | : AllocTypeNotColdThinBackend++; |
| 7480 | OrigAllocsThinBackend++; |
| 7481 | AllocVersionsThinBackend++; |
| 7482 | if (!MaxAllocVersionsThinBackend) |
| 7483 | MaxAllocVersionsThinBackend = 1; |
| 7484 | continue; |
| 7485 | } |
| 7486 | |
| 7487 | if (MemProfMD) { |
| 7488 | // Consult the next alloc node. |
| 7489 | assert(AI != FS->allocs().end()); |
| 7490 | auto &AllocNode = *(AI++); |
| 7491 | |
| 7492 | #ifndef NDEBUG |
| 7493 | checkAllocContextIds(AllocNode, MemProfMD, CallsiteContext, |
| 7494 | ImportSummary); |
| 7495 | #endif |
| 7496 | |
| 7497 | // Perform cloning if not yet done. |
| 7498 | CloneFuncIfNeeded(/*NumClones=*/AllocNode.Versions.size(), FS); |
| 7499 | |
| 7500 | OrigAllocsThinBackend++; |
| 7501 | AllocVersionsThinBackend += AllocNode.Versions.size(); |
| 7502 | if (MaxAllocVersionsThinBackend < AllocNode.Versions.size()) |
| 7503 | MaxAllocVersionsThinBackend = AllocNode.Versions.size(); |
| 7504 | |
| 7505 | // If there is only one version that means we didn't end up |
| 7506 | // considering this function for cloning. Count that as unclonable |
| 7507 | // unless the single version is cold due to the |
| 7508 | // MinClonedColdBytePercent heuristic. The alloc may still be none |
| 7509 | // type (not reached during whole program analysis), in which case |
| 7510 | // there is nothing to apply; otherwise it got the type the whole |
| 7511 | // program analysis resolved it to (not cold, or both types when the |
| 7512 | // contexts could not be disambiguated), which is applied below like |
| 7513 | // for cloned functions so that the attribute matches the regular |
| 7514 | // LTO result. |
| 7515 | if (AllocNode.Versions.size() == 1 && |
| 7516 | (AllocationType)AllocNode.Versions[0] != AllocationType::Cold) |
| 7517 | UnclonableAllocsThinBackend++; |
| 7518 | |
| 7519 | // Update the allocation types per the summary info. |
| 7520 | for (unsigned J = 0; J < AllocNode.Versions.size(); J++) { |
| 7521 | // If the VMap is empty, this clone was a duplicate of another and |
| 7522 | // was created as an alias or a declaration. |
| 7523 | if (J > 0 && VMaps[J - 1]->empty()) |
| 7524 | continue; |
| 7525 | // Ignore any that didn't get an assigned allocation type. |
| 7526 | if (AllocNode.Versions[J] == (uint8_t)AllocationType::None) |
| 7527 | continue; |
| 7528 | AllocationType AllocTy = (AllocationType)AllocNode.Versions[J]; |
| 7529 | // The whole program analysis only records both types when |
| 7530 | // ambiguous attributes are enabled; if the backend runs without |
| 7531 | // them, fall back to the not cold default. |
| 7532 | if (isAmbiguousAllocType(AllocTypes: (uint8_t)AllocTy) && |
| 7533 | !MemProfUseAmbiguousAttributes) |
| 7534 | AllocTy = AllocationType::NotCold; |
| 7535 | const bool IsAmbiguous = isAmbiguousAllocType(AllocTypes: (uint8_t)AllocTy); |
| 7536 | if (AllocTy == AllocationType::Cold) |
| 7537 | AllocTypeColdThinBackend++; |
| 7538 | else if (IsAmbiguous) |
| 7539 | AllocTypeAmbiguousThinBackend++; |
| 7540 | else |
| 7541 | AllocTypeNotColdThinBackend++; |
| 7542 | std::string AllocTypeString = getAllocTypeAttributeString(Type: AllocTy); |
| 7543 | CallBase *CBClone; |
| 7544 | // Copy 0 is the original function. |
| 7545 | if (!J) |
| 7546 | CBClone = CB; |
| 7547 | else |
| 7548 | // Since VMaps are only created for new clones, we index with |
| 7549 | // clone J-1 (J==0 is the original clone and does not have a VMaps |
| 7550 | // entry). |
| 7551 | CBClone = cast<CallBase>(Val&: (*VMaps[J - 1])[CB]); |
| 7552 | if (IsAmbiguous) { |
| 7553 | // The contexts could not be disambiguated: keep (or add) the |
| 7554 | // ambiguous attribute from matching instead of a specific type. |
| 7555 | addAmbiguousAttribute(CB: CBClone); |
| 7556 | } else { |
| 7557 | removeAnyExistingAmbiguousAttribute(CB: CBClone); |
| 7558 | CBClone->addFnAttr(Attr: llvm::Attribute::get(Context&: F.getContext(), Kind: "memprof" , |
| 7559 | Val: AllocTypeString)); |
| 7560 | } |
| 7561 | ORE.emit(OptDiag: OptimizationRemark(DEBUG_TYPE, "MemprofAttribute" , CBClone) |
| 7562 | << ore::NV("AllocationCall" , CBClone) << " in clone " |
| 7563 | << ore::NV("Caller" , CBClone->getFunction()) |
| 7564 | << " marked with memprof allocation attribute " |
| 7565 | << ore::NV("Attribute" , AllocTypeString)); |
| 7566 | } |
| 7567 | } else if (!CallsiteContext.empty()) { |
