Implement Import Call Optimization for x64

Author: dpaoliello

llvm/include/llvm/Transforms/CFGuard.h

@@ -16,6 +16,7 @@
 namespace llvm {
 
 class FunctionPass;
+class GlobalValue;
 
 class CFGuardPass : public PassInfoMixin<CFGuardPass> {
 public:
@@ -34,6 +35,8 @@ FunctionPass *createCFGuardCheckPass();
 /// Insert Control FLow Guard dispatches on indirect function calls.
 FunctionPass *createCFGuardDispatchPass();
 
+bool isCFGuardFunction(const GlobalValue *GV);
+
 } // namespace llvm
 
 #endif

llvm/lib/MC/MCObjectFileInfo.cpp

@@ -599,6 +599,11 @@ void MCObjectFileInfo::initCOFFMCObjectFileInfo(const Triple &T) {
   if (T.getArch() == Triple::aarch64) {
     ImportCallSection =
         Ctx->getCOFFSection(".impcall", COFF::IMAGE_SCN_LNK_INFO);
+  } else if (T.getArch() == Triple::x86_64) {
+    // Import Call Optimization on x64 leverages the same metadata as the
+    // retpoline mitigation, hence the unusual section name.
+    ImportCallSection =
+        Ctx->getCOFFSection(".retplne", COFF::IMAGE_SCN_LNK_INFO);
   }
 
   // Debug info.

llvm/lib/Target/X86/X86AsmPrinter.cpp

@@ -464,7 +464,8 @@ static bool isIndirectBranchOrTailCall(const MachineInstr &MI) {
          Opc == X86::TAILJMPr64 || Opc == X86::TAILJMPm64 ||
          Opc == X86::TCRETURNri || Opc == X86::TCRETURNmi ||
          Opc == X86::TCRETURNri64 || Opc == X86::TCRETURNmi64 ||
-         Opc == X86::TAILJMPr64_REX || Opc == X86::TAILJMPm64_REX;
+         Opc == X86::TCRETURNri64_ImpCall || Opc == X86::TAILJMPr64_REX ||
+         Opc == X86::TAILJMPm64_REX;
 }
 
 void X86AsmPrinter::emitBasicBlockEnd(const MachineBasicBlock &MBB) {
@@ -945,6 +946,9 @@ void X86AsmPrinter::emitStartOfAsmFile(Module &M) {
     OutStreamer->emitSymbolAttribute(S, MCSA_Global);
     OutStreamer->emitAssignment(
         S, MCConstantExpr::create(Feat00Value, MMI->getContext()));
+
+    if (M.getModuleFlag("import-call-optimization"))
+      EnableImportCallOptimization = true;
   }
   OutStreamer->emitSyntaxDirective();
 
@@ -1046,6 +1050,35 @@ void X86AsmPrinter::emitEndOfAsmFile(Module &M) {
     // safe to set.
     OutStreamer->emitAssemblerFlag(MCAF_SubsectionsViaSymbols);
   } else if (TT.isOSBinFormatCOFF()) {
+    // If import call optimization is enabled, emit the appropriate section.
+    // We do this whether or not we recorded any items.
+    if (EnableImportCallOptimization) {
+      OutStreamer->switchSection(getObjFileLowering().getImportCallSection());
+
+      // Section always starts with some magic.
+      constexpr char ImpCallMagic[12] = "RetpolineV1";
+      OutStreamer->emitBytes(StringRef{ImpCallMagic, sizeof(ImpCallMagic)});
+
+      // Layout of this section is:
+      // Per section that contains an item to record:
+      //  uint32_t SectionSize: Size in bytes for information in this section.
+      //  uint32_t Section Number
+      //  Per call to imported function in section:
+      //    uint32_t Kind: the kind of item.
+      //    uint32_t InstOffset: the offset of the instr in its parent section.
+      for (auto &[Section, CallsToImportedFuncs] :
+           SectionToImportedFunctionCalls) {
+        unsigned SectionSize =
+            sizeof(uint32_t) * (2 + 2 * CallsToImportedFuncs.size());
+        OutStreamer->emitInt32(SectionSize);
+        OutStreamer->emitCOFFSecNumber(Section->getBeginSymbol());
+        for (auto &[CallsiteSymbol, Kind] : CallsToImportedFuncs) {
+          OutStreamer->emitInt32(Kind);
+          OutStreamer->emitCOFFSecOffset(CallsiteSymbol);
+        }
+      }
+    }
+
     if (usesMSVCFloatingPoint(TT, M)) {
       // In Windows' libcmt.lib, there is a file which is linked in only if the
       // symbol _fltused is referenced. Linking this in causes some

