Add new pass LazyMachineBlockFrequencyInfo

[anemet]

Summary

And use it in MachineOptimizationRemarkEmitter. A test will follow on top of
Justin's changes to enable MachineORE in AsmPrinter.

The approach is similar to the IR-level pass. It's a bit simpler because BPI
is immutable at the Machine level so we don't need to make that lazy.

Because of this a new function mapping is introduced (calculateBPI) at the
IR-agnostic LazyBlockFrequencyInfo class. This function calculates and
extracts BPI from the pass. For Machine-level, this is the identity function

[davidxl]

include/llvm/Analysis/LazyBlockFrequencyInfo.h

37

It is quite confusing to read the code here regarding xxxInfo, xxxInfoPass and default parameters. It may be more readable to get rid of the last parameter, but instead do this:

class LazyBranchProbabilityInfoPass ... {
public:

   typedef BranchProbablityInfo BPIType;
...

};

class MachineBranchProbablityInfo ... {
public:

typedef MachineBranchProbabilityInfo BPIType;

..
};

and then

template <...>
class LazyBlockFrequencyInfo {

typedef typename BranchProbablityInfoPassT::BPIType BranchProbablityInfoT;
 ...

};

[anemet]

Redo the mapping from the BPI pass to the BPInfo

[anemet]

include/llvm/Analysis/LazyBlockFrequencyInfo.h

37

I agree that it's not the most readable but I wanted to keep the type info of the mapping and the actual mapping function together.

I've completely redone this part of the patch. Now we just have a trait class that provides the mapping function for each BPI pass. Let me know if this works for you.

[anemet]

No need to include <functional> anymore.

[davidxl]

lgtm

include/llvm/Analysis/LazyBlockFrequencyInfo.h

 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_ANALYSIS_LAZYBLOCKFREQUENCYINFO_H
 #define LLVM_ANALYSIS_LAZYBLOCKFREQUENCYINFO_H
 
 #include "llvm/Analysis/BlockFrequencyInfo.h"
 #include "llvm/Analysis/LazyBranchProbabilityInfo.h"
 #include "llvm/Pass.h"
+#include <functional>
 
 namespace llvm {
 class AnalysisUsage;
 class BranchProbabilityInfo;
 class Function;
 class LoopInfo;
 
 /// Wraps a BFI to allow lazy computation of the block frequencies.
 ///
 /// A pass that only conditionally uses BFI can uncondtionally require the
 /// analysis without paying for the overhead if BFI doesn't end up being used.
-template <typename FunctionT, typename LazyBranchProbabilityInfoPassT,
-          typename LoopInfoT, typename BlockFrequencyInfoT>
+template <typename FunctionT, typename LoopInfoT, typename BlockFrequencyInfoT,
+          typename LazyBranchProbabilityInfoPassT,
+          typename BranchProbabilityInfoT = LazyBranchProbabilityInfoPassT>
 class LazyBlockFrequencyInfo {
 public:
-  LazyBlockFrequencyInfo()
-      : Calculated(false), F(nullptr), BPIPass(nullptr), LI(nullptr) {}
+  typedef std::function<BranchProbabilityInfoT &(
+      LazyBranchProbabilityInfoPassT *)>
+      CalculateBPI;
+
+  LazyBlockFrequencyInfo(CalculateBPI calculateBPI)
+      : Calculated(false), F(nullptr), BPIPass(nullptr),
+        calculateBPI(calculateBPI), LI(nullptr) {}
 
   /// Set up the per-function input.
   void setAnalysis(const FunctionT *F, LazyBranchProbabilityInfoPassT *BPIPass,
                    const LoopInfoT *LI) {
     this->F = F;
     this->BPIPass = BPIPass;
     this->LI = LI;
   }
 
   /// Retrieve the BFI with the block frequencies computed.
   BlockFrequencyInfoT &getCalculated() {
     if (!Calculated) {
       assert(F && BPIPass && LI && "call setAnalysis");
-      BFI.calculate(*F, BPIPass->getBPI(), *LI);
+      BFI.calculate(*F, calculateBPI(BPIPass), *LI);
       Calculated = true;
     }
     return BFI;
   }
 
   const BlockFrequencyInfoT &getCalculated() const {
     return const_cast<LazyBlockFrequencyInfo *>(this)->getCalculated();
   }
 
   void releaseMemory() {
     BFI.releaseMemory();
     Calculated = false;
     setAnalysis(nullptr, nullptr, nullptr);
   }
 
 private:
   BlockFrequencyInfoT BFI;
   bool Calculated;
   const FunctionT *F;
   LazyBranchProbabilityInfoPassT *BPIPass;
+  CalculateBPI calculateBPI;
   const LoopInfoT *LI;
 };
 