| 7568 | if (!CalledFunction) { |
| 7569 | #ifndef NDEBUG |
| 7570 | // We should have skipped inline assembly calls. |
| 7571 | auto *CI = dyn_cast<CallInst>(CB); |
| 7572 | assert(!CI || !CI->isInlineAsm()); |
| 7573 | #endif |
| 7574 | // We should have skipped direct calls via a Constant. |
| 7575 | assert(CalledValue && !isa<Constant>(CalledValue)); |
| 7576 | |
| 7577 | // This is an indirect call, see if we have profile information and |
| 7578 | // whether any clones were recorded for the profiled targets (that |
| 7579 | // we synthesized CallsiteInfo summary records for when building the |
| 7580 | // index). These records are the next ones in the function's |
| 7581 | // callsites, one per profiled target. |
| 7582 | auto NumClones = recordICPInfo( |
| 7583 | CB, |
| 7584 | GetCallsite: [&](const InstrProfValueData &Candidate) { |
| 7585 | assert(SI != FS->callsites().end()); |
| 7586 | #ifndef NDEBUG |
| 7587 | auto CalleeValueInfo = |
| 7588 | ImportSummary->getValueInfo(Candidate.Value); |
| 7589 | auto LocalGUID = |
| 7590 | ImportSummary->getGUIDFromOriginalID(Candidate.Value); |
| 7591 | #endif |
| 7592 | // We might not have a ValueInfo if this is a distributed |
| 7593 | // ThinLTO backend and decided not to import that function. |
| 7594 | // The thin link may also have updated the callee from the |
| 7595 | // original name GUID recorded in the VP metadata to the GUID |
| 7596 | // of the local target function (see |
| 7597 | // updateValueInfoForIndirectCalls). |
| 7598 | assert(!CalleeValueInfo || SI->Callee == CalleeValueInfo || |
| 7599 | (SI->Callee && SI->Callee.getGUID() == LocalGUID)); |
| 7600 | return &*(SI++); |
| 7601 | }, |
| 7602 | ICallAnalysisInfo); |
| 7603 | |
| 7604 | // Perform cloning if not yet done. This is done here in case |
| 7605 | // we don't need to do ICP, but might need to clone this |
| 7606 | // function as it is the target of other cloned calls. |
| 7607 | if (NumClones) |
| 7608 | CloneFuncIfNeeded(NumClones, FS); |
| 7609 | } |
| 7610 | |
| 7611 | else { |
| 7612 | // Consult the next callsite node. |
| 7613 | assert(SI != FS->callsites().end()); |
| 7614 | auto &StackNode = *(SI++); |
| 7615 | |
| 7616 | #ifndef NDEBUG |
| 7617 | // Sanity check that the stack ids match between the summary and |
| 7618 | // instruction metadata. |
| 7619 | auto StackIdIndexIter = StackNode.StackIdIndices.begin(); |
| 7620 | for (auto StackId : CallsiteContext) { |
| 7621 | assert(StackIdIndexIter != StackNode.StackIdIndices.end()); |
| 7622 | assert(ImportSummary->getStackIdAtIndex(*StackIdIndexIter) == |
| 7623 | StackId); |
| 7624 | StackIdIndexIter++; |
| 7625 | } |
| 7626 | #endif |
| 7627 | |
| 7628 | CloneCallsite(StackNode, CB, CalledFunction, FS); |
| 7629 | } |
| 7630 | } else if (CB->isTailCall() && CalledFunction) { |
| 7631 | // Locate the synthesized callsite info for the callee VI, if any was |
| 7632 | // created, and use that for cloning. |
| 7633 | ValueInfo CalleeVI = findValueInfoForGV( |
| 7634 | GV: *cast<GlobalValue>(Val: CalledValue), M, ImportSummary, CallingFunc: &F); |
| 7635 | if (CalleeVI && MapTailCallCalleeVIToCallsite.count(Val: CalleeVI)) { |
| 7636 | auto Callsite = MapTailCallCalleeVIToCallsite.find(Val: CalleeVI); |
| 7637 | assert(Callsite != MapTailCallCalleeVIToCallsite.end()); |
| 7638 | CloneCallsite(Callsite->second, CB, CalledFunction, FS); |
| 7639 | } |
| 7640 | } else if (CB->isTailCall() && !CalledFunction && |
| 7641 | !MapTailCallCalleeVIToCallsite.empty()) { |
| 7642 | // An indirect tail call, which the thin link may have identified as |
| 7643 | // being on profiled contexts with a missing frame, via the profiled |
| 7644 | // targets in its VP metadata. If so, the callsite infos synthesized |
| 7645 | // for those targets are used to perform ICP so that each clone of |
| 7646 | // this function calls the assigned clone of the target. |
| 7647 | auto PrevNumICallAnalysisInfo = ICallAnalysisInfo.size(); |
| 7648 | auto NumClones = recordICPInfo( |
| 7649 | CB, |
| 7650 | GetCallsite: [&](const InstrProfValueData &Candidate) { |
| 7651 | return findIndirectTailCallCallsite( |
| 7652 | CandidateGUID: Candidate.Value, MapTailCallCalleeVIToCallsite, |
| 7653 | OrigNameGUIDToTailCallCallsite, AliasVIToGV, M, |
| 7654 | ImportSummary, Symtab&: *Symtab); |
| 7655 | }, |
| 7656 | ICallAnalysisInfo); |
