[MC] Adding code padding for performance stability - infrastructure. NFC.

Infrastructure designed for padding code with nop instructions in key places such that preformance improvement will be achieved.
The infrastructure is implemented such that the padding is done in the Assembler after the layout is done and all IPs and alignments are known.
This patch by itself in a NFC. Future patches will make use of this infrastructure to implement required policies for code padding.

Reviewers:
aaboud
zvi
craig.topper
gadi.haber

Differential revision: https://reviews.llvm.org/D34393

Change-Id: I92110d0c0a757080a8405636914a93ef6f8ad00e
llvm-svn: 316413

include/llvm/CodeGen/AsmPrinter.h

@@ -58,6 +58,7 @@ class MachineModuleInfo;
 class MachineOptimizationRemarkEmitter;
 class MCAsmInfo;
 class MCCFIInstruction;
+struct MCCodePaddingContext;
 class MCContext;
 class MCExpr;
 class MCInst;
@@ -377,7 +378,7 @@ public:
   virtual void EmitBasicBlockStart(const MachineBasicBlock &MBB) const;
 
   /// Targets can override this to emit stuff at the end of a basic block.
-  virtual void EmitBasicBlockEnd(const MachineBasicBlock &MBB) {}
+  virtual void EmitBasicBlockEnd(const MachineBasicBlock &MBB);
 
   /// Targets should implement this to emit instructions.
   virtual void EmitInstruction(const MachineInstr *) {
@@ -626,6 +627,8 @@ private:
   /// Emit GlobalAlias or GlobalIFunc.
   void emitGlobalIndirectSymbol(Module &M,
                                 const GlobalIndirectSymbol& GIS);
+  void setupCodePaddingContext(const MachineBasicBlock &MBB,
+                               MCCodePaddingContext &Context) const;
 };
 
 } // end namespace llvm

include/llvm/MC/MCAsmBackend.h

@@ -15,17 +15,22 @@
 #include "llvm/ADT/StringRef.h"
 #include "llvm/MC/MCDirectives.h"
 #include "llvm/MC/MCFixup.h"
+#include "llvm/MC/MCFragment.h"
 #include <cstdint>
+#include <memory>
 
 namespace llvm {
 
 class MCAsmLayout;
 class MCAssembler;
 class MCCFIInstruction;
+class MCCodePadder;
 struct MCFixupKindInfo;
 class MCFragment;
 class MCInst;
+class MCObjectStreamer;
 class MCObjectWriter;
+struct MCCodePaddingContext;
 class MCRelaxableFragment;
 class MCSubtargetInfo;
 class MCValue;
@@ -33,8 +38,11 @@ class raw_pwrite_stream;
 
 /// Generic interface to target specific assembler backends.
 class MCAsmBackend {
+  std::unique_ptr<MCCodePadder> CodePadder;
+
 protected: // Can only create subclasses.
   MCAsmBackend();
+  MCAsmBackend(std::unique_ptr<MCCodePadder> TargetCodePadder);
 
 public:
   MCAsmBackend(const MCAsmBackend &) = delete;
@@ -133,6 +141,40 @@ public:
       generateCompactUnwindEncoding(ArrayRef<MCCFIInstruction>) const {
     return 0;
   }
+
+  /// Handles all target related code padding when starting to write a new
+  /// basic block to an object file.
+  ///
+  /// \param OS The streamer used for writing the padding data and function.
+  /// \param Context the context of the padding, Embeds the basic block's
+  /// parameters.
+  void handleCodePaddingBasicBlockStart(MCObjectStreamer *OS,
+                                        const MCCodePaddingContext &Context);
+  /// Handles all target related code padding after writing a block to an object
+  /// file.
+  ///
+  /// \param Context the context of the padding, Embeds the basic block's
+  /// parameters.
+  void handleCodePaddingBasicBlockEnd(const MCCodePaddingContext &Context);
+  /// Handles all target related code padding before writing a new instruction
+  /// to an object file.
+  ///
+  /// \param Inst the instruction.
+  void handleCodePaddingInstructionBegin(const MCInst &Inst);
+  /// Handles all target related code padding after writing an instruction to an
+  /// object file.
+  ///
+  /// \param Inst the instruction.
+  void handleCodePaddingInstructionEnd(const MCInst &Inst);
+
+  /// Relaxes a fragment (changes the size of the padding) according to target
+  /// requirements. The new size computation is done w.r.t a layout.
+  ///
+  /// \param Fragment The fragment to relax.
+  /// \param Layout Code layout information.
+  ///
+  /// \returns true iff any relaxation occured.
+  bool relaxFragment(MCPaddingFragment *PF, MCAsmLayout &Layout);
 };
 
 } // end namespace llvm

include/llvm/MC/MCAssembler.h

@@ -183,6 +183,8 @@ private:
 
   bool relaxInstruction(MCAsmLayout &Layout, MCRelaxableFragment &IF);
 
+  bool relaxPaddingFragment(MCAsmLayout &Layout, MCPaddingFragment &PF);
+
   bool relaxLEB(MCAsmLayout &Layout, MCLEBFragment &IF);
 
   bool relaxDwarfLineAddr(MCAsmLayout &Layout, MCDwarfLineAddrFragment &DF);