 /// \brief This is an alternative analysis pass to
 /// BlockFrequencyInfoWrapperPass.  The difference is that with this pass the
 /// block frequencies are not computed when the analysis pass is executed but
 /// rather when the BFI results is explicitly requested by the analysis client.
 ///

 ///    getAnalysis<LazyBlockFrequencyInfoPass>().getBFI() before either LoopInfo
 ///    or BPI could be invalidated for example by changing the CFG.
 ///
 /// Note that it is expected that we wouldn't need this functionality for the
 /// new PM since with the new PM, analyses are executed on demand.
 
 class LazyBlockFrequencyInfoPass : public FunctionPass {
 private:
-  LazyBlockFrequencyInfo<Function, LazyBranchProbabilityInfoPass, LoopInfo,
-                         BlockFrequencyInfo>
+  LazyBlockFrequencyInfo<Function, LoopInfo, BlockFrequencyInfo,
+                         LazyBranchProbabilityInfoPass, BranchProbabilityInfo>
       LBFI;
 
 public:
   static char ID;
 
   LazyBlockFrequencyInfoPass();
 
   /// \brief Compute and return the block frequencies.

include/llvm/CodeGen/LazyMachineBlockFrequencyInfo.h

+///===- LazyMachineBlockFrequencyInfo.h - Lazy Block Frequency -*- C++ -*--===//
+///
+///                     The LLVM Compiler Infrastructure
+///
+/// This file is distributed under the University of Illinois Open Source
+/// License. See LICENSE.TXT for details.
+///
+///===---------------------------------------------------------------------===//
+/// \file
+/// This is an alternative analysis pass to MachineBlockFrequencyInfo.  The
+/// difference is that with this pass the block frequencies are not computed
+/// when the analysis pass is executed but rather when the BFI result is
+/// explicitly requested by the analysis client.
+///
+///===---------------------------------------------------------------------===//
+
+#ifndef LLVM_ANALYSIS_LAZYMACHINEBLOCKFREQUENCYINFO_H
+#define LLVM_ANALYSIS_LAZYMACHINEBLOCKFREQUENCYINFO_H
+
+#include "llvm/Analysis/LazyBlockFrequencyInfo.h"
+#include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
+#include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
+#include "llvm/CodeGen/MachineLoopInfo.h"
+
+namespace llvm {
+/// \brief This is an alternative analysis pass to MachineBlockFrequencyInfo.
+/// The difference is that with this pass, the block frequencies are not
+/// computed when the analysis pass is executed but rather when the BFI result
+/// is explicitly requested by the analysis client.
+///
+/// There are some additional requirements for any client pass that wants to use
+/// the analysis:
+///
+/// 1. The pass needs to initialize dependent passes with:
+///
+///   INITIALIZE_PASS_DEPENDENCY(LazyMachineBFIPass)
+///
+/// 2. Similarly, getAnalysisUsage should call:
+///
+///   LazyMachineBlockFrequencyInfoPass::getLazyMachineBFIAnalysisUsage(AU)
+///
+/// 3. The computed MachineBFI should be requested with
+///    getAnalysis<LazyMachineBlockFrequencyInfoPass>().getBFI() before
+///    MachineLoopInfo could be invalidated for example by changing the CFG.
+///
+/// Note that it is expected that we wouldn't need this functionality for the
+/// new PM since with the new PM, analyses are executed on demand.
+
+class LazyMachineBlockFrequencyInfoPass : public MachineFunctionPass {
+private:
+  /// \brief Machine BPI is an immutable pass, no need to use it lazily.
+  LazyBlockFrequencyInfo<MachineFunction, MachineLoopInfo,
+                         MachineBlockFrequencyInfo,
+                         MachineBranchProbabilityInfo>
+      LMBFI;
+
+public:
+  static char ID;
+
+  LazyMachineBlockFrequencyInfoPass();
+
+  /// \brief Compute and return the block frequencies.
+  MachineBlockFrequencyInfo &getBFI() { return LMBFI.getCalculated(); }
+
+  /// \brief Compute and return the block frequencies.
+  const MachineBlockFrequencyInfo &getBFI() const {
+    return LMBFI.getCalculated();
+  }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override;
+
+  /// Helper for client passes to set up the analysis usage on behalf of this
+  /// pass.
+  static void getLazyMachineBFIAnalysisUsage(AnalysisUsage &AU);
+
+  bool runOnMachineFunction(MachineFunction &F) override;
+  void releaseMemory() override;
+  void print(raw_ostream &OS, const Module *M) const override;
+};
+
+/// \brief Helper for client passes to initialize dependent passes for LMBFI.
+void initializeLazyMachineBFIPassPass(PassRegistry &Registry);
+}
+#endif

