MC: Store MCRelaxableFragment MCInst out-of-line

Follow-up to llvm#146307

Moved MCInst storage to MCSection, enabling trivial ~MCRelaxableFragment
and eliminating the need for a fragment walk in ~MCSection.

Updated MCRelaxableFragment::getInst to construct an MCInst on demand.
Modified MCAssembler::relaxInstruction's mayNeedRelaxation to accept
opcode and operands instead of an MCInst, avoiding redundant MCInst
creation. Note that MCObjectStreamer::emitInstructionImpl calls
mayNeedRelaxation before determining the target fragment for the MCInst.

Unfortunately, we also have to encode `MCInst::Flags` to support
the EVEX prefix, e.g. `{evex} xorw $foo, %ax`

There is a small decrease in max-rss (stage1-ReleaseLTO-g (link only))
with negligible instructions:u change.
https://llvm-compile-time-tracker.com/compare.php?from=408e87184f8274e30d188a2fe198facf55243733&to=5e0997dca139e08c6ace0d404a311e1f1b2d3b3f&stat=max-rss&linkStats=on

Next: Enable MCFragment to store fixed-size data (was MCDataFragment's job)
and optional Opcode/Operands data (was MCRelaxableFragment's job),
and delete MCDataFragment/MCRelaxableFragment.
This will allow re-encoding of Data+Relax+Data+Relax sequences as
Frag+Frag. The saving should outweigh the downside of larger
MCFragment.

Pull Request: llvm#147229

Author: MaskRay

llvm/include/llvm/MC/MCAsmBackend.h

@@ -32,6 +32,7 @@ class MCInst;
 class MCObjectStreamer;
 class MCObjectTargetWriter;
 class MCObjectWriter;
+class MCOperand;
 class MCSubtargetInfo;
 class MCValue;
 class raw_pwrite_stream;
@@ -147,12 +148,9 @@ class LLVM_ABI MCAsmBackend {
   /// \name Target Relaxation Interfaces
   /// @{
 
-  /// Check whether the given instruction may need relaxation.
-  ///
-  /// \param Inst - The instruction to test.
-  /// \param STI - The MCSubtargetInfo in effect when the instruction was
-  /// encoded.
-  virtual bool mayNeedRelaxation(const MCInst &Inst,
+  /// Check whether the given instruction (encoded as Opcode+Operands) may need
+  /// relaxation.
+  virtual bool mayNeedRelaxation(unsigned Opcode, ArrayRef<MCOperand> Operands,
                                  const MCSubtargetInfo &STI) const {
     return false;
   }

llvm/include/llvm/MC/MCInst.h

@@ -15,6 +15,7 @@
 #ifndef LLVM_MC_MCINST_H
 #define LLVM_MC_MCINST_H
 
+#include "llvm/ADT/ArrayRef.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/ADT/bit.h"
@@ -210,7 +211,11 @@ class MCInst {
   MCOperand &getOperand(unsigned i) { return Operands[i]; }
   unsigned getNumOperands() const { return Operands.size(); }
 
+  ArrayRef<MCOperand> getOperands() const { return Operands; }
   void addOperand(const MCOperand Op) { Operands.push_back(Op); }
+  void setOperands(ArrayRef<MCOperand> Ops) {
+    Operands.assign(Ops.begin(), Ops.end());
+  }
 
   using iterator = SmallVectorImpl<MCOperand>::iterator;
   using const_iterator = SmallVectorImpl<MCOperand>::const_iterator;

llvm/include/llvm/MC/MCSection.h

@@ -39,8 +39,8 @@ class raw_ostream;
 class Triple;
 
 // Represents a contiguous piece of code or data within a section. Its size is
-// determined by MCAssembler::layout. All subclasses (except
-// MCRelaxableFragment, which stores a MCInst) must have trivial destructors.
+// determined by MCAssembler::layout. All subclasses must have trivial
+// destructors.
 //
 // Declaration order: MCFragment, MCSection, then MCFragment's derived classes.
 // This allows MCSection's inline functions to access MCFragment members and
@@ -101,12 +101,6 @@ class MCFragment {
   MCFragment(const MCFragment &) = delete;
   MCFragment &operator=(const MCFragment &) = delete;
 
-  /// Destroys the current fragment.
-  ///
-  /// This must be used instead of delete as MCFragment is non-virtual.
-  /// This method will dispatch to the appropriate subclass.
-  LLVM_ABI void destroy();
-
   MCFragment *getNext() const { return Next; }
 
   FragmentType getKind() const { return Kind; }
@@ -133,6 +127,7 @@ class LLVM_ABI MCSection {
   friend MCAssembler;
   friend MCObjectStreamer;
   friend class MCEncodedFragment;
+  friend class MCRelaxableFragment;
   static constexpr unsigned NonUniqueID = ~0U;
 
   enum SectionVariant {
@@ -213,6 +208,7 @@ class LLVM_ABI MCSection {
   // Content and fixup storage for fragments
   SmallVector<char, 0> ContentStorage;
   SmallVector<MCFixup, 0> FixupStorage;
+  SmallVector<MCOperand, 0> MCOperandStorage;
 
 protected:
   // TODO Make Name private when possible.
@@ -221,7 +217,8 @@ class LLVM_ABI MCSection {
 