include/llvm/MC/MCCodePadder.h

@@ -0,0 +1,243 @@
+//===- llvm/MC/CodePadder.h - MC Code Padder --------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_MC_MCCODEPADDER_H
+#define LLVM_MC_MCCODEPADDER_H
+
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SmallVector.h"
+
+namespace llvm {
+
+class MCAsmLayout;
+class MCCodePaddingPolicy;
+class MCFragment;
+class MCInst;
+class MCObjectStreamer;
+class MCPaddingFragment;
+class MCSection;
+
+typedef SmallVector<const MCPaddingFragment *, 8> MCPFRange;
+
+struct MCCodePaddingContext {
+  bool IsPaddingActive;
+  bool IsBasicBlockInsideInnermostLoop;
+  bool IsBasicBlockReachableViaFallthrough;
+  bool IsBasicBlockReachableViaBranch;
+};
+
+/// Target-independent base class incharge of all code padding decisions for a
+/// target. During encoding it determines if and where MCPaddingFragments will
+/// be located, as later on, when layout information is available, it determines
+/// their sizes.
+class MCCodePadder {
+  MCCodePadder(const MCCodePadder &) = delete;
+  void operator=(const MCCodePadder &) = delete;
+
+  /// Determines if the MCCodePaddingPolicies are active.
+  bool ArePoliciesActive;
+
+  /// All the supported MCCodePaddingPolicies.
+  SmallPtrSet<MCCodePaddingPolicy *, 4> CodePaddingPolicies;
+
+  /// A pointer to the fragment of the instruction whose padding is currently
+  /// done for.
+  MCPaddingFragment *CurrHandledInstFragment;
+
+  /// A map holding the jurisdiction for each padding fragment. Key: padding
+  /// fragment. Value: The fragment's jurisdiction. A jurisdiction is a vector
+  /// of padding fragments whose conditions are being controlled by another
+  /// fragment, the key fragment.
+  DenseMap<MCPaddingFragment *, MCPFRange> FragmentToJurisdiction;
+  MCPFRange &getJurisdiction(MCPaddingFragment *Fragment, MCAsmLayout &Layout);
+
+  /// A map holding the maximal instruction window size relevant for a padding
+  /// fragment.
+  DenseMap<MCPaddingFragment *, uint64_t> FragmentToMaxWindowSize;
+  uint64_t getMaxWindowSize(MCPaddingFragment *Fragment, MCAsmLayout &Layout);
+
+protected:
+  /// The current streamer, used to stream code padding.
+  MCObjectStreamer *OS;
+
+  bool addPolicy(MCCodePaddingPolicy *Policy);
+
+  virtual bool
+  basicBlockRequiresInsertionPoint(const MCCodePaddingContext &Context) {
+    return false;
+  }
+
+  virtual bool instructionRequiresInsertionPoint(const MCInst &Inst) {
+    return false;
+  }
+
+  virtual bool usePoliciesForBasicBlock(const MCCodePaddingContext &Context) {
+    return Context.IsPaddingActive;
+  }
+
+public:
+  MCCodePadder()
+      : ArePoliciesActive(false), CurrHandledInstFragment(nullptr),
+        OS(nullptr) {}
+  virtual ~MCCodePadder();
+
+  /// Handles all target related code padding when starting to write a new
+  /// basic block to an object file.
+  ///
+  /// \param OS The streamer used for writing the padding data and function.
+  /// \param Context the context of the padding, Embeds the basic block's
+  /// parameters.
+  void handleBasicBlockStart(MCObjectStreamer *OS,
+                             const MCCodePaddingContext &Context);
+  /// Handles all target related code padding when done writing a block to an
+  /// object file.
+  ///
+  /// \param Context the context of the padding, Embeds the basic block's
+  /// parameters.
+  void handleBasicBlockEnd(const MCCodePaddingContext &Context);
+  /// Handles all target related code padding before writing a new instruction
+  /// to an object file.
+  ///
+  /// \param Inst the instruction.
+  void handleInstructionBegin(const MCInst &Inst);
+  /// Handles all target related code padding after writing an instruction to an
+  /// object file.
+  ///
+  /// \param Inst the instruction.
+  void handleInstructionEnd(const MCInst &Inst);
+
+  /// Relaxes a fragment (changes the size of the padding) according to target
+  /// requirements. The new size computation is done w.r.t a layout.
+  ///
+  /// \param Fragment The fragment to relax.
+  /// \param Layout Code layout information.
+  ///
+  /// \returns true iff any relaxation occured.
+  bool relaxFragment(MCPaddingFragment *Fragment, MCAsmLayout &Layout);
+};
+
+/// The base class for all padding policies, i.e. a rule or set of rules to pad
+/// the generated code.
+class MCCodePaddingPolicy {
+  MCCodePaddingPolicy() = delete;
+  MCCodePaddingPolicy(const MCCodePaddingPolicy &) = delete;
+  void operator=(const MCCodePaddingPolicy &) = delete;
+
+protected:
+  /// A mask holding the kind of this policy, i.e. only the i'th bit will be set
+  /// where i is the kind number.
+  const uint64_t KindMask;
+  /// Instruction window size relevant to this policy.
+  const uint64_t WindowSize;
+  /// A boolean indicating which byte of the instruction determies its
+  /// instruction window. If true - the last byte of the instructions, o.w. -
+  /// the first byte of the instruction.
+  const bool InstByteIsLastByte;
+
+  MCCodePaddingPolicy(uint64_t Kind, uint64_t WindowSize,
+                      bool InstByteIsLastByte)
+      : KindMask(UINT64_C(1) << Kind), WindowSize(WindowSize),
+        InstByteIsLastByte(InstByteIsLastByte) {}
+
+  /// Computes and returns the offset of the consecutive fragment of a given
+  /// fragment.
+  ///
+  /// \param Fragment The fragment whose consecutive offset will be computed.
+  /// \param Layout Code layout information.
+  ///
+  /// \returns the offset of the consecutive fragment of \p Fragment.
+  static uint64_t getNextFragmentOffset(const MCFragment *Fragment,
+                                        const MCAsmLayout &Layout);
+  /// Returns the instruction byte of an instruction pointed by a given
+  /// MCPaddingFragment. An instruction byte is the address of the byte of an
+  /// instruction which determines its instruction window.
+  ///
+  /// \param Fragment The fragment pointing to the instruction.
+  /// \param Layout Code layout information.
+  ///
+  /// \returns the instruction byte of an instruction pointed by \p Fragment.
+  uint64_t getFragmentInstByte(const MCPaddingFragment *Fragment,
+                               MCAsmLayout &Layout) const;
+  uint64_t computeWindowEndAddress(const MCPaddingFragment *Fragment,
+                                   uint64_t Offset, MCAsmLayout &Layout) const;
+
+  /// Computes and returns the penalty weight of a first instruction window in a
+  /// range. This requires a special function since the first window does not
+  /// contain all the padding fragments in that window. It only contains all the
+  /// padding fragments starting from the relevant insertion point.
+  ///
+  /// \param Window The first window.
+  /// \param Offset The offset of the parent section relative to the beginning
+  /// of the file, mod the window size.
+  /// \param Layout Code layout information.
+  ///
+  /// \returns the penalty weight of a first instruction window in a range, \p
+  /// Window.
+  double computeFirstWindowPenaltyWeight(const MCPFRange &Window,
+                                         uint64_t Offset,
+                                         MCAsmLayout &Layout) const;
+  /// Computes and returns the penalty caused by an instruction window.
+  ///
+  /// \param Window The instruction window.
+  /// \param Offset The offset of the parent section relative to the beginning
+  /// of the file, mod the window size.
+  /// \param Layout Code layout information.
+  ///
+  /// \returns the penalty caused by \p Window.
+  virtual double computeWindowPenaltyWeight(const MCPFRange &Window,
+                                            uint64_t Offset,
+                                            MCAsmLayout &Layout) const = 0;
+
+public:
+  virtual ~MCCodePaddingPolicy() {}
+
+  /// Returns the kind mask of this policy -  A mask holding the kind of this
+  /// policy, i.e. only the i'th bit will be set where i is the kind number.
+  uint64_t getKindMask() const { return KindMask; }
+  /// Returns the instruction window size relevant to this policy.
+  uint64_t getWindowSize() const { return WindowSize; }
+  /// Returns true if the last byte of an instruction determines its instruction
+  /// window, or false if the first of an instruction determines it.
+  bool isInstByteLastByte() const { return InstByteIsLastByte; }
+
+  /// Returns true iff this policy needs padding for a given basic block.
+  ///
+  /// \param Context the context of the padding, Embeds the basic block's
+  /// parameters.
+  ///
+  /// \returns true iff this policy needs padding for the basic block.
+  virtual bool
+  basicBlockRequiresPaddingFragment(const MCCodePaddingContext &Context) const {
+    return false;
+  }
+  /// Returns true iff this policy needs padding for a given instruction.
+  ///
+  /// \param Inst The given instruction.
+  ///
+  /// \returns true iff this policy needs padding for \p Inst.
+  virtual bool instructionRequiresPaddingFragment(const MCInst &Inst) const {
+    return false;
+  }
+  /// Computes and returns the penalty caused by a range of instruction windows.
+  /// The weight is computed for each window separelty and then accumulated.
+  ///
+  /// \param Range The range.
+  /// \param Offset The offset of the parent section relative to the beginning
+  /// of the file, mod the window size.
+  /// \param Layout Code layout information.
+  ///
+  /// \returns the penalty caused by \p Range.
+  double computeRangePenaltyWeight(const MCPFRange &Range, uint64_t Offset,
+                                   MCAsmLayout &Layout) const;
+};
+
+} // namespace llvm
+
+#endif // LLVM_MC_MCCODEPADDER_H