include/llvm/CodeGen/MachineBlockFrequencyInfo.h

 #include "llvm/CodeGen/MachineFunctionPass.h"
 #include "llvm/Support/BlockFrequency.h"
 #include <climits>
 
 namespace llvm {
 
 class MachineBasicBlock;
 class MachineBranchProbabilityInfo;
+class MachineLoopInfo;
 template <class BlockT> class BlockFrequencyInfoImpl;
 
 /// MachineBlockFrequencyInfo pass uses BlockFrequencyInfoImpl implementation
 /// to estimate machine basic block frequencies.
 class MachineBlockFrequencyInfo : public MachineFunctionPass {
   typedef BlockFrequencyInfoImpl<MachineBasicBlock> ImplType;
   std::unique_ptr<ImplType> MBFI;
 
 public:
   static char ID;
 
   MachineBlockFrequencyInfo();
 
   ~MachineBlockFrequencyInfo() override;
 
   void getAnalysisUsage(AnalysisUsage &AU) const override;
 
   bool runOnMachineFunction(MachineFunction &F) override;
 
+  /// calculate - compute block frequency info for the given function.
+  void calculate(const MachineFunction &F,
+                 const MachineBranchProbabilityInfo &MBPI,
+                 const MachineLoopInfo &MLI);
+
   void releaseMemory() override;
 
   /// getblockFreq - Return block frequency. Return 0 if we don't have the
   /// information. Please note that initial frequency is equal to 1024. It means
   /// that we should not rely on the value itself, but only on the comparison to
   /// the other block frequencies. We do this to avoid using of floating points.
   ///
   BlockFrequency getBlockFreq(const MachineBasicBlock *MBB) const;

include/llvm/InitializePasses.h

 void initializeInternalizeLegacyPassPass(PassRegistry&);
 void initializeIntervalPartitionPass(PassRegistry&);
 void initializeJumpThreadingPass(PassRegistry&);
 void initializeLCSSAWrapperPassPass(PassRegistry&);
 void initializeLCSSAVerificationPassPass(PassRegistry&);
 void initializeLegacyLICMPassPass(PassRegistry&);
 void initializeLegacyLoopSinkPassPass(PassRegistry&);
 void initializeLazyBranchProbabilityInfoPassPass(PassRegistry&);
+void initializeLazyMachineBlockFrequencyInfoPassPass(PassRegistry&);
 void initializeLazyBlockFrequencyInfoPassPass(PassRegistry&);
 void initializeLazyValueInfoWrapperPassPass(PassRegistry&);
 void initializeLegalizerPass(PassRegistry&);
 void initializeLibCallsShrinkWrapLegacyPassPass(PassRegistry&);
 void initializeLintPass(PassRegistry&);
 void initializeLiveDebugValuesPass(PassRegistry&);
 void initializeLiveDebugVariablesPass(PassRegistry&);
 void initializeLiveIntervalsPass(PassRegistry&);

lib/Analysis/LazyBlockFrequencyInfo.cpp

INITIALIZE_PASS_BEGIN(LazyBlockFrequencyInfoPass, DEBUG_TYPE,
                       "Lazy Block Frequency Analysis", true, true)
 INITIALIZE_PASS_DEPENDENCY(LazyBPIPass)
 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
 INITIALIZE_PASS_END(LazyBlockFrequencyInfoPass, DEBUG_TYPE,
                     "Lazy Block Frequency Analysis", true, true)
 
 char LazyBlockFrequencyInfoPass::ID = 0;
 