llvm/lib/Target/X86/X86AsmPrinter.h

@@ -35,6 +35,26 @@ class LLVM_LIBRARY_VISIBILITY X86AsmPrinter : public AsmPrinter {
   bool EmitFPOData = false;
   bool ShouldEmitWeakSwiftAsyncExtendedFramePointerFlags = false;
   bool IndCSPrefix = false;
+  bool EnableImportCallOptimization = false;
+
+  enum ImportCallKind : unsigned {
+    IMAGE_RETPOLINE_AMD64_IMPORT_BR = 0x02,
+    IMAGE_RETPOLINE_AMD64_IMPORT_CALL = 0x03,
+    IMAGE_RETPOLINE_AMD64_INDIR_BR = 0x04,
+    IMAGE_RETPOLINE_AMD64_INDIR_CALL = 0x05,
+    IMAGE_RETPOLINE_AMD64_INDIR_BR_REX = 0x06,
+    IMAGE_RETPOLINE_AMD64_CFG_BR = 0x08,
+    IMAGE_RETPOLINE_AMD64_CFG_CALL = 0x09,
+    IMAGE_RETPOLINE_AMD64_CFG_BR_REX = 0x0A,
+    IMAGE_RETPOLINE_AMD64_SWITCHTABLE_FIRST = 0x010,
+    IMAGE_RETPOLINE_AMD64_SWITCHTABLE_LAST = 0x01F,
+  };
+  struct ImportCallInfo {
+    MCSymbol *CalleeSymbol;
+    ImportCallKind Kind;
+  };
+  DenseMap<MCSection *, std::vector<ImportCallInfo>>
+      SectionToImportedFunctionCalls;
 
   // This utility class tracks the length of a stackmap instruction's 'shadow'.
   // It is used by the X86AsmPrinter to ensure that the stackmap shadow
@@ -49,7 +69,7 @@ class LLVM_LIBRARY_VISIBILITY X86AsmPrinter : public AsmPrinter {
     void startFunction(MachineFunction &MF) {
       this->MF = &MF;
     }
-    void count(MCInst &Inst, const MCSubtargetInfo &STI,
+    void count(const MCInst &Inst, const MCSubtargetInfo &STI,
                MCCodeEmitter *CodeEmitter);
 
     // Called to signal the start of a shadow of RequiredSize bytes.
@@ -130,6 +150,12 @@ class LLVM_LIBRARY_VISIBILITY X86AsmPrinter : public AsmPrinter {
   void emitMachOIFuncStubHelperBody(Module &M, const GlobalIFunc &GI,
                                     MCSymbol *LazyPointer) override;
 
+  void emitCallInstruction(const llvm::MCInst &MCI);
+
+  // Emits a label to mark the next instruction as being relevant to Import Call
+  // Optimization.
+  void emitLabelAndRecordForImportCallOptimization(ImportCallKind Kind);
+
 public:
   X86AsmPrinter(TargetMachine &TM, std::unique_ptr<MCStreamer> Streamer);
 