include/llvm/MC/MCFragment.h

@@ -41,6 +41,7 @@ public:
     FT_Dwarf,
     FT_DwarfFrame,
     FT_LEB,
+    FT_Padding,
     FT_SafeSEH,
     FT_CVInlineLines,
     FT_CVDefRange,
@@ -323,6 +324,98 @@ public:
   }
 };
 
+/// Fragment for adding required padding.
+/// This fragment is always inserted before an instruction, and holds that
+/// instruction as context information (as well as a mask of kinds) for
+/// determining the padding size.
+///
+class MCPaddingFragment : public MCFragment {
+  /// A mask containing all the kinds relevant to this fragment. i.e. the i'th
+  /// bit will be set iff kind i is relevant to this fragment.
+  uint64_t PaddingPoliciesMask;
+  /// A boolean indicating if this fragment will actually hold padding. If its
+  /// value is false, then this fragment serves only as a placeholder,
+  /// containing data to assist other insertion point in their decision making.
+  bool IsInsertionPoint;
+
+  uint64_t Size;
+
+  struct MCInstInfo {
+    bool IsInitialized;
+    MCInst Inst;
+    /// A boolean indicating whether the instruction pointed by this fragment is
+    /// a fixed size instruction or a relaxable instruction held by a
+    /// MCRelaxableFragment.
+    bool IsImmutableSizedInst;
+    union {
+      /// If the instruction is a fixed size instruction, hold its size.
+      size_t InstSize;
+      /// Otherwise, hold a pointer to the MCRelaxableFragment holding it.
+      MCRelaxableFragment *InstFragment;
+    };
+  };
+  MCInstInfo InstInfo;
+
+public:
+  static const uint64_t PFK_None = UINT64_C(0);
+
+  enum MCPaddingFragmentKind {
+    // values 0-7 are reserved for future target independet values.
+
+    FirstTargetPerfNopFragmentKind = 8,
+
+    /// Limit range of target MCPerfNopFragment kinds to fit in uint64_t
+    MaxTargetPerfNopFragmentKind = 63
+  };
+
+  MCPaddingFragment(MCSection *Sec = nullptr)
+      : MCFragment(FT_Padding, false, 0, Sec), PaddingPoliciesMask(PFK_None),
+        IsInsertionPoint(false), Size(UINT64_C(0)),
+        InstInfo({false, MCInst(), false, {0}}) {}
+
+  bool isInsertionPoint() const { return IsInsertionPoint; }
+  void setAsInsertionPoint() { IsInsertionPoint = true; }
+  uint64_t getPaddingPoliciesMask() const { return PaddingPoliciesMask; }
+  void setPaddingPoliciesMask(uint64_t Value) { PaddingPoliciesMask = Value; }
+  bool hasPaddingPolicy(uint64_t PolicyMask) const {
+    assert(isPowerOf2_64(PolicyMask) &&
+           "Policy mask must contain exactly one policy");
+    return (getPaddingPoliciesMask() & PolicyMask) != PFK_None;
+  }
+  const MCInst &getInst() const {
+    assert(isInstructionInitialized() && "Fragment has no instruction!");
+    return InstInfo.Inst;
+  }
+  size_t getInstSize() const {
+    assert(isInstructionInitialized() && "Fragment has no instruction!");
+    if (InstInfo.IsImmutableSizedInst)
+      return InstInfo.InstSize;
+    assert(InstInfo.InstFragment != nullptr &&
+           "Must have a valid InstFragment to retrieve InstSize from");
+    return InstInfo.InstFragment->getContents().size();
+  }
+  void setInstAndInstSize(const MCInst &Inst, size_t InstSize) {
+	InstInfo.IsInitialized = true;
+    InstInfo.IsImmutableSizedInst = true;
+    InstInfo.Inst = Inst;
+    InstInfo.InstSize = InstSize;
+  }
+  void setInstAndInstFragment(const MCInst &Inst,
+                              MCRelaxableFragment *InstFragment) {
+    InstInfo.IsInitialized = true;
+    InstInfo.IsImmutableSizedInst = false;
+    InstInfo.Inst = Inst;
+    InstInfo.InstFragment = InstFragment;
+  }
+  uint64_t getSize() const { return Size; }
+  void setSize(uint64_t Value) { Size = Value; }
+  bool isInstructionInitialized() const { return InstInfo.IsInitialized; }
+
+  static bool classof(const MCFragment *F) {
+    return F->getKind() == MCFragment::FT_Padding;
+  }
+};
+
 class MCFillFragment : public MCFragment {
   /// Value to use for filling bytes.
   uint8_t Value;

include/llvm/MC/MCObjectStreamer.h

@@ -47,6 +47,7 @@ class MCObjectStreamer : public MCStreamer {
   void EmitCFIStartProcImpl(MCDwarfFrameInfo &Frame) override;
   void EmitCFIEndProcImpl(MCDwarfFrameInfo &Frame) override;
   MCSymbol *EmitCFILabel() override;
+  void EmitInstructionImpl(const MCInst &Inst, const MCSubtargetInfo &STI);
 
 protected:
   MCObjectStreamer(MCContext &Context, std::unique_ptr<MCAsmBackend> TAB,
@@ -76,6 +77,7 @@ public:
   /// Get a data fragment to write into, creating a new one if the current
   /// fragment is not a data fragment.
   MCDataFragment *getOrCreateDataFragment();
+  MCPaddingFragment *getOrCreatePaddingFragment();
 
 protected:
   bool changeSectionImpl(MCSection *Section, const MCExpr *Subsection);
@@ -121,6 +123,10 @@ public:
                          unsigned MaxBytesToEmit = 0) override;
   void emitValueToOffset(const MCExpr *Offset, unsigned char Value,
                          SMLoc Loc) override;
+  void
+  EmitCodePaddingBasicBlockStart(const MCCodePaddingContext &Context) override;
+  void
+  EmitCodePaddingBasicBlockEnd(const MCCodePaddingContext &Context) override;
   void EmitDwarfLocDirective(unsigned FileNo, unsigned Line,
                              unsigned Column, unsigned Flags,
                              unsigned Isa, unsigned Discriminator,

include/llvm/MC/MCStreamer.h

@@ -38,6 +38,7 @@ class AssemblerConstantPools;
 class formatted_raw_ostream;
 class MCAsmBackend;
 class MCCodeEmitter;
+struct MCCodePaddingContext;
 class MCContext;
 class MCExpr;
 class MCInst;
@@ -716,6 +717,12 @@ public:
   virtual void emitValueToOffset(const MCExpr *Offset, unsigned char Value,
                                  SMLoc Loc);
 
