[llvm] Add a pass to collect dropped var stats for MIR. (PR #115566)

via llvm-commits llvm-commits at lists.llvm.org
Fri Nov 8 15:30:33 PST 2024


llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-ir

Author: Shubham Sandeep Rastogi (rastogishubham)

<details>
<summary>Changes</summary>

This patch uses the DroppedVariableStats class to add dropped variable
statistics for MIR passes.

---

Patch is 75.74 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/115566.diff


11 Files Affected:

- (added) llvm/include/llvm/CodeGen/DroppedVariableStats.h (+199) 
- (modified) llvm/include/llvm/CodeGen/MachineFunctionPass.h (+2) 
- (modified) llvm/include/llvm/Passes/StandardInstrumentations.h (+2-78) 
- (modified) llvm/lib/CodeGen/CMakeLists.txt (+1) 
- (added) llvm/lib/CodeGen/DroppedVariableStats.cpp (+222) 
- (modified) llvm/lib/CodeGen/MachineFunctionPass.cpp (+14-2) 
- (modified) llvm/lib/Passes/StandardInstrumentations.cpp (+2-178) 
- (modified) llvm/unittests/IR/CMakeLists.txt (+1-1) 
- (renamed) llvm/unittests/IR/DroppedVariableStatsIRTest.cpp (+28-42) 
- (modified) llvm/unittests/MIR/CMakeLists.txt (+1) 
- (added) llvm/unittests/MIR/DroppedVariableStatsMIRTest.cpp (+1067) 


``````````diff
diff --git a/llvm/include/llvm/CodeGen/DroppedVariableStats.h b/llvm/include/llvm/CodeGen/DroppedVariableStats.h
new file mode 100644
index 00000000000000..8969105dcb08c2
--- /dev/null
+++ b/llvm/include/llvm/CodeGen/DroppedVariableStats.h
@@ -0,0 +1,199 @@
+///===- DroppedVariableStats.h - Opt Diagnostics -*- C++ -*----------------===//
+///
+/// Part of the LLVM Project, under the Apache License v2.0 with LLVM
+/// Exceptions. See https://llvm.org/LICENSE.txt for license information.
+/// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+///
+///===---------------------------------------------------------------------===//
+/// \file
+/// Dropped Variable Statistics for Debug Information. Reports any number
+/// of #dbg_values or DBG_VALUEs that get dropped due to an optimization pass.
+///
+///===---------------------------------------------------------------------===//
+
+#ifndef LLVM_CODEGEN_DROPPEDVARIABLESTATS_H
+#define LLVM_CODEGEN_DROPPEDVARIABLESTATS_H
+
+#include "llvm/CodeGen/MachinePassManager.h"
+#include "llvm/IR/DebugInfoMetadata.h"
+#include "llvm/IR/DiagnosticInfo.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/PassInstrumentation.h"
+
+namespace llvm {
+
+/// A unique key that represents a #dbg_value.
+using VarID =
+    std::tuple<const DIScope *, const DIScope *, const DILocalVariable *>;
+
+/// A base class to collect and print dropped debug information variable
+/// statistics.
+class DroppedVariableStats {
+public:
+  DroppedVariableStats(bool DroppedVarStatsEnabled)
+      : DroppedVariableStatsEnabled(DroppedVarStatsEnabled) {
+    if (DroppedVarStatsEnabled)
+      llvm::outs()
+          << "Pass Level, Pass Name, Num of Dropped Variables, Func or "
+             "Module Name\n";
+  };
+  // We intend this to be unique per-compilation, thus no copies.
+  DroppedVariableStats(const DroppedVariableStats &) = delete;
+  void operator=(const DroppedVariableStats &) = delete;
+
+  void setup() {
+    DebugVariablesStack.push_back(
+        {DenseMap<const Function *, DebugVariables>()});
+    InlinedAts.push_back(
+        {DenseMap<StringRef, DenseMap<VarID, DILocation *>>()});
+    return;
+  }
+
+  void cleanup() {
+    DebugVariablesStack.pop_back();
+    InlinedAts.pop_back();
+    return;
+  }
+
+  bool getPassDroppedVariables() { return PassDroppedVariables; }
+
+protected:
+  bool PassDroppedVariables = false;