| 7657 | if (ICallAnalysisInfo.size() > PrevNumICallAnalysisInfo) |
| 7658 | IndirectTailCallsRecordedForICPThinBackend++; |
| 7659 | if (NumClones) |
| 7660 | CloneFuncIfNeeded(NumClones, FS); |
| 7661 | } |
| 7662 | } |
| 7663 | } |
| 7664 | |
| 7665 | // Now do any promotion required for cloning. |
| 7666 | performICP(M, VMaps, ICallAnalysisInfo, ORE); |
| 7667 | } |
| 7668 | |
| 7669 | // We skip some of the functions and instructions above, so remove all the |
| 7670 | // metadata in a single sweep here. |
| 7671 | for (auto &F : M) { |
| 7672 | // We can skip memprof clones because createFunctionClones already strips |
| 7673 | // the metadata from the newly created clones. |
| 7674 | if (F.isDeclaration() || isMemProfClone(GV: F)) |
| 7675 | continue; |
| 7676 | for (auto &BB : F) { |
| 7677 | for (auto &I : BB) { |
| 7678 | if (!isa<CallBase>(Val: I)) |
| 7679 | continue; |
| 7680 | I.setMetadata(KindID: LLVMContext::MD_memprof, Node: nullptr); |
| 7681 | I.setMetadata(KindID: LLVMContext::MD_callsite, Node: nullptr); |
| 7682 | } |
| 7683 | } |
| 7684 | } |
| 7685 | |
| 7686 | return Changed; |
| 7687 | } |
| 7688 | |
| 7689 | unsigned MemProfContextDisambiguation::recordICPInfo( |
| 7690 | CallBase *CB, |
| 7691 | function_ref<const CallsiteInfo *(const InstrProfValueData &)> GetCallsite, |
| 7692 | SmallVector<ICallAnalysisData> &ICallAnalysisInfo) { |
| 7693 | // First see if we have profile information for this indirect call. |
| 7694 | uint32_t NumCandidates; |
| 7695 | uint64_t TotalCount; |
| 7696 | auto CandidateProfileData = |
| 7697 | ICallAnalysis->getPromotionCandidatesForInstruction( |
| 7698 | I: CB, TotalCount, NumCandidates, |
| 7699 | MaxNumValueData: getMaxSummaryIndirectEdges(C: CB->getContext())); |
| 7700 | if (CandidateProfileData.empty()) |
| 7701 | return 0; |
| 7702 | |
| 7703 | // Iterate through all of the candidate profiled targets along with the |
| 7704 | // CallsiteInfo summary records synthesized for them, and see if any are |
| 7705 | // cloned and/or refer to clones. |
| 7706 | bool ICPNeeded = false; |
| 7707 | unsigned NumClones = 0; |
| 7708 | SmallVector<const CallsiteInfo *> Callsites; |
| 7709 | for (const auto &Candidate : CandidateProfileData) { |
| 7710 | const CallsiteInfo *Callsite = GetCallsite(Candidate); |
| 7711 | Callsites.push_back(Elt: Callsite); |
| 7712 | if (!Callsite) |
| 7713 | continue; |
| 7714 | // See if any of the clones of the indirect callsite for this |
| 7715 | // profiled target should call a cloned version of the profiled |
| 7716 | // target. We only need to do the ICP here if so. |
| 7717 | ICPNeeded |= llvm::any_of(Range: Callsite->Clones, |
| 7718 | P: [](unsigned CloneNo) { return CloneNo != 0; }); |
| 7719 | // Every callsite in the same function should have been cloned the same |
| 7720 | // number of times. |
| 7721 | assert(!NumClones || NumClones == Callsite->Clones.size()); |
| 7722 | NumClones = Callsite->Clones.size(); |
| 7723 | } |
| 7724 | if (!ICPNeeded) |
| 7725 | return NumClones; |
| 7726 | // Save information for ICP, which is performed later to avoid messing up the |
| 7727 | // current function traversal. |
| 7728 | ICallAnalysisInfo.push_back(Elt: {.CB: CB, .CandidateProfileData: CandidateProfileData.vec(), .NumCandidates: NumCandidates, |
| 7729 | .TotalCount: TotalCount, .NumClones: NumClones, .Callsites: std::move(Callsites)}); |
| 7730 | return NumClones; |
| 7731 | } |
| 7732 | |
| 7733 | void MemProfContextDisambiguation::performICP( |
| 7734 | Module &M, ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps, |
| 7735 | ArrayRef<ICallAnalysisData> ICallAnalysisInfo, |
| 7736 | OptimizationRemarkEmitter &ORE) { |
| 7737 | // Now do any promotion required for cloning. Specifically, for each |
| 7738 | // recorded ICP candidate (which was only recorded because one clone of that |
| 7739 | // candidate should call a cloned target), we perform ICP (speculative |
| 7740 | // devirtualization) for each clone of the callsite, and update its callee |
| 7741 | // to the appropriate clone. Note that the ICP compares against the original |
| 7742 | // version of the target, which is what is in the vtable. |
| 7743 | for (auto &Info : ICallAnalysisInfo) { |
| 7744 | auto *CB = Info.CB; |
| 7745 | auto TotalCount = Info.TotalCount; |
| 7746 | unsigned NumClones = Info.NumClones; |
| 7747 | SmallVector<InstrProfValueData, 8> RemainingCandidates; |