+  virtual void
+  EmitCodePaddingBasicBlockStart(const MCCodePaddingContext &Context) {}
+
+  virtual void
+  EmitCodePaddingBasicBlockEnd(const MCCodePaddingContext &Context) {}
+
   /// @}
 
   /// \brief Switch to a new logical file.  This is used to implement the '.file

lib/CodeGen/AsmPrinter/AsmPrinter.cpp

@@ -73,6 +73,7 @@
 #include "llvm/IR/Type.h"
 #include "llvm/IR/Value.h"
 #include "llvm/MC/MCAsmInfo.h"
+#include "llvm/MC/MCCodePadder.h"
 #include "llvm/MC/MCContext.h"
 #include "llvm/MC/MCDirectives.h"
 #include "llvm/MC/MCDwarf.h"
@@ -233,8 +234,7 @@ void AsmPrinter::getAnalysisUsage(AnalysisUsage &AU) const {
   AU.addRequired<MachineModuleInfo>();
   AU.addRequired<MachineOptimizationRemarkEmitterPass>();
   AU.addRequired<GCModuleInfo>();
-  if (isVerbose())
-    AU.addRequired<MachineLoopInfo>();
+  AU.addRequired<MachineLoopInfo>();
 }
 
 bool AsmPrinter::doInitialization(Module &M) {
@@ -1440,8 +1440,7 @@ void AsmPrinter::SetupMachineFunction(MachineFunction &MF) {
   }
 
   ORE = &getAnalysis<MachineOptimizationRemarkEmitterPass>().getORE();
-  if (isVerbose())
-    LI = &getAnalysis<MachineLoopInfo>();
+  LI = &getAnalysis<MachineLoopInfo>();
 
   const TargetSubtargetInfo &STI = MF.getSubtarget();
   EnablePrintSchedInfo = PrintSchedule.getNumOccurrences()
@@ -2614,6 +2613,23 @@ static void emitBasicBlockLoopComments(const MachineBasicBlock &MBB,
   PrintChildLoopComment(OS, Loop, AP.getFunctionNumber());
 }
 
+void AsmPrinter::setupCodePaddingContext(const MachineBasicBlock &MBB,
+                                         MCCodePaddingContext &Context) const {
+  assert(MF != nullptr && "Machine function must be valid");
+  assert(LI != nullptr && "Loop info must be valid");
+  Context.IsPaddingActive = !MF->hasInlineAsm() &&
+                            !MF->getFunction()->optForSize() &&
+                            TM.getOptLevel() != CodeGenOpt::None;
+  const MachineLoop *CurrentLoop = LI->getLoopFor(&MBB);
+  Context.IsBasicBlockInsideInnermostLoop =
+      CurrentLoop != nullptr && CurrentLoop->getSubLoops().empty();
+  Context.IsBasicBlockReachableViaFallthrough =
+      std::find(MBB.pred_begin(), MBB.pred_end(), MBB.getPrevNode()) !=
+      MBB.pred_end();
+  Context.IsBasicBlockReachableViaBranch =
+      MBB.pred_size() > 0 && !isBlockOnlyReachableByFallthrough(&MBB);
+}
+
 /// EmitBasicBlockStart - This method prints the label for the specified
 /// MachineBasicBlock, an alignment (if present) and a comment describing
 /// it if appropriate.
@@ -2629,6 +2645,9 @@ void AsmPrinter::EmitBasicBlockStart(const MachineBasicBlock &MBB) const {
   // Emit an alignment directive for this block, if needed.
   if (unsigned Align = MBB.getAlignment())
     EmitAlignment(Align);
+  MCCodePaddingContext Context;
+  setupCodePaddingContext(MBB, Context);
+  OutStreamer->EmitCodePaddingBasicBlockStart(Context);
 
   // If the block has its address taken, emit any labels that were used to
   // reference the block.  It is possible that there is more than one label
@@ -2670,6 +2689,12 @@ void AsmPrinter::EmitBasicBlockStart(const MachineBasicBlock &MBB) const {
   }
 }
 
+void AsmPrinter::EmitBasicBlockEnd(const MachineBasicBlock &MBB) {
+  MCCodePaddingContext Context;
+  setupCodePaddingContext(MBB, Context);
+  OutStreamer->EmitCodePaddingBasicBlockEnd(Context);
+}
+
 void AsmPrinter::EmitVisibility(MCSymbol *Sym, unsigned Visibility,
                                 bool IsDefinition) const {
   MCSymbolAttr Attr = MCSA_Invalid;

lib/MC/CMakeLists.txt

@@ -10,6 +10,7 @@ add_llvm_library(LLVMMC
   MCAsmStreamer.cpp
   MCAssembler.cpp
   MCCodeEmitter.cpp
+  MCCodePadder.cpp
   MCCodeView.cpp
   MCContext.cpp
   MCDwarf.cpp

lib/MC/MCAsmBackend.cpp

@@ -10,6 +10,7 @@
 #include "llvm/MC/MCAsmBackend.h"
 #include "llvm/ADT/None.h"
 #include "llvm/ADT/STLExtras.h"
+#include "llvm/MC/MCCodePadder.h"
 #include "llvm/MC/MCFixupKindInfo.h"
 #include <cassert>
 #include <cstddef>
@@ -17,7 +18,10 @@
 
 using namespace llvm;
 
-MCAsmBackend::MCAsmBackend() = default;
+MCAsmBackend::MCAsmBackend() : CodePadder(new MCCodePadder()) {}
+
+MCAsmBackend::MCAsmBackend(std::unique_ptr<MCCodePadder> TargetCodePadder)
+    : CodePadder(std::move(TargetCodePadder)) {}
 
 MCAsmBackend::~MCAsmBackend() = default;
 