+  bool DroppedVariableStatsEnabled = false;
+
+  struct DebugVariables {
+    /// DenseSet of VarIDs before an optimization pass has run.
+    DenseSet<VarID> DebugVariablesBefore;
+    /// DenseSet of VarIDs after an optimization pass has run.
+    DenseSet<VarID> DebugVariablesAfter;
+  };
+
+  /// A stack of a DenseMap, that maps DebugVariables for every pass to an
+  /// llvm::Function. A stack is used because an optimization pass can call
+  /// other passes.
+  SmallVector<DenseMap<const Function *, DebugVariables>> DebugVariablesStack;
+
+  /// A DenseSet tracking whether a scope was visited before.
+  DenseSet<const DIScope *> VisitedScope;
+  /// A stack of DenseMaps, which map the name of an llvm::Function to a
+  /// DenseMap of VarIDs and their inlinedAt locations before an optimization
+  /// pass has run.
+  SmallVector<DenseMap<StringRef, DenseMap<VarID, DILocation *>>> InlinedAts;
+  /// Remove a dropped #dbg_value VarID from all Sets in the
+  /// DroppedVariablesBefore stack.
+  void removeVarFromAllSets(VarID Var, const Function *F) {
+    // Do not remove Var from the last element, it will be popped from the
+    // stack.
+    for (auto &DebugVariablesMap : llvm::drop_end(DebugVariablesStack))
+      DebugVariablesMap[F].DebugVariablesBefore.erase(Var);
+  }
+  /// Return true if \p Scope is the same as \p DbgValScope or a child scope of
+  /// \p DbgValScope, return false otherwise.
+  bool isScopeChildOfOrEqualTo(DIScope *Scope, const DIScope *DbgValScope);
+  /// Return true if \p InlinedAt is the same as \p DbgValInlinedAt or part of
+  /// the InlinedAt chain, return false otherwise.
+  bool isInlinedAtChildOfOrEqualTo(const DILocation *InlinedAt,
+                                   const DILocation *DbgValInlinedAt);
+};
+
+/// A class to collect and print dropped debug information due to LLVM IR
+/// optimization passes. After every LLVM IR pass is run, it will print how many
+/// #dbg_values were dropped due to that pass.
+class DroppedVariableStatsIR : public DroppedVariableStats {
+public:
+  DroppedVariableStatsIR(bool DroppedVarStatsEnabled)
+      : llvm::DroppedVariableStats(DroppedVarStatsEnabled) {}
+
+  void runBeforePass(Any IR) {
+    setup();
+    if (const auto *M = unwrapIR<Module>(IR))
+      return this->runOnModule(M, true);
+    if (const auto *F = unwrapIR<Function>(IR))
+      return this->runOnFunction(F, true);
+  }
+
+  void runAfterPass(StringRef P, Any IR) {
+    if (const auto *M = unwrapIR<Module>(IR))
+      runAfterPassModule(P, M);
+    else if (const auto *F = unwrapIR<Function>(IR))
+      runAfterPassFunction(P, F);
+    return cleanup();
+  }
+
+  void registerCallbacks(PassInstrumentationCallbacks &PIC);
+
+private:
+  void runAfterPassFunction(StringRef PassID, const Function *F) {
+    runOnFunction(F, false);
+    calculateDroppedVarStatsOnFunction(F, PassID, F->getName().str(),
+                                       "Function");
+  }
+
+  void runAfterPassModule(StringRef PassID, const Module *M) {
+    runOnModule(M, false);
+    calculateDroppedVarStatsOnModule(M, PassID, M->getName().str(), "Module");
+  }
+  /// Populate DebugVariablesBefore, DebugVariablesAfter, InlinedAts before or
+  /// after a pass has run to facilitate dropped variable calculation for an
+  /// llvm::Function.
+  void runOnFunction(const Function *F, bool Before);
+  /// Iterate over all Instructions in a Function and report any dropped debug
+  /// information.
+  void calculateDroppedVarStatsOnFunction(const Function *F, StringRef PassID,
+                                          std::string FuncOrModName,