| 7748 | |
| 7749 | assert(Info.CandidateProfileData.size() == Info.Callsites.size()); |
| 7750 | for (unsigned I = 0; I < Info.CandidateProfileData.size(); I++) { |
| 7751 | auto &Candidate = Info.CandidateProfileData[I]; |
| 7752 | const CallsiteInfo *StackNode = Info.Callsites[I]; |
| 7753 | // All calls in the same function must have the same number of clones. |
| 7754 | assert(!StackNode || NumClones == StackNode->Clones.size()); |
| 7755 | |
| 7756 | // See if the target is in the module. If it wasn't imported, it is |
| 7757 | // possible that this profile could have been collected on a different |
| 7758 | // target (or version of the code), and we need to be conservative |
| 7759 | // (similar to what is done in the ICP pass). |
| 7760 | Function *TargetFunction = Symtab->getFunction(FuncMD5Hash: Candidate.Value); |
| 7761 | // If not found, try the GUID of the callee recorded in the summary. For |
| 7762 | // local targets the VP metadata may contain the GUID of the original |
| 7763 | // name without the source file prefix (e.g. if it was synthesized from |
| 7764 | // the MemProf profile's callee GUIDs or from a SamplePGO profile), which |
| 7765 | // the thin link translates to the actual GUID of the local function |
| 7766 | // (see updateValueInfoForIndirectCalls). The latter is recorded in the |
| 7767 | // function's !guid metadata, which is preserved across promotion and |
| 7768 | // importing, and is added to the Symtab. If there is no summary record, |
| 7769 | // the target was not cloned, so we conservatively only promote it if |
| 7770 | // found directly via the VP metadata GUID. |
| 7771 | if (!TargetFunction && StackNode && StackNode->Callee && |
| 7772 | StackNode->Callee.getGUID() != Candidate.Value) |
| 7773 | TargetFunction = Symtab->getFunction(FuncMD5Hash: StackNode->Callee.getGUID()); |
| 7774 | // The target may be the aliasee of an alias imported as a copy of the |
| 7775 | // aliasee, without the aliasee itself being imported (e.g. an IRPGO value |
| 7776 | // profile records the aliasee for calls through the alias, as the alias |
| 7777 | // has no profile data of its own). The Symtab keys such a copy by the |
| 7778 | // alias GUID (see initializeIndirectCallPromotionInfo), so look for one |
| 7779 | // with the target's GUID as its aliasee GUID. Promoting to the copy is |
| 7780 | // correct as it has the same address, and its clones are aliases of the |
| 7781 | // aliasee's clones in the defining module. |
| 7782 | if (!TargetFunction) { |
| 7783 | if (!AliaseeGUIDToImportedAlias) { |
| 7784 | AliaseeGUIDToImportedAlias.emplace(); |
| 7785 | for (Function &Fn : M) { |
| 7786 | if (isMemProfClone(GV: Fn) || !getImportedAliasGUID(F: Fn)) |
| 7787 | continue; |
| 7788 | if (auto AliaseeGUID = Fn.getGUIDIfAssigned()) |
| 7789 | AliaseeGUIDToImportedAlias->try_emplace(Key: *AliaseeGUID, Args: &Fn); |
| 7790 | } |
| 7791 | } |
| 7792 | TargetFunction = AliaseeGUIDToImportedAlias->lookup(Val: Candidate.Value); |
| 7793 | if (!TargetFunction && StackNode && StackNode->Callee && |
| 7794 | StackNode->Callee.getGUID() != Candidate.Value) |
| 7795 | TargetFunction = |
| 7796 | AliaseeGUIDToImportedAlias->lookup(Val: StackNode->Callee.getGUID()); |
| 7797 | } |
| 7798 | if (TargetFunction == nullptr || |
| 7799 | // Any ThinLTO global dead symbol removal should have already |
| 7800 | // occurred, so it should be safe to promote when the target is a |
| 7801 | // declaration. |
| 7802 | // TODO: Remove internal option once more fully tested. |
| 7803 | (MemProfRequireDefinitionForPromotion && |
| 7804 | TargetFunction->isDeclaration())) { |
| 7805 | ORE.emit(RemarkBuilder: [&]() { |
| 7806 | return OptimizationRemarkMissed(DEBUG_TYPE, "UnableToFindTarget" , CB) |
| 7807 | << "Memprof cannot promote indirect call: target with md5sum " |
| 7808 | << ore::NV("target md5sum" , Candidate.Value) << " not found" ; |
| 7809 | }); |
| 7810 | // FIXME: See if we can use the new declaration importing support to |
| 7811 | // at least get the declarations imported for this case. Hot indirect |
| 7812 | // targets should have been imported normally, however. |
| 7813 | RemainingCandidates.push_back(Elt: Candidate); |
| 7814 | continue; |
| 7815 | } |
| 7816 | |
| 7817 | // Check if legal to promote |
| 7818 | const char *Reason = nullptr; |
| 7819 | if (!isLegalToPromote(CB: *CB, Callee: TargetFunction, FailureReason: &Reason)) { |
| 7820 | ORE.emit(RemarkBuilder: [&]() { |
| 7821 | return OptimizationRemarkMissed(DEBUG_TYPE, "UnableToPromote" , CB) |
| 7822 | << "Memprof cannot promote indirect call to " |
| 7823 | << ore::NV("TargetFunction" , TargetFunction) |
| 7824 | << " with count of " << ore::NV("TotalCount" , TotalCount) |
| 7825 | << ": " << Reason; |
| 7826 | }); |
| 7827 | RemainingCandidates.push_back(Elt: Candidate); |
| 7828 | continue; |
| 7829 | } |
| 7830 | |
| 7831 | assert(!isMemProfClone(*TargetFunction)); |
| 7832 | |
| 7833 | // Handle each call clone, applying ICP so that each clone directly |
| 7834 | // calls the specified callee clone, guarded by the appropriate ICP |
| 7835 | // check. |
| 7836 | CallBase *CBClone = CB; |
| 7837 | for (unsigned J = 0; J < NumClones; J++) { |
| 7838 | // If the VMap is empty, this clone was a duplicate of another and was |
| 7839 | // created as an alias or a declaration. |
| 7840 | if (J > 0 && VMaps[J - 1]->empty()) |
| 7841 | continue; |
| 7842 | // Copy 0 is the original function. |
| 7843 | if (J > 0) |
| 7844 | CBClone = cast<CallBase>(Val&: (*VMaps[J - 1])[CB]); |
| 7845 | // We do the promotion using the original name, so that the comparison |
| 7846 | // is against the name in the vtable. Then just below, change the new |
| 7847 | // direct call to call the cloned function. |
| 7848 | auto &DirectCall = |
| 7849 | pgo::promoteIndirectCall(CB&: *CBClone, F: TargetFunction, Count: Candidate.Count, |
| 7850 | TotalCount, AttachProfToDirectCall: isSamplePGO, ORE: &ORE); |
| 7851 | auto *TargetToUse = TargetFunction; |
| 7852 | // Call original if this version calls the original version of its |
| 7853 | // callee (always the case if there is no summary record). |
| 7854 | if (StackNode && StackNode->Clones[J]) { |
| 7855 | TargetToUse = |
| 7856 | cast<Function>(Val: M.getOrInsertFunction( |
| 7857 | Name: getMemProfFuncName(Base: TargetFunction->getName(), |
| 7858 | CloneNo: StackNode->Clones[J]), |
| 7859 | T: TargetFunction->getFunctionType()) |
| 7860 | .getCallee()); |
| 7861 | } |
| 7862 | DirectCall.setCalledFunction(TargetToUse); |
| 7863 | // During matching we generate synthetic VP metadata for indirect calls |
| 7864 | // not already having any, from the memprof profile's callee GUIDs. If |
| 7865 | // we subsequently promote and inline those callees, we currently lose |
| 7866 | // the ability to generate this synthetic VP metadata. Optionally apply |
| 7867 | // a noinline attribute to promoted direct calls, where the threshold is |
| 7868 | // set to capture synthetic VP metadata targets which get a count of 1. |
| 7869 | if (MemProfICPNoInlineThreshold && |
| 7870 | Candidate.Count < MemProfICPNoInlineThreshold) |
| 7871 | DirectCall.setIsNoInline(); |
| 7872 | ORE.emit(OptDiag: OptimizationRemark(DEBUG_TYPE, "MemprofCall" , CBClone) |
| 7873 | << ore::NV("Call" , CBClone) << " in clone " |
| 7874 | << ore::NV("Caller" , CBClone->getFunction()) |
| 7875 | << " promoted and assigned to call function clone " |
| 7876 | << ore::NV("Callee" , TargetToUse)); |
| 7877 | } |
| 7878 | |
| 7879 | // Update TotalCount (all clones should get same count above) |
| 7880 | TotalCount -= Candidate.Count; |
| 7881 | } |
| 7882 | // Adjust the MD.prof metadata for all clones, now that we have the new |
| 7883 | // TotalCount and the remaining candidates. |
| 7884 | CallBase *CBClone = CB; |
| 7885 | for (unsigned J = 0; J < NumClones; J++) { |
| 7886 | // If the VMap is empty, this clone was a duplicate of another and was |
| 7887 | // created as an alias or a declaration. |
| 7888 | if (J > 0 && VMaps[J - 1]->empty()) |
| 7889 | continue; |
| 7890 | // Copy 0 is the original function. |
| 7891 | if (J > 0) |
| 7892 | CBClone = cast<CallBase>(Val&: (*VMaps[J - 1])[CB]); |
| 7893 | // First delete the old one. |
| 7894 | CBClone->setMetadata(KindID: LLVMContext::MD_prof, Node: nullptr); |
| 7895 | // If all promoted, we don't need the MD.prof metadata. |
| 7896 | // Otherwise we need update with the un-promoted records back. |
| 7897 | if (TotalCount != 0) |
| 7898 | annotateValueSite(M, Inst&: *CBClone, VDs: RemainingCandidates, Sum: TotalCount, |
| 7899 | ValueKind: IPVK_IndirectCallTarget, MaxMDCount: Info.NumCandidates); |
| 7900 | } |
| 7901 | } |
| 7902 | } |
| 7903 | |