@@ -59,3 +63,25 @@ bool MCAsmBackend::fixupNeedsRelaxationAdvanced(
     return true;
   return fixupNeedsRelaxation(Fixup, Value, DF, Layout);
 }
+
+void MCAsmBackend::handleCodePaddingBasicBlockStart(
+    MCObjectStreamer *OS, const MCCodePaddingContext &Context) {
+  CodePadder->handleBasicBlockStart(OS, Context);
+}
+
+void MCAsmBackend::handleCodePaddingBasicBlockEnd(
+    const MCCodePaddingContext &Context) {
+  CodePadder->handleBasicBlockEnd(Context);
+}
+
+void MCAsmBackend::handleCodePaddingInstructionBegin(const MCInst &Inst) {
+  CodePadder->handleInstructionBegin(Inst);
+}
+
+void MCAsmBackend::handleCodePaddingInstructionEnd(const MCInst &Inst) {
+  CodePadder->handleInstructionEnd(Inst);
+}
+
+bool MCAsmBackend::relaxFragment(MCPaddingFragment *PF, MCAsmLayout &Layout) {
+  return CodePadder->relaxFragment(PF, Layout);
+}

lib/MC/MCAssembler.cpp

@@ -68,6 +68,10 @@ STATISTIC(FragmentLayouts, "Number of fragment layouts");
 STATISTIC(ObjectBytes, "Number of emitted object file bytes");
 STATISTIC(RelaxationSteps, "Number of assembler layout and relaxation steps");
 STATISTIC(RelaxedInstructions, "Number of relaxed instructions");
+STATISTIC(PaddingFragmentsRelaxations,
+          "Number of Padding Fragments relaxations");
+STATISTIC(PaddingFragmentsBytes,
+          "Total size of all padding from adding Fragments");
 
 } // end namespace stats
 } // end anonymous namespace
@@ -283,6 +287,9 @@ uint64_t MCAssembler::computeFragmentSize(const MCAsmLayout &Layout,
   case MCFragment::FT_LEB:
     return cast<MCLEBFragment>(F).getContents().size();
 
+  case MCFragment::FT_Padding:
+    return cast<MCPaddingFragment>(F).getSize();
+
   case MCFragment::FT_SafeSEH:
     return 4;
 
@@ -549,6 +556,13 @@ static void writeFragment(const MCAssembler &Asm, const MCAsmLayout &Layout,
     break;
   }
 
+  case MCFragment::FT_Padding: {
+    if (!Asm.getBackend().writeNopData(FragmentSize, OW))
+      report_fatal_error("unable to write nop sequence of " +
+                         Twine(FragmentSize) + " bytes");
+    break;
+  }
+
   case MCFragment::FT_SafeSEH: {
     const MCSafeSEHFragment &SF = cast<MCSafeSEHFragment>(F);
     OW->write32(SF.getSymbol()->getIndex());
@@ -822,6 +836,19 @@ bool MCAssembler::relaxInstruction(MCAsmLayout &Layout,
   return true;
 }
 
+bool MCAssembler::relaxPaddingFragment(MCAsmLayout &Layout,
+                                       MCPaddingFragment &PF) {
+  uint64_t OldSize = PF.getSize();
+  if (!getBackend().relaxFragment(&PF, Layout))
+    return false;
+  uint64_t NewSize = PF.getSize();
+
+  ++stats::PaddingFragmentsRelaxations;
+  stats::PaddingFragmentsBytes += NewSize;
+  stats::PaddingFragmentsBytes -= OldSize;
+  return true;
+}
+
 bool MCAssembler::relaxLEB(MCAsmLayout &Layout, MCLEBFragment &LF) {
   uint64_t OldSize = LF.getContents().size();
   int64_t Value;
@@ -916,6 +943,9 @@ bool MCAssembler::layoutSectionOnce(MCAsmLayout &Layout, MCSection &Sec) {
     case MCFragment::FT_LEB:
       RelaxedFrag = relaxLEB(Layout, *cast<MCLEBFragment>(I));
       break;
+    case MCFragment::FT_Padding:
+      RelaxedFrag = relaxPaddingFragment(Layout, *cast<MCPaddingFragment>(I));
+      break;
     case MCFragment::FT_CVInlineLines:
       RelaxedFrag =
           relaxCVInlineLineTable(Layout, *cast<MCCVInlineLineTableFragment>(I));