+                                          std::string PassLevel);
+  /// Populate DebugVariablesBefore, DebugVariablesAfter, InlinedAts before or
+  /// after a pass has run to facilitate dropped variable calculation for an
+  /// llvm::Module. Calls runOnFunction on every Function in the Module.
+  void runOnModule(const Module *M, bool Before);
+  /// Iterate over all Functions in a Module and report any dropped debug
+  /// information. Will call calculateDroppedVarStatsOnFunction on every
+  /// Function.
+  void calculateDroppedVarStatsOnModule(const Module *M, StringRef PassID,
+                                        std::string FuncOrModName,
+                                        std::string PassLevel);
+
+  template <typename IRUnitT> static const IRUnitT *unwrapIR(Any IR) {
+    const IRUnitT **IRPtr = llvm::any_cast<const IRUnitT *>(&IR);
+    return IRPtr ? *IRPtr : nullptr;
+  }
+};
+
+/// A class to collect and print dropped debug information due to MIR
+/// optimization passes. After every MIR pass is run, it will print how many
+/// #DBG_VALUEs were dropped due to that pass.
+class DroppedVariableStatsMIR : public DroppedVariableStats {
+public:
+  DroppedVariableStatsMIR() : llvm::DroppedVariableStats(false) {}
+
+  void runBeforePass(StringRef PassID, MachineFunction *MF) {
+    if (PassID == "Debug Variable Analysis")
+      return;
+    setup();
+    return runOnMachineFunction(MF, true);
+  }
+
+  void runAfterPass(StringRef PassID, MachineFunction *MF) {
+    if (PassID == "Debug Variable Analysis")
+      return;
+    runOnMachineFunction(MF, false);
+    calculateDroppedVarStatsOnMachineFunction(MF, PassID, MF->getName().str());
+    cleanup();
+  }
+
+private:
+  /// Populate DebugVariablesBefore, DebugVariablesAfter, InlinedAts before or
+  /// after a pass has run to facilitate dropped variable calculation for an
+  /// llvm::MachineFunction.
+  void runOnMachineFunction(const MachineFunction *MF, bool Before);
+  /// Iterate over all Instructions in a MachineFunction and report any dropped
+  /// debug information.
+  void calculateDroppedVarStatsOnMachineFunction(const MachineFunction *MF,
+                                                 StringRef PassID,
+                                                 std::string FuncOrModName);
+};
+
+} // namespace llvm
+
+#endif
diff --git a/llvm/include/llvm/CodeGen/MachineFunctionPass.h b/llvm/include/llvm/CodeGen/MachineFunctionPass.h
index caaf22c2139e31..2ab6fb203da874 100644
--- a/llvm/include/llvm/CodeGen/MachineFunctionPass.h
+++ b/llvm/include/llvm/CodeGen/MachineFunctionPass.h
@@ -18,6 +18,7 @@
 #ifndef LLVM_CODEGEN_MACHINEFUNCTIONPASS_H
 #define LLVM_CODEGEN_MACHINEFUNCTIONPASS_H
 
+#include "llvm/CodeGen/DroppedVariableStats.h"
 #include "llvm/CodeGen/MachineFunction.h"
 #include "llvm/Pass.h"
 
@@ -73,6 +74,7 @@ class MachineFunctionPass : public FunctionPass {
                           const std::string &Banner) const override;
 
   bool runOnFunction(Function &F) override;
+  DroppedVariableStatsMIR DroppedVarStatsMF;
 };
 
 } // End llvm namespace
diff --git a/llvm/include/llvm/Passes/StandardInstrumentations.h b/llvm/include/llvm/Passes/StandardInstrumentations.h
index 9301a12c740eec..12a34c099eaffe 100644
--- a/llvm/include/llvm/Passes/StandardInstrumentations.h
+++ b/llvm/include/llvm/Passes/StandardInstrumentations.h
@@ -19,6 +19,7 @@
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/ADT/StringSet.h"
+#include "llvm/CodeGen/DroppedVariableStats.h"
 #include "llvm/CodeGen/MachineBasicBlock.h"
 #include "llvm/IR/BasicBlock.h"
 #include "llvm/IR/DebugInfoMetadata.h"
@@ -579,83 +580,6 @@ class PrintCrashIRInstrumentation {