| 7904 | template <typename DerivedCCG, typename FuncTy, typename CallTy> |
| 7905 | bool CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::process( |
| 7906 | function_ref<void(StringRef, StringRef, const Twine &)> , |
| 7907 | bool ) { |
| 7908 | if (DumpCCG) { |
| 7909 | dbgs() << "CCG before cloning:\n" ; |
| 7910 | dbgs() << *this; |
| 7911 | } |
| 7912 | if (ExportToDot) |
| 7913 | exportToDot(Label: "postbuild" ); |
| 7914 | |
| 7915 | if (VerifyCCG) { |
| 7916 | check(); |
| 7917 | } |
| 7918 | |
| 7919 | // The detail suffix needs the pre-cloning state of each context; only |
| 7920 | // capture it (extra memory and time) when it will be printed. |
| 7921 | if (MemProfReportHintedSizes && MemProfReportHintedSizesDetail) |
| 7922 | capturePreCloningContextInfo(); |
| 7923 | |
| 7924 | identifyClones(); |
| 7925 | |
| 7926 | if (VerifyCCG) { |
| 7927 | check(); |
| 7928 | } |
| 7929 | |
| 7930 | if (DumpCCG) { |
| 7931 | dbgs() << "CCG after cloning:\n" ; |
| 7932 | dbgs() << *this; |
| 7933 | } |
| 7934 | if (ExportToDot) |
| 7935 | exportToDot(Label: "cloned" ); |
| 7936 | |
| 7937 | bool Changed = assignFunctions(); |
| 7938 | |
| 7939 | if (DumpCCG) { |
| 7940 | dbgs() << "CCG after assigning function clones:\n" ; |
| 7941 | dbgs() << *this; |
| 7942 | } |
| 7943 | if (ExportToDot) |
| 7944 | exportToDot(Label: "clonefuncassign" ); |
| 7945 | |
| 7946 | // AllowExtraAnalysis is set when the hinted size messages were enabled via |
| 7947 | // remarks (-pass-remarks=memprof-context-disambiguation). |
| 7948 | // TODO: Consider also supporting the detail suffix for the remark output. |
| 7949 | if (MemProfReportHintedSizes || AllowExtraAnalysis) { |
| 7950 | // The referenced-function set can be large; build it only now, when it is |
| 7951 | // used by the detail suffix, rather than keeping it across cloning. |
| 7952 | if (MemProfReportHintedSizes && MemProfReportHintedSizesDetail) |
| 7953 | populateFunctionsWithCallers(); |
| 7954 | printTotalSizes(OS&: errs(), EmitRemark); |
| 7955 | } |
| 7956 | |
| 7957 | return Changed; |
| 7958 | } |
| 7959 | |
| 7960 | bool MemProfContextDisambiguation::( |
| 7961 | Module &M, |
| 7962 | llvm::function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter) { |
| 7963 | |
| 7964 | // If we have an import summary, then the cloning decisions were made during |
| 7965 | // the thin link on the index. Apply them and return. |
| 7966 | if (ImportSummary) |
| 7967 | return applyImport(M); |
| 7968 | |
| 7969 | // TODO: If/when other types of memprof cloning are enabled beyond just for |
| 7970 | // hot and cold, we will need to change this to individually control the |
| 7971 | // AllocationType passed to addStackNodesForMIB during CCG construction. |
| 7972 | // Note that we specifically check this after applying imports above, so that |
| 7973 | // the option isn't needed to be passed to distributed ThinLTO backend |
| 7974 | // clang processes, which won't necessarily have visibility into the linker |
| 7975 | // dependences. Instead the information is communicated from the LTO link to |
| 7976 | // the backends via the combined summary index. |
| 7977 | if (!SupportsHotColdNew) |
| 7978 | return false; |
| 7979 | |
| 7980 | ModuleCallsiteContextGraph CCG(M, OREGetter); |
| 7981 | // TODO: Set up remarks for regular LTO. We need to decide what function to |
| 7982 | // use in the callback. |
| 7983 | return CCG.process(); |
| 7984 | } |
| 7985 | |
| 7986 | MemProfContextDisambiguation::MemProfContextDisambiguation( |
| 7987 | const ModuleSummaryIndex *Summary, bool isSamplePGO) |
| 7988 | : ImportSummary(Summary), isSamplePGO(isSamplePGO) { |
| 7989 | // Check the dot graph printing options once here, to make sure we have valid |
| 7990 | // and expected combinations. |
| 7991 | if (DotGraphScope == DotScope::Alloc && !AllocIdForDot.getNumOccurrences()) |
| 7992 | llvm::report_fatal_error( |
| 7993 | reason: "-memprof-dot-scope=alloc requires -memprof-dot-alloc-id" ); |
| 7994 | if (DotGraphScope == DotScope::Context && |
| 7995 | !ContextIdForDot.getNumOccurrences()) |
| 7996 | llvm::report_fatal_error( |
| 7997 | reason: "-memprof-dot-scope=context requires -memprof-dot-context-id" ); |
| 7998 | if (DotGraphScope == DotScope::All && AllocIdForDot.getNumOccurrences() && |
| 7999 | ContextIdForDot.getNumOccurrences()) |
| 8000 | llvm::report_fatal_error( |
| 8001 | reason: "-memprof-dot-scope=all can't have both -memprof-dot-alloc-id and " |
| 8002 | "-memprof-dot-context-id" ); |
| 8003 | if (ImportSummary) { |
| 8004 | // The MemProfImportSummary should only be used for testing ThinLTO |
| 8005 | // distributed backend handling via opt, in which case we don't have a |
| 8006 | // summary from the pass pipeline. |
| 8007 | assert(MemProfImportSummary.empty()); |
| 8008 | return; |
| 8009 | } |
| 8010 | if (MemProfImportSummary.empty()) |
| 8011 | return; |
| 8012 | |
| 8013 | auto ReadSummaryFile = |
| 8014 | errorOrToExpected(EO: MemoryBuffer::getFile(Filename: MemProfImportSummary)); |
| 8015 | if (!ReadSummaryFile) { |
| 8016 | logAllUnhandledErrors(E: ReadSummaryFile.takeError(), OS&: errs(), |
| 8017 | ErrorBanner: "Error loading file '" + MemProfImportSummary + |
| 8018 | "': " ); |
| 8019 | return; |
| 8020 | } |
| 8021 | auto ImportSummaryForTestingOrErr = getModuleSummaryIndex(Buffer: **ReadSummaryFile); |
| 8022 | if (!ImportSummaryForTestingOrErr) { |
| 8023 | logAllUnhandledErrors(E: ImportSummaryForTestingOrErr.takeError(), OS&: errs(), |
| 8024 | ErrorBanner: "Error parsing file '" + MemProfImportSummary + |
| 8025 | "': " ); |
| 8026 | return; |
| 8027 | } |
| 8028 | ImportSummaryForTesting = std::move(*ImportSummaryForTestingOrErr); |
| 8029 | ImportSummary = ImportSummaryForTesting.get(); |
| 8030 | } |
| 8031 | |
| 8032 | PreservedAnalyses MemProfContextDisambiguation::run(Module &M, |
| 8033 | ModuleAnalysisManager &AM) { |
| 8034 | auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager(); |
| 8035 | auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & { |
| 8036 | return FAM.getResult<OptimizationRemarkEmitterAnalysis>(IR&: *F); |
| 8037 | }; |
| 8038 | if (!processModule(M, OREGetter)) |
| 8039 | return PreservedAnalyses::all(); |
| 8040 | return PreservedAnalyses::none(); |
| 8041 | } |
| 8042 | |
| 8043 | // After the thin link cloning decisions have been recorded in the summary |
| 8044 | // callsite records, mark the summary indirect call edges to any callee that a |
| 8045 | // callsite was assigned to call a (non-original) clone of as requiring a |
| 8046 | // declaration import. |
| 8047 | // See the comments for the MemProfImportClonedIndirectTargets option. |
| 8048 | static void updateClonedCalleeEdgesForImport( |
| 8049 | ModuleSummaryIndex &Index, |
| 8050 | llvm::function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> |
| 8051 | isPrevailing) { |
| 8052 | // Map from callee to the corresponding call edge in the current function |
| 8053 | // summary, so that we don't need to walk the function's call edges for each |
| 8054 | // callsite assigned to call a cloned callee. Note that there is a single call |
| 8055 | // edge per callee. The map is built lazily, the first time we encounter such |
| 8056 | // a callsite in a given function summary, since most functions have none |
| 8057 | // (the map is cleared before processing each function summary). The |
| 8058 | // pointers into the call edge vector remain valid since it isn't modified |
| 8059 | // while the map is in use. |
| 8060 | SmallDenseMap<ValueInfo, CalleeInfo *> CalleeToEdge; |
| 8061 | for (auto &I : Index) { |
| 8062 | auto VI = Index.getValueInfo(R: I); |
| 8063 | for (auto &S : VI.getSummaryList()) { |
| 8064 | // Only the prevailing copies were added to the CCG (see the |
| 8065 | // IndexCallsiteContextGraph constructor). |
| 8066 | if (!GlobalValue::isLocalLinkage(Linkage: S->linkage()) && |
| 8067 | !isPrevailing(VI.getGUID(), S.get())) |
| 8068 | continue; |
| 8069 | auto *FS = dyn_cast<FunctionSummary>(Val: S.get()); |
| 8070 | if (!FS || FS->callsites().empty()) |
| 8071 | continue; |
| 8072 | CalleeToEdge.clear(); |
| 8073 | for (auto &CI : FS->callsites()) { |
| 8074 | if (!CI.Callee || llvm::all_of(Range: CI.Clones, P: [](unsigned CloneNo) { |
| 8075 | return CloneNo == 0; |
| 8076 | })) |
| 8077 | continue; |
| 8078 | // Note that we mark the edge even if the callee is defined in the |
| 8079 | // same module as the caller, since the caller may be imported into |
| 8080 | // other modules. The function importer will skip importing the callee |
| 8081 | // into any module that already defines it. |
| 8082 | NumClonedCalleeCallsites++; |
| 8083 | // Lazily build the map on the first such callsite in this function. |