lib/MC/MCCodePadder.cpp

@@ -0,0 +1,371 @@
+//===- MCCodePadder.cpp - Target MC Code Padder ---------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/MC/MCAsmLayout.h"
+#include "llvm/MC/MCCodePadder.h"
+#include "llvm/MC/MCObjectStreamer.h"
+#include <algorithm>
+#include <limits>
+#include <numeric>
+
+using namespace llvm;
+
+//---------------------------------------------------------------------------
+// MCCodePadder
+//
+
+MCCodePadder::~MCCodePadder() {
+  for (auto *Policy : CodePaddingPolicies)
+    delete Policy;
+}
+
+bool MCCodePadder::addPolicy(MCCodePaddingPolicy *Policy) {
+  assert(Policy && "Policy must be valid");
+  return CodePaddingPolicies.insert(Policy).second;
+}
+
+void MCCodePadder::handleBasicBlockStart(MCObjectStreamer *OS,
+                                         const MCCodePaddingContext &Context) {
+  assert(OS != nullptr && "OS must be valid");
+  assert(this->OS == nullptr && "Still handling another basic block");
+  this->OS = OS;
+
+  ArePoliciesActive = usePoliciesForBasicBlock(Context);
+
+  bool InsertionPoint = basicBlockRequiresInsertionPoint(Context);
+  assert((!InsertionPoint ||
+          OS->getCurrentFragment()->getKind() != MCFragment::FT_Align) &&
+         "Cannot insert padding nops right after an alignment fragment as it "
+         "will ruin the alignment");
+
+  uint64_t PoliciesMask = MCPaddingFragment::PFK_None;
+  if (ArePoliciesActive) {
+    PoliciesMask = std::accumulate(
+        CodePaddingPolicies.begin(), CodePaddingPolicies.end(),
+        MCPaddingFragment::PFK_None,
+        [&Context](uint64_t Mask,
+                   const MCCodePaddingPolicy *Policy) -> uint64_t {
+          return Policy->basicBlockRequiresPaddingFragment(Context)
+                     ? (Mask | Policy->getKindMask())
+                     : Mask;
+        });
+  }
+
+  if (InsertionPoint || PoliciesMask != MCPaddingFragment::PFK_None) {
+    MCPaddingFragment *PaddingFragment = OS->getOrCreatePaddingFragment();
+    if (InsertionPoint)
+      PaddingFragment->setAsInsertionPoint();
+    PaddingFragment->setPaddingPoliciesMask(
+        PaddingFragment->getPaddingPoliciesMask() | PoliciesMask);
+  }
+}
+
+void MCCodePadder::handleBasicBlockEnd(const MCCodePaddingContext &Context) {
+  assert(this->OS != nullptr && "Not handling a basic block");
+  OS = nullptr;
+}
+
+void MCCodePadder::handleInstructionBegin(const MCInst &Inst) {
+  if (!OS)
+    return; // instruction was emitted outside a function
+
+  assert(CurrHandledInstFragment == nullptr && "Can't start handling an "
+                                               "instruction while still "
+                                               "handling another instruction");
+
+  bool InsertionPoint = instructionRequiresInsertionPoint(Inst);
+  assert((!InsertionPoint ||
+          OS->getCurrentFragment()->getKind() != MCFragment::FT_Align) &&
+         "Cannot insert padding nops right after an alignment fragment as it "
+         "will ruin the alignment");
+
+  uint64_t PoliciesMask = MCPaddingFragment::PFK_None;
+  if (ArePoliciesActive) {
+    PoliciesMask = std::accumulate(
+        CodePaddingPolicies.begin(), CodePaddingPolicies.end(),
+        MCPaddingFragment::PFK_None,
+        [&Inst](uint64_t Mask, const MCCodePaddingPolicy *Policy) -> uint64_t {
+          return Policy->instructionRequiresPaddingFragment(Inst)
+                     ? (Mask | Policy->getKindMask())
+                     : Mask;
+        });
+  }
+  MCFragment *CurrFragment = OS->getCurrentFragment();
+  // CurrFragment can be a previously created MCPaddingFragment. If so, let's
+  // update it with the information we have, such as the instruction that it
+  // should point to.
+  bool needToUpdateCurrFragment =
+      CurrFragment != nullptr &&
+      CurrFragment->getKind() == MCFragment::FT_Padding;
+  if (InsertionPoint || PoliciesMask != MCPaddingFragment::PFK_None ||
+      needToUpdateCurrFragment) {
+    // temporarily holding the fragment as CurrHandledInstFragment, to be
+    // updated after the instruction will be written
+    CurrHandledInstFragment = OS->getOrCreatePaddingFragment();
+    if (InsertionPoint)
+      CurrHandledInstFragment->setAsInsertionPoint();
+    CurrHandledInstFragment->setPaddingPoliciesMask(
+        CurrHandledInstFragment->getPaddingPoliciesMask() | PoliciesMask);
+  }
+}
+
+void MCCodePadder::handleInstructionEnd(const MCInst &Inst) {
+  if (!OS)
+    return; // instruction was emitted outside a function
+  if (CurrHandledInstFragment == nullptr)
+    return;
+
+  MCFragment *InstFragment = OS->getCurrentFragment();
+  if (MCDataFragment *InstDataFragment =
+          dyn_cast_or_null<MCDataFragment>(InstFragment))
+    // Inst is a fixed size instruction and was encoded into a MCDataFragment.
+    // Let the fragment hold it and its size. Its size is the current size of
+    // the data fragment, as the padding fragment was inserted right before it
+    // and nothing was written yet except Inst
+    CurrHandledInstFragment->setInstAndInstSize(
+        Inst, InstDataFragment->getContents().size());
+  else if (MCRelaxableFragment *InstRelaxableFragment =
+               dyn_cast_or_null<MCRelaxableFragment>(InstFragment))
+    // Inst may be relaxed and its size may vary.
+    // Let the fragment hold the instruction and the MCRelaxableFragment
+    // that's holding it.
+    CurrHandledInstFragment->setInstAndInstFragment(Inst,
+                                                    InstRelaxableFragment);
+  else
+    llvm_unreachable("After encoding an instruction current fragment must be "
+                     "either a MCDataFragment or a MCRelaxableFragment");
+
+  CurrHandledInstFragment = nullptr;
+}
+
+MCPFRange &MCCodePadder::getJurisdiction(MCPaddingFragment *Fragment,
+                                         MCAsmLayout &Layout) {
+  auto JurisdictionLocation = FragmentToJurisdiction.find(Fragment);
+  if (JurisdictionLocation != FragmentToJurisdiction.end())
+    return JurisdictionLocation->second;
+
+  MCPFRange Jurisdiction;
+
+  // Forward scanning the fragments in this section, starting from the given
+  // fragments, and adding relevant MCPaddingFragments to the Jurisdiction
+  for (MCFragment *CurrFragment = Fragment; CurrFragment != nullptr;
+       CurrFragment = CurrFragment->getNextNode()) {
+
+    MCPaddingFragment *CurrPaddingFragment =
+        dyn_cast<MCPaddingFragment>(CurrFragment);
+    if (CurrPaddingFragment == nullptr)
+      continue;
+
+    if (CurrPaddingFragment != Fragment &&
+        CurrPaddingFragment->isInsertionPoint())
+      // Found next insertion point Fragment. From now on it's its jurisdiction.
+      break;
+    for (const auto *Policy : CodePaddingPolicies) {
+      if (CurrPaddingFragment->hasPaddingPolicy(Policy->getKindMask())) {
+        Jurisdiction.push_back(CurrPaddingFragment);
+        break;
+      }
+    }
+  }
+
+  auto InsertionResult =
+      FragmentToJurisdiction.insert(std::make_pair(Fragment, Jurisdiction));
+  assert(InsertionResult.second &&
+         "Insertion to FragmentToJurisdiction failed");
+  return InsertionResult.first->second;
+}
+
+uint64_t MCCodePadder::getMaxWindowSize(MCPaddingFragment *Fragment,
+                                        MCAsmLayout &Layout) {
+  auto MaxFragmentSizeLocation = FragmentToMaxWindowSize.find(Fragment);
+  if (MaxFragmentSizeLocation != FragmentToMaxWindowSize.end())
+    return MaxFragmentSizeLocation->second;
+
+  MCPFRange &Jurisdiction = getJurisdiction(Fragment, Layout);