   static void SignalHandler(void *);
 };
 
-/// A class to collect and print dropped debug information variable statistics.
-/// After every LLVM IR pass is run, it will print how many #dbg_values were
-/// dropped due to that pass.
-class DroppedVariableStats {
-public:
-  DroppedVariableStats(bool DroppedVarStatsEnabled) {
-    if (DroppedVarStatsEnabled)
-      llvm::outs()
-          << "Pass Level, Pass Name, Num of Dropped Variables, Func or "
-             "Module Name\n";
-  };
-  // We intend this to be unique per-compilation, thus no copies.
-  DroppedVariableStats(const DroppedVariableStats &) = delete;
-  void operator=(const DroppedVariableStats &) = delete;
-
-  void registerCallbacks(PassInstrumentationCallbacks &PIC);
-  void runBeforePass(StringRef PassID, Any IR);
-  void runAfterPass(StringRef PassID, Any IR, const PreservedAnalyses &PA);
-  void runAfterPassInvalidated(StringRef PassID, const PreservedAnalyses &PA);
-  bool getPassDroppedVariables() { return PassDroppedVariables; }
-
-private:
-  bool PassDroppedVariables = false;
-  /// A unique key that represents a #dbg_value.
-  using VarID =
-      std::tuple<const DIScope *, const DIScope *, const DILocalVariable *>;
-
-  struct DebugVariables {
-    /// DenseSet of VarIDs before an optimization pass has run.
-    DenseSet<VarID> DebugVariablesBefore;
-    /// DenseSet of VarIDs after an optimization pass has run.
-    DenseSet<VarID> DebugVariablesAfter;
-  };
-
-  /// A stack of a DenseMap, that maps DebugVariables for every pass to an
-  /// llvm::Function. A stack is used because an optimization pass can call
-  /// other passes.
-  SmallVector<DenseMap<const Function *, DebugVariables>> DebugVariablesStack;
-
-  /// A DenseSet tracking whether a scope was visited before.
-  DenseSet<const DIScope *> VisitedScope;
-  /// A stack of DenseMaps, which map the name of an llvm::Function to a
-  /// DenseMap of VarIDs and their inlinedAt locations before an optimization
-  /// pass has run.
-  SmallVector<DenseMap<StringRef, DenseMap<VarID, DILocation *>>> InlinedAts;
-
-  /// Iterate over all Functions in a Module and report any dropped debug
-  /// information. Will call calculateDroppedVarStatsOnFunction on every
-  /// Function.
-  void calculateDroppedVarStatsOnModule(const Module *M, StringRef PassID,
-                                        std::string FuncOrModName,
-                                        std::string PassLevel);
-  /// Iterate over all Instructions in a Function and report any dropped debug
-  /// information.
-  void calculateDroppedVarStatsOnFunction(const Function *F, StringRef PassID,
-                                          std::string FuncOrModName,
-                                          std::string PassLevel);
-  /// Populate DebugVariablesBefore, DebugVariablesAfter, InlinedAts before or
-  /// after a pass has run to facilitate dropped variable calculation for an
-  /// llvm::Function.
-  void runOnFunction(const Function *F, bool Before);
-  /// Populate DebugVariablesBefore, DebugVariablesAfter, InlinedAts before or
-  /// after a pass has run to facilitate dropped variable calculation for an
-  /// llvm::Module. Calls runOnFunction on every Function in the Module.
-  void runOnModule(const Module *M, bool Before);
-  /// Remove a dropped #dbg_value VarID from all Sets in the
-  /// DroppedVariablesBefore stack.
-  void removeVarFromAllSets(VarID Var, const Function *F);
-  /// Return true if \p Scope is the same as \p DbgValScope or a child scope of
-  /// \p DbgValScope, return false otherwise.