llvm/lib/Target/X86/X86ExpandPseudo.cpp

@@ -274,6 +274,7 @@ bool X86ExpandPseudo::expandMI(MachineBasicBlock &MBB,
   case X86::TCRETURNdi64:
   case X86::TCRETURNdi64cc:
   case X86::TCRETURNri64:
+  case X86::TCRETURNri64_ImpCall:
   case X86::TCRETURNmi64: {
     bool isMem = Opcode == X86::TCRETURNmi || Opcode == X86::TCRETURNmi64;
     MachineOperand &JumpTarget = MBBI->getOperand(0);
@@ -345,12 +346,14 @@ bool X86ExpandPseudo::expandMI(MachineBasicBlock &MBB,
       MachineInstrBuilder MIB = BuildMI(MBB, MBBI, DL, TII->get(Op));
       for (unsigned i = 0; i != X86::AddrNumOperands; ++i)
         MIB.add(MBBI->getOperand(i));
-    } else if (Opcode == X86::TCRETURNri64) {
+    } else if ((Opcode == X86::TCRETURNri64) ||
+               (Opcode == X86::TCRETURNri64_ImpCall)) {
       JumpTarget.setIsKill();
       BuildMI(MBB, MBBI, DL,
               TII->get(IsWin64 ? X86::TAILJMPr64_REX : X86::TAILJMPr64))
           .add(JumpTarget);
     } else {
+      assert(!IsWin64 && "Win64 requires REX for indirect jumps.");
       JumpTarget.setIsKill();
       BuildMI(MBB, MBBI, DL, TII->get(X86::TAILJMPr))
           .add(JumpTarget);
@@ -875,6 +878,9 @@ bool X86ExpandPseudo::expandMI(MachineBasicBlock &MBB,
   case X86::CALL64m_RVMARKER:
     expandCALL_RVMARKER(MBB, MBBI);
     return true;
+  case X86::CALL64r_ImpCall:
+    MI.setDesc(TII->get(X86::CALL64r));
+    return true;
   case X86::ADD32mi_ND:
   case X86::ADD64mi32_ND:
   case X86::SUB32mi_ND:

llvm/lib/Target/X86/X86FastISel.cpp

@@ -34,6 +34,7 @@
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/IntrinsicsX86.h"
+#include "llvm/IR/Module.h"
 #include "llvm/IR/Operator.h"
 #include "llvm/MC/MCAsmInfo.h"
 #include "llvm/MC/MCSymbol.h"
@@ -3316,6 +3317,11 @@ bool X86FastISel::fastLowerCall(CallLoweringInfo &CLI) {
     if (Flag.isSwiftError() || Flag.isPreallocated())
       return false;
 
+  // Can't handle import call optimization.
+  if (Is64Bit &&
+      MF->getFunction().getParent()->getModuleFlag("import-call-optimization"))
+    return false;
+
   SmallVector<MVT, 16> OutVTs;
   SmallVector<Register, 16> ArgRegs;
 

llvm/lib/Target/X86/X86FrameLowering.cpp

@@ -2399,7 +2399,8 @@ X86FrameLowering::getWinEHFuncletFrameSize(const MachineFunction &MF) const {
 static bool isTailCallOpcode(unsigned Opc) {
   return Opc == X86::TCRETURNri || Opc == X86::TCRETURNdi ||
          Opc == X86::TCRETURNmi || Opc == X86::TCRETURNri64 ||
-         Opc == X86::TCRETURNdi64 || Opc == X86::TCRETURNmi64;
+         Opc == X86::TCRETURNri64_ImpCall || Opc == X86::TCRETURNdi64 ||
+         Opc == X86::TCRETURNmi64;
 }
 
 void X86FrameLowering::emitEpilogue(MachineFunction &MF,

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -19166,7 +19166,7 @@ SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
 
 SDValue X86TargetLowering::LowerExternalSymbol(SDValue Op,
                                                SelectionDAG &DAG) const {
-  return LowerGlobalOrExternal(Op, DAG, /*ForCall=*/false);
+  return LowerGlobalOrExternal(Op, DAG, /*ForCall=*/false, nullptr);
 }
 
 SDValue
@@ -19194,7 +19194,8 @@ X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
 /// Creates target global address or external symbol nodes for calls or
 /// other uses.
 SDValue X86TargetLowering::LowerGlobalOrExternal(SDValue Op, SelectionDAG &DAG,
-                                                 bool ForCall) const {
+                                                 bool ForCall,
+                                                 bool *IsImpCall) const {
   // Unpack the global address or external symbol.
   SDLoc dl(Op);
   const GlobalValue *GV = nullptr;
@@ -19244,6 +19245,16 @@ SDValue X86TargetLowering::LowerGlobalOrExternal(SDValue Op, SelectionDAG &DAG,
   if (ForCall && !NeedsLoad && !HasPICReg && Offset == 0)
     return Result;
 
+  // If Import Call Optimization is enabled and this is an imported function
+  // then make a note of it and return the global address without wrapping.
+  if (IsImpCall && (OpFlags == X86II::MO_DLLIMPORT) &&
+      Mod.getModuleFlag("import-call-optimization")) {
+    assert(ForCall && "Should only enable import call optimization if we are "
+                      "lowering a call");
+    *IsImpCall = true;
+    return Result;
+  }
+
   Result = DAG.getNode(getGlobalWrapperKind(GV, OpFlags), dl, PtrVT, Result);
 