-LazyBlockFrequencyInfoPass::LazyBlockFrequencyInfoPass() : FunctionPass(ID) {
+LazyBlockFrequencyInfoPass::LazyBlockFrequencyInfoPass()
+    : FunctionPass(ID),
+      LBFI([](LazyBranchProbabilityInfoPass *P) -> BranchProbabilityInfo & {
+        return P->getBPI();
+      }) {
   initializeLazyBlockFrequencyInfoPassPass(*PassRegistry::getPassRegistry());
 }
 
 void LazyBlockFrequencyInfoPass::print(raw_ostream &OS, const Module *) const {
   LBFI.getCalculated().print(OS);
 }
 
 void LazyBlockFrequencyInfoPass::getAnalysisUsage(AnalysisUsage &AU) const {

lib/CodeGen/CMakeLists.txt

add_llvm_library(LLVMCodeGen
   GlobalMerge.cpp
   IfConversion.cpp
   ImplicitNullChecks.cpp
   InlineSpiller.cpp
   InterferenceCache.cpp
   InterleavedAccessPass.cpp
   IntrinsicLowering.cpp
   LatencyPriorityQueue.cpp
+  LazyMachineBlockFrequencyInfo.cpp
   LexicalScopes.cpp
   LiveDebugValues.cpp
   LiveDebugVariables.cpp
   LiveIntervalAnalysis.cpp
   LiveInterval.cpp
   LiveIntervalUnion.cpp
   LivePhysRegs.cpp
   LiveRangeCalc.cpp

lib/CodeGen/LazyMachineBlockFrequencyInfo.cpp

+///===- LazyMachineBlockFrequencyInfo.cpp - Lazy Machine Block Frequency --===//
+///
+///                     The LLVM Compiler Infrastructure
+///
+/// This file is distributed under the University of Illinois Open Source
+/// License. See LICENSE.TXT for details.
+///
+///===---------------------------------------------------------------------===//
+/// \file
+/// This is an alternative analysis pass to MachineBlockFrequencyInfo.  The
+/// difference is that with this pass the block frequencies are not computed
+/// when the analysis pass is executed but rather when the BFI result is
+/// explicitly requested by the analysis client.
+///
+///===---------------------------------------------------------------------===//
+
+#include "llvm/CodeGen/LazyMachineBlockFrequencyInfo.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "lazy-machine-block-freq"
+
+INITIALIZE_PASS_BEGIN(LazyMachineBlockFrequencyInfoPass, DEBUG_TYPE,
+                      "Lazy Machine Block Frequency Analysis", true, true)
+INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
+INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
+INITIALIZE_PASS_END(LazyMachineBlockFrequencyInfoPass, DEBUG_TYPE,
+                    "Lazy Machine Block Frequency Analysis", true, true)
+
+char LazyMachineBlockFrequencyInfoPass::ID = 0;
+
+LazyMachineBlockFrequencyInfoPass::LazyMachineBlockFrequencyInfoPass()
+    : MachineFunctionPass(ID),
+      /// Machine BPI is an immutable pass, so this is effectively just the
+      /// identity function.
+      LMBFI([](MachineBranchProbabilityInfo *I)
+                -> MachineBranchProbabilityInfo & { return *I; }) {
+  initializeLazyMachineBlockFrequencyInfoPassPass(
+      *PassRegistry::getPassRegistry());
+}
+
+void LazyMachineBlockFrequencyInfoPass::print(raw_ostream &OS,
+                                              const Module *M) const {
+  LMBFI.getCalculated().print(OS, M);
+}
+
+void LazyMachineBlockFrequencyInfoPass::getAnalysisUsage(
+    AnalysisUsage &AU) const {
+  AU.addRequired<MachineBranchProbabilityInfo>();
+  AU.addRequired<MachineLoopInfo>();
+  AU.setPreservesAll();
+  MachineFunctionPass::getAnalysisUsage(AU);
+}
+
+void LazyMachineBlockFrequencyInfoPass::releaseMemory() {
+  LMBFI.releaseMemory();
+}
+
+bool LazyMachineBlockFrequencyInfoPass::runOnMachineFunction(
+    MachineFunction &MF) {
+  auto &BPIPass = getAnalysis<MachineBranchProbabilityInfo>();
+  auto &LI = getAnalysis<MachineLoopInfo>();
+  LMBFI.setAnalysis(&MF, &BPIPass, &LI);
+  return false;
+}
+
+void LazyMachineBlockFrequencyInfoPass::getLazyMachineBFIAnalysisUsage(
+    AnalysisUsage &AU) {
+  AU.addRequired<LazyMachineBlockFrequencyInfoPass>();
+  AU.addRequired<MachineBranchProbabilityInfo>();
+  AU.addRequired<LoopInfoWrapperPass>();
+}
+
+void llvm::initializeLazyMachineBFIPassPass(PassRegistry &Registry) {
+  INITIALIZE_PASS_DEPENDENCY(LazyMachineBlockFrequencyInfoPass);
+  INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo);
+  INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo);
+}

lib/CodeGen/MachineBlockFrequencyInfo.cpp

 
 void MachineBlockFrequencyInfo::getAnalysisUsage(AnalysisUsage &AU) const {
   AU.addRequired<MachineBranchProbabilityInfo>();
   AU.addRequired<MachineLoopInfo>();
   AU.setPreservesAll();
   MachineFunctionPass::getAnalysisUsage(AU);
 }
 
-bool MachineBlockFrequencyInfo::runOnMachineFunction(MachineFunction &F) {
-  MachineBranchProbabilityInfo &MBPI =
-      getAnalysis<MachineBranchProbabilityInfo>();
-  MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>();
+void MachineBlockFrequencyInfo::calculate(
+    const MachineFunction &F, const MachineBranchProbabilityInfo &MBPI,
+    const MachineLoopInfo &MLI) {
   if (!MBFI)
     MBFI.reset(new ImplType);
   MBFI->calculate(F, MBPI, MLI);
 #ifndef NDEBUG
   if (ViewMachineBlockFreqPropagationDAG != GVDT_None &&
       (ViewBlockFreqFuncName.empty() ||
        F.getName().equals(ViewBlockFreqFuncName))) {
     view();
   }
 #endif
+}
+
+bool MachineBlockFrequencyInfo::runOnMachineFunction(MachineFunction &F) {
+  MachineBranchProbabilityInfo &MBPI =
+      getAnalysis<MachineBranchProbabilityInfo>();
+  MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>();
+  calculate(F, MBPI, MLI);
   return false;
 }
 
 void MachineBlockFrequencyInfo::releaseMemory() { MBFI.reset(); }
 
 /// Pop up a ghostview window with the current block frequency propagation
 /// rendered using dot.
 void MachineBlockFrequencyInfo::view(bool isSimple) const {

lib/CodeGen/MachineOptimizationRemarkEmitter.cpp

 ///===- MachineOptimizationRemarkEmitter.cpp - Opt Diagnostic -*- C++ -*---===//
 ///
 ///                     The LLVM Compiler Infrastructure
 ///
 /// This file is distributed under the University of Illinois Open Source
 /// License. See LICENSE.TXT for details.
 ///
 ///===---------------------------------------------------------------------===//
 /// \file
 /// Optimization diagnostic interfaces for machine passes.  It's packaged as an
 /// analysis pass so that by using this service passes become dependent on MBFI
 /// as well.  MBFI is used to compute the "hotness" of the diagnostic message.
 ///
 ///===---------------------------------------------------------------------===//
 
 #include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h"
-#include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
+#include "llvm/CodeGen/LazyMachineBlockFrequencyInfo.h"
 #include "llvm/IR/DebugInfo.h"
 #include "llvm/IR/DiagnosticInfo.h"
 #include "llvm/IR/LLVMContext.h"
 
 using namespace llvm;
 
 Optional<uint64_t>
 MachineOptimizationRemarkEmitter::computeHotness(const MachineBasicBlock &MBB) {
initializeMachineOptimizationRemarkEmitterPassPass(
       *PassRegistry::getPassRegistry());
 }
 
 bool MachineOptimizationRemarkEmitterPass::runOnMachineFunction(
     MachineFunction &MF) {
   MachineBlockFrequencyInfo *MBFI;
 
   if (MF.getFunction()->getContext().getDiagnosticHotnessRequested())
-    MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
+    MBFI = &getAnalysis<LazyMachineBlockFrequencyInfoPass>().getBFI();
   else
     MBFI = nullptr;
 
   ORE = llvm::make_unique<MachineOptimizationRemarkEmitter>(MF, MBFI);
   return false;
 }
 
 void MachineOptimizationRemarkEmitterPass::getAnalysisUsage(
     AnalysisUsage &AU) const {
-  AU.addRequired<MachineBlockFrequencyInfo>();
+  LazyMachineBlockFrequencyInfoPass::getLazyMachineBFIAnalysisUsage(AU);
   AU.setPreservesAll();
   MachineFunctionPass::getAnalysisUsage(AU);
 }
 
 char MachineOptimizationRemarkEmitterPass::ID = 0;
 static const char ore_name[] = "Machine Optimization Remark Emitter";
 #define ORE_NAME "machine-opt-remark-emitter"
 
 INITIALIZE_PASS_BEGIN(MachineOptimizationRemarkEmitterPass, ORE_NAME, ore_name,
                       false, true)
-INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
+INITIALIZE_PASS_DEPENDENCY(LazyMachineBFIPass)
 INITIALIZE_PASS_END(MachineOptimizationRemarkEmitterPass, ORE_NAME, ore_name,
                     false, true)