-  bool isScopeChildOfOrEqualTo(DIScope *Scope, const DIScope *DbgValScope);
-  /// Return true if \p InlinedAt is the same as \p DbgValInlinedAt or part of
-  /// the InlinedAt chain, return false otherwise.
-  bool isInlinedAtChildOfOrEqualTo(const DILocation *InlinedAt,
-                                   const DILocation *DbgValInlinedAt);
-};
-
 /// This class provides an interface to register all the standard pass
 /// instrumentations and manages their state (if any).
 class StandardInstrumentations {
@@ -673,7 +597,7 @@ class StandardInstrumentations {
   PrintCrashIRInstrumentation PrintCrashIR;
   IRChangedTester ChangeTester;
   VerifyInstrumentation Verify;
-  DroppedVariableStats DroppedStats;
+  DroppedVariableStatsIR DroppedStatsIR;
 
   bool VerifyEach;
 
diff --git a/llvm/lib/CodeGen/CMakeLists.txt b/llvm/lib/CodeGen/CMakeLists.txt
index 5a17944db0ae03..3eebeaaf04b3a2 100644
--- a/llvm/lib/CodeGen/CMakeLists.txt
+++ b/llvm/lib/CodeGen/CMakeLists.txt
@@ -50,6 +50,7 @@ add_llvm_component_library(LLVMCodeGen
   DeadMachineInstructionElim.cpp
   DetectDeadLanes.cpp
   DFAPacketizer.cpp
+  DroppedVariableStats.cpp
   DwarfEHPrepare.cpp
   EarlyIfConversion.cpp
   EdgeBundles.cpp
diff --git a/llvm/lib/CodeGen/DroppedVariableStats.cpp b/llvm/lib/CodeGen/DroppedVariableStats.cpp
new file mode 100644
index 00000000000000..b8130784d28874
--- /dev/null
+++ b/llvm/lib/CodeGen/DroppedVariableStats.cpp
@@ -0,0 +1,222 @@
+///===- DroppedVariableStats.cpp - Opt Diagnostic -*- C++ -*---------------===//
+///
+/// Part of the LLVM Project, under the Apache License v2.0 with LLVM
+/// Exceptions. See https://llvm.org/LICENSE.txt for license information.
+/// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+///
+///===---------------------------------------------------------------------===//
+/// \file
+/// Dropped Variable Statistics for Debug Information. Reports any number
+/// of #dbg_values or DBG_VALUEs that get dropped due to an optimization pass.
+///
+///===---------------------------------------------------------------------===//
+
+#include "llvm/CodeGen/DroppedVariableStats.h"
+#include "llvm/IR/DebugInfoMetadata.h"
+#include "llvm/IR/InstIterator.h"
+#include "llvm/IR/Module.h"
+
+using namespace llvm;
+
+bool DroppedVariableStats::isScopeChildOfOrEqualTo(DIScope *Scope,
+                                                   const DIScope *DbgValScope) {
+  while (Scope != nullptr) {
+    if (VisitedScope.find(Scope) == VisitedScope.end()) {
+      VisitedScope.insert(Scope);
+      if (Scope == DbgValScope) {
+        VisitedScope.clear();
+        return true;
+      }
+      Scope = Scope->getScope();
+    } else {
+      VisitedScope.clear();
+      return false;
+    }
+  }
+  return false;
+}
+
+bool DroppedVariableStats::isInlinedAtChildOfOrEqualTo(
+    const DILocation *InlinedAt, const DILocation *DbgValInlinedAt) {
+  if (DbgValInlinedAt == InlinedAt)
+    return true;
+  if (!DbgValInlinedAt)
+    return false;
+  if (!InlinedAt)
+    return false;
+  auto *IA = InlinedAt;
+  while (IA) {
+    if (IA == DbgValInlinedAt)
+      return true;
+    IA = IA->getInlinedAt();
+  }
+  return false;
+}
+
+void DroppedVariableStatsIR::runOnFunction(const Function *F, bool Before) {
+  auto &DebugVariables = DebugVariablesStack.back()[F];
+  auto &VarIDSet = (Before ? DebugVariables.DebugVariablesBefore
+                           : DebugVariables.DebugVariablesAfter);
+  auto &InlinedAtsMap = InlinedAts.back();
+  auto FuncName = F->getName();
+  if (Before)
+    InlinedAtsMap.try_emplace(FuncName, DenseMap<VarID, DILocation *>());
+  VarIDSet = DenseSet<VarID>();