   // With PIC, the address is actually $g + Offset.
@@ -19269,7 +19280,7 @@ SDValue X86TargetLowering::LowerGlobalOrExternal(SDValue Op, SelectionDAG &DAG,
 
 SDValue
 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
-  return LowerGlobalOrExternal(Op, DAG, /*ForCall=*/false);
+  return LowerGlobalOrExternal(Op, DAG, /*ForCall=*/false, nullptr);
 }
 
 static SDValue GetTLSADDR(SelectionDAG &DAG, GlobalAddressSDNode *GA,
@@ -34807,6 +34818,7 @@ const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
   NODE_NAME_CASE(FST)
   NODE_NAME_CASE(CALL)
   NODE_NAME_CASE(CALL_RVMARKER)
+  NODE_NAME_CASE(IMP_CALL)
   NODE_NAME_CASE(BT)
   NODE_NAME_CASE(CMP)
   NODE_NAME_CASE(FCMP)
@@ -62003,6 +62015,7 @@ X86TargetLowering::EmitKCFICheck(MachineBasicBlock &MBB,
   Register TargetReg;
   switch (MBBI->getOpcode()) {
   case X86::CALL64r:
+  case X86::CALL64r_ImpCall:
   case X86::CALL64r_NT:
   case X86::TAILJMPr64:
   case X86::TAILJMPr64_REX:

llvm/lib/Target/X86/X86ISelLowering.h

@@ -90,6 +90,10 @@ namespace llvm {
     /// POP_FROM_X87_REG (which may remove a required FPU stack pop).
     POP_FROM_X87_REG,
 
+    // Pseudo for a call to an imported function to ensure the correct machine
+    // instruction is emitted for Import Call Optimization.
+    IMP_CALL,
+
     /// X86 compare and logical compare instructions.
     CMP,
     FCMP,
@@ -1746,8 +1750,8 @@ namespace llvm {
 
     /// Creates target global address or external symbol nodes for calls or
     /// other uses.
-    SDValue LowerGlobalOrExternal(SDValue Op, SelectionDAG &DAG,
-                                  bool ForCall) const;
+    SDValue LowerGlobalOrExternal(SDValue Op, SelectionDAG &DAG, bool ForCall,
+                                  bool *IsImpCall) const;
 
     SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;

llvm/lib/Target/X86/X86ISelLoweringCall.cpp

@@ -2050,6 +2050,12 @@ X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
   if (CallConv == CallingConv::X86_INTR)
     report_fatal_error("X86 interrupts may not be called directly");
 
+  if (IsIndirectCall && !IsWin64 &&
+      M->getModuleFlag("import-call-optimization"))
+    errorUnsupported(DAG, dl,
+                     "Indirect calls must have a normal calling convention if "
+                     "Import Call Optimization is enabled");
+
   // Analyze operands of the call, assigning locations to each operand.
   SmallVector<CCValAssign, 16> ArgLocs;
   CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
@@ -2421,6 +2427,7 @@ X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
     InGlue = Chain.getValue(1);
   }
 
+  bool IsImpCall = false;
   if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
     assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
     // In the 64-bit large code model, we have to make all calls
@@ -2433,7 +2440,7 @@ X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
     // ForCall to true here has the effect of removing WrapperRIP when possible
     // to allow direct calls to be selected without first materializing the
     // address into a register.
-    Callee = LowerGlobalOrExternal(Callee, DAG, /*ForCall=*/true);
+    Callee = LowerGlobalOrExternal(Callee, DAG, /*ForCall=*/true, &IsImpCall);
   } else if (Subtarget.isTarget64BitILP32() &&
              Callee.getValueType() == MVT::i32) {
     // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
@@ -2555,7 +2562,9 @@ X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
 
   // Returns a chain & a glue for retval copy to use.
   SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
-  if (IsNoTrackIndirectCall) {
+  if (IsImpCall) {
+    Chain = DAG.getNode(X86ISD::IMP_CALL, dl, NodeTys, Ops);
+  } else if (IsNoTrackIndirectCall) {
     Chain = DAG.getNode(X86ISD::NT_CALL, dl, NodeTys, Ops);
   } else if (CLI.CB && objcarc::hasAttachedCallOpBundle(CLI.CB)) {
     // Calls with a "clang.arc.attachedcall" bundle are special. They should be