| 8084 | if (CalleeToEdge.empty()) |
| 8085 | for (auto &[CalleeVI, Edge] : FS->mutableCalls()) |
| 8086 | CalleeToEdge[CalleeVI] = &Edge; |
| 8087 | auto It = CalleeToEdge.find(Val: CI.Callee); |
| 8088 | if (It == CalleeToEdge.end()) |
| 8089 | continue; |
| 8090 | CalleeInfo *Edge = It->second; |
| 8091 | // Direct calls don't require importing the callee. Note that the |
| 8092 | // edge is marked indirect if any call to the callee in this function |
| 8093 | // is indirect, in which case we can't distinguish whether this |
| 8094 | // callsite is direct and conservatively treat it as indirect. |
| 8095 | if (!Edge->hasIndirectCall()) |
| 8096 | continue; |
| 8097 | NumClonedCalleeIndirectCallsites++; |
| 8098 | if (!Edge->requiresDeclImport()) { |
| 8099 | Edge->setRequiresDeclImport(true); |
| 8100 | NumClonedCalleeEdgesMarkedRequiresDeclImport++; |
| 8101 | } |
| 8102 | } |
| 8103 | } |
| 8104 | } |
| 8105 | } |
| 8106 | |
| 8107 | void MemProfContextDisambiguation::run( |
| 8108 | ModuleSummaryIndex &Index, |
| 8109 | llvm::function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> |
| 8110 | isPrevailing, |
| 8111 | LLVMContext &Ctx, |
| 8112 | function_ref<void(StringRef, StringRef, const Twine &)> ) { |
| 8113 | // TODO: If/when other types of memprof cloning are enabled beyond just for |
| 8114 | // hot and cold, we will need to change this to individually control the |
| 8115 | // AllocationType passed to addStackNodesForMIB during CCG construction. |
| 8116 | // The index was set from the option, so these should be in sync. |
| 8117 | assert(Index.withSupportsHotColdNew() == SupportsHotColdNew); |
| 8118 | if (!SupportsHotColdNew) |
| 8119 | return; |
| 8120 | |
| 8121 | bool = |
| 8122 | OptimizationRemarkEmitter::allowExtraAnalysis(Ctx, DEBUG_TYPE); |
| 8123 | |
| 8124 | bool Changed = false; |
| 8125 | { |
| 8126 | IndexCallsiteContextGraph CCG(Index, isPrevailing); |
| 8127 | Changed = CCG.process(EmitRemark, AllowExtraAnalysis); |
| 8128 | } |
| 8129 | // Note this must be done after the CCG is destroyed, which is when the |
| 8130 | // callsite records synthesized for missing tail call frames are added to the |
| 8131 | // function summaries. Importing declarations is only useful if MemProf ICP |
| 8132 | // is allowed to promote to a declaration of the target. |
| 8133 | if (Changed && MemProfImportClonedIndirectTargets && |
| 8134 | !MemProfRequireDefinitionForPromotion) |
| 8135 | updateClonedCalleeEdgesForImport(Index, isPrevailing); |
| 8136 | } |
| 8137 | |
| 8138 | // Strips MemProf attributes and metadata. Can be invoked by the pass pipeline |
| 8139 | // when we don't have an index that has recorded that we are linking with |
| 8140 | // allocation libraries containing the necessary APIs for downstream |
| 8141 | // transformations. |
| 8142 | PreservedAnalyses MemProfRemoveInfo::run(Module &M, ModuleAnalysisManager &AM) { |
| 8143 | // The profile matcher applies hotness attributes directly for allocations, |
| 8144 | // and those will cause us to generate calls to the hot/cold interfaces |
| 8145 | // unconditionally. If supports-hot-cold-new was not enabled in the LTO |
| 8146 | // link then assume we don't want these calls (e.g. not linking with |
| 8147 | // the appropriate library, or otherwise trying to disable this behavior). |
| 8148 | bool Changed = false; |
| 8149 | for (auto &F : M) { |
| 8150 | for (auto &BB : F) { |
| 8151 | for (auto &I : BB) { |
| 8152 | auto *CI = dyn_cast<CallBase>(Val: &I); |
| 8153 | if (!CI) |
| 8154 | continue; |
| 8155 | if (CI->hasFnAttr(Kind: "memprof" )) { |
| 8156 | CI->removeFnAttr(Kind: "memprof" ); |
| 8157 | Changed = true; |
| 8158 | } |
| 8159 | if (!CI->hasMetadata(KindID: LLVMContext::MD_callsite)) { |
| 8160 | assert(!CI->hasMetadata(LLVMContext::MD_memprof)); |
| 8161 | continue; |
| 8162 | } |
| 8163 | // Strip off all memprof metadata as it is no longer needed. |
| 8164 | // Importantly, this avoids the addition of new memprof attributes |
| 8165 | // after inlining propagation. |
| 8166 | CI->setMetadata(KindID: LLVMContext::MD_memprof, Node: nullptr); |
| 8167 | CI->setMetadata(KindID: LLVMContext::MD_callsite, Node: nullptr); |
| 8168 | Changed = true; |
| 8169 | } |
| 8170 | } |
| 8171 | } |
| 8172 | if (!Changed) |
| 8173 | return PreservedAnalyses::all(); |
| 8174 | return PreservedAnalyses::none(); |
| 8175 | } |
| 8176 | |