+  for (const auto &I : instructions(F)) {
+    for (DbgRecord &DR : I.getDbgRecordRange()) {
+      if (auto *Dbg = dyn_cast<DbgVariableRecord>(&DR)) {
+        auto *DbgVar = Dbg->getVariable();
+        auto DbgLoc = DR.getDebugLoc();
+        VarID Key{DbgVar->getScope(), DbgLoc->getInlinedAtScope(), DbgVar};
+        VarIDSet.insert(Key);
+        if (Before)
+          InlinedAtsMap[FuncName].try_emplace(Key, DbgLoc.getInlinedAt());
+      }
+    }
+  }
+}
+
+void DroppedVariableStatsIR::calculateDroppedVarStatsOnFunction(
+    const Function *F, StringRef PassID, std::string FuncOrModName,
+    std::string PassLevel) {
+  unsigned DroppedCount = 0;
+  StringRef FuncName = F->getName();
+  DebugVariables &DbgVariables = DebugVariablesStack.back()[F];
+  DenseSet<VarID> &DebugVariablesBeforeSet = DbgVariables.DebugVariablesBefore;
+  DenseSet<VarID> &DebugVariablesAfterSet = DbgVariables.DebugVariablesAfter;
+  DenseMap<VarID, DILocation *> &InlinedAtsMap = InlinedAts.back()[FuncName];
+  // Find an Instruction that shares the same scope as the dropped #dbg_value or
+  // has a scope that is the child of the scope of the #dbg_value, and has an
+  // inlinedAt equal to the inlinedAt of the #dbg_value or it's inlinedAt chain
+  // contains the inlinedAt of the #dbg_value, if such an Instruction is found,
+  // debug information is dropped.
+  for (VarID Var : DebugVariablesBeforeSet) {
+    if (DebugVariablesAfterSet.contains(Var))
+      continue;
+    const DIScope *DbgValScope = std::get<0>(Var);
+    for (const auto &I : instructions(F)) {
+      auto *DbgLoc = I.getDebugLoc().get();
+      if (!DbgLoc)
+        continue;
+
+      auto *Scope = DbgLoc->getScope();
+      if (isScopeChildOfOrEqualTo(Scope, DbgValScope)) {
+        if (isInlinedAtChildOfOrEqualTo(DbgLoc->getInlinedAt(),
+                                        InlinedAtsMap[Var])) {
+          // Found another instruction in the variable's scope, so there exists
+          // a break point at which the variable could be observed. Count it as
+          // dropped.
+          DroppedCount++;
+          break;
+        }
+      }
+    }
+    removeVarFromAllSets(Var, F);
+  }
+  if (DroppedCount > 0) {
+    llvm::outs() << PassLevel << ", " << PassID << ", " << DroppedCount << ", "
+                 << FuncOrModName << "\n";
+    PassDroppedVariables = true;
+  } else
+    PassDroppedVariables = false;
+}
+
+void DroppedVariableStatsIR::runOnModule(const Module *M, bool Before) {
+  for (auto &F : *M)
+    runOnFunction(&F, Before);
+}
+
+void DroppedVariableStatsIR::calculateDroppedVarStatsOnModule(
+    const Module *M, StringRef PassID, std::string FuncOrModName,
+    std::string PassLevel) {
+  for (auto &F : *M) {
+    calculateDroppedVarStatsOnFunction(&F, PassID, FuncOrModName, PassLevel);
+  }
+}
+
+void DroppedVariableStatsIR::registerCallbacks(
+    PassInstrumentationCallbacks &PIC) {
+  if (!DroppedVariableStatsEnabled)
+    return;
+
+  PIC.registerBeforeNonSkippedPassCallback(
+      [this](StringRef P, Any IR) { return runBeforePass(IR); });
+  PIC.registerAfterPassCallback(
+      [this](StringRef P, Any IR, const PreservedAnalyses &PA) {
+        return runAfterPass(P, IR);
+      });
+  PIC.registerAfterPassInvalidatedCallback(
+      [this](StringRef P, const PreservedAnalyses &PA) { return cleanup(); });
+}
+
+void DroppedVariableStatsMIR::runOnMachineFunction(const MachineFunction *MF,
+                                                   boo...
[truncated]

``````````

</details>


https://github.com/llvm/llvm-project/pull/115566


More information about the llvm-commits mailing list