   MCSection(SectionVariant V, StringRef Name, bool IsText, bool IsVirtual,
             MCSymbol *Begin);
-  ~MCSection();
+  // Protected non-virtual dtor prevents destroy through a base class pointer.
+  ~MCSection() {}
 
 public:
   MCSection(const MCSection &) = delete;
@@ -430,17 +427,41 @@ class MCDataFragment : public MCEncodedFragment {
 /// relaxed during the assembler layout and relaxation stage.
 ///
 class MCRelaxableFragment : public MCEncodedFragment {
-  /// The instruction this is a fragment for.
-  MCInst Inst;
+  uint32_t Opcode = 0;
+  uint32_t Flags = 0;
+  uint32_t OperandStart = 0;
+  uint32_t OperandSize = 0;
 
 public:
-  MCRelaxableFragment(const MCInst &Inst, const MCSubtargetInfo &STI)
-      : MCEncodedFragment(FT_Relaxable, true), Inst(Inst) {
+  MCRelaxableFragment(const MCSubtargetInfo &STI)
+      : MCEncodedFragment(FT_Relaxable, true) {
     this->STI = &STI;
   }
 
-  const MCInst &getInst() const { return Inst; }
-  void setInst(const MCInst &Value) { Inst = Value; }
+  unsigned getOpcode() const { return Opcode; }
+  ArrayRef<MCOperand> getOperands() const {
+    return MutableArrayRef(getParent()->MCOperandStorage)
+        .slice(OperandStart, OperandSize);
+  }
+  MCInst getInst() const {
+    MCInst Inst;
+    Inst.setOpcode(Opcode);
+    Inst.setFlags(Flags);
+    Inst.setOperands(ArrayRef(getParent()->MCOperandStorage)
+                         .slice(OperandStart, OperandSize));
+    return Inst;
+  }
+  void setInst(const MCInst &Inst) {
+    Opcode = Inst.getOpcode();
+    Flags = Inst.getFlags();
+    auto &S = getParent()->MCOperandStorage;
+    if (Inst.getNumOperands() > OperandSize) {
+      OperandStart = S.size();
+      S.resize_for_overwrite(S.size() + Inst.getNumOperands());
+    }
+    OperandSize = Inst.getNumOperands();
+    llvm::copy(Inst, S.begin() + OperandStart);
+  }
 
   bool getAllowAutoPadding() const { return AllowAutoPadding; }
   void setAllowAutoPadding(bool V) { AllowAutoPadding = V; }

llvm/lib/MC/MCAssembler.cpp

@@ -869,7 +869,8 @@ bool MCAssembler::relaxInstruction(MCRelaxableFragment &F) {
   // If this inst doesn't ever need relaxation, ignore it. This occurs when we
   // are intentionally pushing out inst fragments, or because we relaxed a
   // previous instruction to one that doesn't need relaxation.
-  if (!getBackend().mayNeedRelaxation(F.getInst(), *F.getSubtargetInfo()))
+  if (!getBackend().mayNeedRelaxation(F.getOpcode(), F.getOperands(),
+                                      *F.getSubtargetInfo()))
     return false;
 
   bool DoRelax = false;
@@ -881,6 +882,8 @@ bool MCAssembler::relaxInstruction(MCRelaxableFragment &F) {
 
   ++stats::RelaxedInstructions;
 
+  // TODO Refactor relaxInstruction to accept MCRelaxableFragment and remove
+  // `setInst`.
   MCInst Relaxed = F.getInst();
   getBackend().relaxInstruction(Relaxed, *F.getSubtargetInfo());
 

llvm/lib/MC/MCObjectStreamer.cpp

@@ -352,7 +352,7 @@ void MCObjectStreamer::emitInstructionImpl(const MCInst &Inst,
   // If this instruction doesn't need relaxation, just emit it as data.
   MCAssembler &Assembler = getAssembler();
   MCAsmBackend &Backend = Assembler.getBackend();
-  if (!(Backend.mayNeedRelaxation(Inst, STI) ||
+  if (!(Backend.mayNeedRelaxation(Inst.getOpcode(), Inst.getOperands(), STI) ||
         Backend.allowEnhancedRelaxation())) {
     emitInstToData(Inst, STI);
     return;
@@ -366,7 +366,8 @@ void MCObjectStreamer::emitInstructionImpl(const MCInst &Inst,
   if (Assembler.getRelaxAll() ||
       (Assembler.isBundlingEnabled() && Sec->isBundleLocked())) {
     MCInst Relaxed = Inst;
-    while (Backend.mayNeedRelaxation(Relaxed, STI))
+    while (Backend.mayNeedRelaxation(Relaxed.getOpcode(), Relaxed.getOperands(),
+                                     STI))
       Backend.relaxInstruction(Relaxed, STI);
     emitInstToData(Relaxed, STI);
     return;
@@ -397,8 +398,9 @@ void MCObjectStreamer::emitInstToFragment(const MCInst &Inst,
   // Always create a new, separate fragment here, because its size can change
   // during relaxation.
   MCRelaxableFragment *IF =
-      getContext().allocFragment<MCRelaxableFragment>(Inst, STI);
+      getContext().allocFragment<MCRelaxableFragment>(STI);
   insert(IF);
+  IF->setInst(Inst);
 