+  uint64_t JurisdictionMask = MCPaddingFragment::PFK_None;
+  for (const auto *Protege : Jurisdiction)
+    JurisdictionMask |= Protege->getPaddingPoliciesMask();
+
+  uint64_t MaxFragmentSize = UINT64_C(0);
+  for (const auto *Policy : CodePaddingPolicies)
+    if ((JurisdictionMask & Policy->getKindMask()) !=
+        MCPaddingFragment::PFK_None)
+      MaxFragmentSize = std::max(MaxFragmentSize, Policy->getWindowSize());
+
+  auto InsertionResult =
+      FragmentToMaxWindowSize.insert(std::make_pair(Fragment, MaxFragmentSize));
+  assert(InsertionResult.second &&
+         "Insertion to FragmentToMaxWindowSize failed");
+  return InsertionResult.first->second;
+}
+
+bool MCCodePadder::relaxFragment(MCPaddingFragment *Fragment,
+                                 MCAsmLayout &Layout) {
+  if (!Fragment->isInsertionPoint())
+    return false;
+  uint64_t OldSize = Fragment->getSize();
+
+  uint64_t MaxWindowSize = getMaxWindowSize(Fragment, Layout);
+  if (MaxWindowSize == UINT64_C(0))
+    return false;
+  assert(isPowerOf2_64(MaxWindowSize) &&
+         "MaxWindowSize must be an integer power of 2");
+  uint64_t SectionAlignment = Fragment->getParent()->getAlignment();
+  assert(isPowerOf2_64(SectionAlignment) &&
+         "SectionAlignment must be an integer power of 2");
+
+  MCPFRange &Jurisdiction = getJurisdiction(Fragment, Layout);
+  uint64_t OptimalSize = UINT64_C(0);
+  double OptimalWeight = std::numeric_limits<double>::max();
+  uint64_t MaxFragmentSize = MaxWindowSize - UINT16_C(1);
+  for (uint64_t Size = UINT64_C(0); Size <= MaxFragmentSize; ++Size) {
+    Fragment->setSize(Size);
+    Layout.invalidateFragmentsFrom(Fragment);
+    double SizeWeight = 0.0;
+    // The section is guaranteed to be aligned to SectionAlignment, but that
+    // doesn't guarantee the exact section offset w.r.t. the policies window
+    // size.
+    // As a concrete example, the section could be aligned to 16B, but a
+    // policy's window size can be 32B. That means that the section actual start
+    // address can either be 0mod32 or 16mod32. The said policy will act
+    // differently for each case, so we need to take both into consideration.
+    for (uint64_t Offset = UINT64_C(0); Offset < MaxWindowSize;
+         Offset += SectionAlignment) {
+      double OffsetWeight = std::accumulate(
+          CodePaddingPolicies.begin(), CodePaddingPolicies.end(), 0.0,
+          [&Jurisdiction, &Offset, &Layout](
+              double Weight, const MCCodePaddingPolicy *Policy) -> double {
+            double PolicyWeight =
+                Policy->computeRangePenaltyWeight(Jurisdiction, Offset, Layout);
+            assert(PolicyWeight >= 0.0 && "A penalty weight must be positive");
+            return Weight + PolicyWeight;
+          });
+      SizeWeight = std::max(SizeWeight, OffsetWeight);
+    }
+    if (SizeWeight < OptimalWeight) {
+      OptimalWeight = SizeWeight;
+      OptimalSize = Size;
+    }
+    if (OptimalWeight == 0.0)
+      break;
+  }
+
+  Fragment->setSize(OptimalSize);
+  Layout.invalidateFragmentsFrom(Fragment);
+  return OldSize != OptimalSize;
+}
+
+//---------------------------------------------------------------------------
+// MCCodePaddingPolicy
+//
+
+uint64_t MCCodePaddingPolicy::getNextFragmentOffset(const MCFragment *Fragment,
+                                                    const MCAsmLayout &Layout) {
+  assert(Fragment != nullptr && "Fragment cannot be null");
+  MCFragment const *NextFragment = Fragment->getNextNode();
+  return NextFragment == nullptr
+             ? Layout.getSectionAddressSize(Fragment->getParent())
+             : Layout.getFragmentOffset(NextFragment);
+}
+
+uint64_t
+MCCodePaddingPolicy::getFragmentInstByte(const MCPaddingFragment *Fragment,
+                                         MCAsmLayout &Layout) const {
+  uint64_t InstByte = getNextFragmentOffset(Fragment, Layout);
+  if (InstByteIsLastByte)
+    InstByte += Fragment->getInstSize() - UINT64_C(1);
+  return InstByte;
+}
+
+uint64_t
+MCCodePaddingPolicy::computeWindowEndAddress(const MCPaddingFragment *Fragment,
+                                             uint64_t Offset,
+                                             MCAsmLayout &Layout) const {
+  uint64_t InstByte = getFragmentInstByte(Fragment, Layout);
+  return alignTo(InstByte + UINT64_C(1) + Offset, WindowSize) - Offset;
+}
+
+double MCCodePaddingPolicy::computeRangePenaltyWeight(
+    const MCPFRange &Range, uint64_t Offset, MCAsmLayout &Layout) const {
+
+  SmallVector<MCPFRange, 8> Windows;
+  SmallVector<MCPFRange, 8>::iterator CurrWindowLocation = Windows.end();
+  for (const MCPaddingFragment *Fragment : Range) {
+    if (!Fragment->hasPaddingPolicy(getKindMask()))
+      continue;
+    uint64_t FragmentWindowEndAddress =
+        computeWindowEndAddress(Fragment, Offset, Layout);
+    if (CurrWindowLocation == Windows.end() ||
+        FragmentWindowEndAddress !=
+            computeWindowEndAddress(*CurrWindowLocation->begin(), Offset,
+                                    Layout)) {
+      // next window is starting
+      Windows.push_back(MCPFRange());
+      CurrWindowLocation = Windows.end() - 1;
+    }
+    CurrWindowLocation->push_back(Fragment);
+  }
+
+  if (Windows.empty())
+    return 0.0;
+
+  double RangeWeight = 0.0;
+  SmallVector<MCPFRange, 8>::iterator I = Windows.begin();
+  RangeWeight += computeFirstWindowPenaltyWeight(*I, Offset, Layout);
+  ++I;
+  RangeWeight += std::accumulate(
+      I, Windows.end(), 0.0,
+      [this, &Layout, &Offset](double Weight, MCPFRange &Window) -> double {
+        return Weight += computeWindowPenaltyWeight(Window, Offset, Layout);
+      });
+  return RangeWeight;
+}
+
+double MCCodePaddingPolicy::computeFirstWindowPenaltyWeight(
+    const MCPFRange &Window, uint64_t Offset, MCAsmLayout &Layout) const {
+  if (Window.empty())
+    return 0.0;
+  uint64_t WindowEndAddress =
+      computeWindowEndAddress(*Window.begin(), Offset, Layout);
+
+  MCPFRange FullWindowFirstPart; // will hold all the fragments that are in the
+								 // same window as the fragments in the given
+								 // window but their penalty weight should not
+								 // be added
+  for (const MCFragment *Fragment = (*Window.begin())->getPrevNode();
+       Fragment != nullptr; Fragment = Fragment->getPrevNode()) {
+    const MCPaddingFragment *PaddingNopFragment =
+        dyn_cast<MCPaddingFragment>(Fragment);
+    if (PaddingNopFragment == nullptr ||
+        !PaddingNopFragment->hasPaddingPolicy(getKindMask()))
+      continue;
+    if (WindowEndAddress !=
+        computeWindowEndAddress(PaddingNopFragment, Offset, Layout))
+      break;
+
+    FullWindowFirstPart.push_back(PaddingNopFragment);
+  }
+
+  std::reverse(FullWindowFirstPart.begin(), FullWindowFirstPart.end());
+  double FullWindowFirstPartWeight =
+      computeWindowPenaltyWeight(FullWindowFirstPart, Offset, Layout);
+
+  MCPFRange FullWindow(
+      FullWindowFirstPart); // will hold all the fragments that are in the
+                            // same window as the fragments in the given
+                            // window, whether their weight should be added
+                            // or not
+  FullWindow.append(Window.begin(), Window.end());
+  double FullWindowWeight =
+      computeWindowPenaltyWeight(FullWindow, Offset, Layout);
+
+  assert(FullWindowWeight >= FullWindowFirstPartWeight &&
+         "More fragments necessarily means bigger weight");
+  return FullWindowWeight - FullWindowFirstPartWeight;
+}

lib/MC/MCFragment.cpp

@@ -278,6 +278,9 @@ void MCFragment::destroy() {
     case FT_LEB:
       delete cast<MCLEBFragment>(this);
       return;
+    case FT_Padding:
+      delete cast<MCPaddingFragment>(this);
+      return;
     case FT_SafeSEH:
       delete cast<MCSafeSEHFragment>(this);
       return;
@@ -322,6 +325,7 @@ LLVM_DUMP_METHOD void MCFragment::dump() const {
   case MCFragment::FT_Dwarf: OS << "MCDwarfFragment"; break;
   case MCFragment::FT_DwarfFrame: OS << "MCDwarfCallFrameFragment"; break;
   case MCFragment::FT_LEB:   OS << "MCLEBFragment"; break;
+  case MCFragment::FT_Padding: OS << "MCPaddingFragment"; break;
   case MCFragment::FT_SafeSEH:    OS << "MCSafeSEHFragment"; break;
   case MCFragment::FT_CVInlineLines: OS << "MCCVInlineLineTableFragment"; break;
   case MCFragment::FT_CVDefRange: OS << "MCCVDefRangeTableFragment"; break;
@@ -419,6 +423,19 @@ LLVM_DUMP_METHOD void MCFragment::dump() const {
     OS << " Value:" << LF->getValue() << " Signed:" << LF->isSigned();
     break;
   }
+  case MCFragment::FT_Padding: {
+    const MCPaddingFragment *F = cast<MCPaddingFragment>(this);
+    OS << "\n       ";
+    OS << " PaddingPoliciesMask:" << F->getPaddingPoliciesMask()
+       << " IsInsertionPoint:" << F->isInsertionPoint()
+       << " Size:" << F->getSize();
+    OS << "\n       ";
+    OS << " Inst:";
+    F->getInst().dump_pretty(OS);
+    OS << " InstSize:" << F->getInstSize();
+    OS << "\n       ";
+    break;
+  }
   case MCFragment::FT_SafeSEH: {
     const MCSafeSEHFragment *F = cast<MCSafeSEHFragment>(this);
     OS << "\n       ";

lib/MC/MCObjectStreamer.cpp

@@ -108,6 +108,16 @@ MCDataFragment *MCObjectStreamer::getOrCreateDataFragment() {
   return F;
 }
 
+MCPaddingFragment *MCObjectStreamer::getOrCreatePaddingFragment() {
+  MCPaddingFragment *F =
+      dyn_cast_or_null<MCPaddingFragment>(getCurrentFragment());
+  if (!F) {
+    F = new MCPaddingFragment();
+    insert(F);
+  }
+  return F;
+}
+
 void MCObjectStreamer::visitUsedSymbol(const MCSymbol &Sym) {
   Assembler->registerSymbol(Sym);
 }
@@ -247,6 +257,13 @@ bool MCObjectStreamer::mayHaveInstructions(MCSection &Sec) const {
 
 void MCObjectStreamer::EmitInstruction(const MCInst &Inst,
                                        const MCSubtargetInfo &STI, bool) {
+  getAssembler().getBackend().handleCodePaddingInstructionBegin(Inst);
+  EmitInstructionImpl(Inst, STI);
+  getAssembler().getBackend().handleCodePaddingInstructionEnd(Inst);
+}
+
+void MCObjectStreamer::EmitInstructionImpl(const MCInst &Inst,
+                                           const MCSubtargetInfo &STI) {
   MCStreamer::EmitInstruction(Inst, STI);
 
   MCSection *Sec = getCurrentSectionOnly();
@@ -467,6 +484,16 @@ void MCObjectStreamer::emitValueToOffset(const MCExpr *Offset,
   insert(new MCOrgFragment(*Offset, Value, Loc));
 }
 
+void MCObjectStreamer::EmitCodePaddingBasicBlockStart(
+    const MCCodePaddingContext &Context) {
+  getAssembler().getBackend().handleCodePaddingBasicBlockStart(this, Context);
+}
+
+void MCObjectStreamer::EmitCodePaddingBasicBlockEnd(
+    const MCCodePaddingContext &Context) {
+  getAssembler().getBackend().handleCodePaddingBasicBlockEnd(Context);
+}
+
 // Associate DTPRel32 fixup with data and resize data area
 void MCObjectStreamer::EmitDTPRel32Value(const MCExpr *Value) {
   MCDataFragment *DF = getOrCreateDataFragment();

lib/Target/Mips/MipsAsmPrinter.cpp

@@ -417,6 +417,7 @@ void MipsAsmPrinter::EmitFunctionBodyEnd() {
 }
 
 void MipsAsmPrinter::EmitBasicBlockEnd(const MachineBasicBlock &MBB) {
+  AsmPrinter::EmitBasicBlockEnd(MBB);
   MipsTargetStreamer &TS = getTargetStreamer();
   if (MBB.empty())
     TS.emitDirectiveInsn();

lib/Target/X86/X86AsmPrinter.h

@@ -119,6 +119,7 @@ public:
   void EmitInstruction(const MachineInstr *MI) override;
 
   void EmitBasicBlockEnd(const MachineBasicBlock &MBB) override {
+    AsmPrinter::EmitBasicBlockEnd(MBB);
     SMShadowTracker.emitShadowPadding(*OutStreamer, getSubtargetInfo());
   }
 

test/CodeGen/AArch64/arm64-opt-remarks-lazy-bfi.ll

@@ -36,7 +36,9 @@
 ; HOTNESS-NOT: Executing Pass
 ; HOTNESS: block-frequency: empty_func
 ; HOTNESS-NOT: Executing Pass
-; HOTNESS: Executing Pass 'AArch64 Assembly Printer'
+; HOTNESS: Executing Pass 'MachineDominator Tree Construction'
+; HOTNESS-NEXT: Executing Pass 'Machine Natural Loop Construction'
+; HOTNESS-NEXT: Executing Pass 'AArch64 Assembly Printer'
 
 ; HOTNESS: arm64-summary-remarks.ll:5:0: 1 instructions in function (hotness: 33)
 
@@ -45,6 +47,8 @@
 ; NO_HOTNESS-NEXT:  Freeing Pass 'Implement the 'patchable-function' attribute'
 ; NO_HOTNESS-NEXT: Executing Pass 'Lazy Machine Block Frequency Analysis'
 ; NO_HOTNESS-NEXT: Executing Pass 'Machine Optimization Remark Emitter'
+; NO_HOTNESS-NEXT: Executing Pass 'MachineDominator Tree Construction'
+; NO_HOTNESS-NEXT: Executing Pass 'Machine Natural Loop Construction'
 ; NO_HOTNESS-NEXT: Executing Pass 'AArch64 Assembly Printer'
 
 ; NO_HOTNESS: arm64-summary-remarks.ll:5:0: 1 instructions in function{{$}}