   SmallVector<MCFixup, 1> Fixups;
   getAssembler().getEmitter().encodeInstruction(

llvm/lib/MC/MCSection.cpp

@@ -36,18 +36,6 @@ MCSymbol *MCSection::getEndSymbol(MCContext &Ctx) {
 
 bool MCSection::hasEnded() const { return End && End->isInSection(); }
 
-MCSection::~MCSection() {
-  // If ~MCRelaxableFragment becomes trivial (no longer store a MCInst member),
-  // this dtor can be made empty.
-  for (auto &[_, Chain] : Subsections) {
-    for (MCFragment *X = Chain.Head, *Y; X; X = Y) {
-      Y = X->Next;
-      if (auto *F = dyn_cast<MCRelaxableFragment>(X))
-        F->~MCRelaxableFragment();
-    }
-  }
-}
-
 void MCSection::setBundleLockState(BundleLockStateType NewState) {
   if (NewState == NotBundleLocked) {
     if (BundleLockNestingDepth == 0) {

llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUAsmBackend.cpp

@@ -40,7 +40,7 @@ class AMDGPUAsmBackend : public MCAsmBackend {
   void relaxInstruction(MCInst &Inst,
                         const MCSubtargetInfo &STI) const override;
 
-  bool mayNeedRelaxation(const MCInst &Inst,
+  bool mayNeedRelaxation(unsigned Opcode, ArrayRef<MCOperand> Operands,
                          const MCSubtargetInfo &STI) const override;
 
   unsigned getMinimumNopSize() const override;
@@ -70,12 +70,13 @@ bool AMDGPUAsmBackend::fixupNeedsRelaxation(const MCFixup &Fixup,
   return (((int64_t(Value)/4)-1) == 0x3f);
 }
 
-bool AMDGPUAsmBackend::mayNeedRelaxation(const MCInst &Inst,
-                       const MCSubtargetInfo &STI) const {
+bool AMDGPUAsmBackend::mayNeedRelaxation(unsigned Opcode,
+                                         ArrayRef<MCOperand> Operands,
+                                         const MCSubtargetInfo &STI) const {
   if (!STI.hasFeature(AMDGPU::FeatureOffset3fBug))
     return false;
 
-  if (AMDGPU::getSOPPWithRelaxation(Inst.getOpcode()) >= 0)
+  if (AMDGPU::getSOPPWithRelaxation(Opcode) >= 0)
     return true;
 
   return false;

llvm/lib/Target/ARM/MCTargetDesc/ARMAsmBackend.cpp

@@ -201,11 +201,9 @@ unsigned ARMAsmBackend::getRelaxedOpcode(unsigned Op,
   }
 }
 
-bool ARMAsmBackend::mayNeedRelaxation(const MCInst &Inst,
+bool ARMAsmBackend::mayNeedRelaxation(unsigned Opcode, ArrayRef<MCOperand>,
                                       const MCSubtargetInfo &STI) const {
-  if (getRelaxedOpcode(Inst.getOpcode(), STI) != Inst.getOpcode())
-    return true;
-  return false;
+  return getRelaxedOpcode(Opcode, STI) != Opcode;
 }
 
 static const char *checkPCRelOffset(uint64_t Value, int64_t Min, int64_t Max) {

llvm/lib/Target/ARM/MCTargetDesc/ARMAsmBackend.h

@@ -45,7 +45,7 @@ class ARMAsmBackend : public MCAsmBackend {
 
   unsigned getRelaxedOpcode(unsigned Op, const MCSubtargetInfo &STI) const;
 
-  bool mayNeedRelaxation(const MCInst &Inst,
+  bool mayNeedRelaxation(unsigned Opcode, ArrayRef<MCOperand> Operands,
                          const MCSubtargetInfo &STI) const override;
 
   const char *reasonForFixupRelaxation(const MCFixup &Fixup,

llvm/lib/Target/CSKY/MCTargetDesc/CSKYAsmBackend.cpp

@@ -239,9 +239,9 @@ void CSKYAsmBackend::applyFixup(const MCFragment &F, const MCFixup &Fixup,
   }
 }
 
-bool CSKYAsmBackend::mayNeedRelaxation(const MCInst &Inst,
+bool CSKYAsmBackend::mayNeedRelaxation(unsigned Opcode, ArrayRef<MCOperand>,
                                        const MCSubtargetInfo &STI) const {
-  switch (Inst.getOpcode()) {
+  switch (Opcode) {
   default:
     return false;
   case CSKY::JBR32: