Add support for stack map generation in the X86 backend.

Originally implemented by Lang Hames.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@193811 91177308-0d34-0410-b5e6-96231b3b80d8

7 files changed, 395 insertions, 5 deletions

diff --git a/lib/CodeGen/CMakeLists.txt b/lib/CodeGen/CMakeLists.txt
index c01875c1f8..10cc9fff9a 100644
--- a/lib/CodeGen/CMakeLists.txt
+++ b/lib/CodeGen/CMakeLists.txt
@@ -97,6 +97,7 @@ add_llvm_library(LLVMCodeGen
   StackColoring.cpp
   StackProtector.cpp
   StackSlotColoring.cpp
+  StackMaps.cpp
   TailDuplication.cpp
   TargetFrameLoweringImpl.cpp
   TargetInstrInfo.cpp
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
index 69c82d4f0f..8ff8427b4d 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -33,6 +33,7 @@
 #include "llvm/CodeGen/MachineModuleInfo.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/SelectionDAG.h"
+#include "llvm/CodeGen/StackMaps.h"
 #include "llvm/DebugInfo.h"
 #include "llvm/IR/CallingConv.h"
 #include "llvm/IR/Constants.h"
@@ -6879,6 +6880,8 @@ void SelectionDAGBuilder::visitPatchpoint(const CallInst &CI) {
     SDValue OpVal = getValue(CI.getArgOperand(i));
     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(OpVal)) {
       Ops.push_back(
+        DAG.getTargetConstant(StackMaps::ConstantOp, MVT::i64));
+      Ops.push_back(
         DAG.getTargetConstant(C->getSExtValue(), MVT::i64));
     } else
       Ops.push_back(OpVal);
diff --git a/lib/CodeGen/StackMaps.cpp b/lib/CodeGen/StackMaps.cpp
new file mode 100644
index 0000000000..ffe46288b9
--- /dev/null
+++ b/lib/CodeGen/StackMaps.cpp
@@ -0,0 +1,213 @@
+//===---------------------------- StackMaps.cpp ---------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "stackmaps"
+
+#include "llvm/CodeGen/StackMaps.h"
+
+#include "llvm/CodeGen/AsmPrinter.h"
+#include "llvm/CodeGen/MachineInstr.h"
+#include "llvm/MC/MCContext.h"
+#include "llvm/MC/MCExpr.h"
+#include "llvm/MC/MCSectionMachO.h"
+#include "llvm/MC/MCStreamer.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetOpcodes.h"
+#include "llvm/Target/TargetMachine.h"
+#include "llvm/Target/TargetRegisterInfo.h"
+
+#include <iterator>
+
+using namespace llvm;
+
+void StackMaps::recordStackMap(const MachineInstr &MI, uint32_t ID,
+                               MachineInstr::const_mop_iterator MOI,
+                               MachineInstr::const_mop_iterator MOE) {
+
+  MCContext &OutContext = AP.OutStreamer.getContext();
+  MCSymbol *MILabel = OutContext.CreateTempSymbol();
+  AP.OutStreamer.EmitLabel(MILabel);
+
+  LocationVec CallsiteLocs;
+
+  while (MOI != MOE) {
+    std::pair<Location, MachineInstr::const_mop_iterator> ParseResult =
+      OpParser(MOI, MOE);
+
+    Location &Loc = ParseResult.first;
+
+    // Move large constants into the constant pool.
+    if (Loc.LocType == Location::Constant && (Loc.Offset & ~0xFFFFFFFFULL)) {
+      Loc.LocType = Location::ConstantIndex;
+      Loc.Offset = ConstPool.getConstantIndex(Loc.Offset);
+    }
+
+    CallsiteLocs.push_back(Loc);
+    MOI = ParseResult.second;
+  }
+
+  const MCExpr *CSOffsetExpr = MCBinaryExpr::CreateSub(
+    MCSymbolRefExpr::Create(MILabel, OutContext),
+    MCSymbolRefExpr::Create(AP.CurrentFnSym, OutContext),
+    OutContext);
+
+  CSInfos.push_back(CallsiteInfo(CSOffsetExpr, ID, CallsiteLocs));
+}
+
+/// serializeToStackMapSection conceptually populates the following fields:
+///
+/// uint32 : Reserved (header)
+/// uint32 : NumConstants
+/// int64  : Constants[NumConstants]
+/// uint32 : NumRecords
+/// StkMapRecord[NumRecords] {
+///   uint32 : PatchPoint ID
+///   uint32 : Instruction Offset
+///   uint16 : Reserved (record flags)
+///   uint16 : NumLocations
+///   Location[NumLocations] {
+///     uint8  : Register | Direct | Indirect | Constant | ConstantIndex
+///     uint8  : Reserved (location flags)
+///     uint16 : Dwarf RegNum
+///     int32  : Offset
+///   }
+/// }
+///
+/// Location Encoding, Type, Value:
+///   0x1, Register, Reg                 (value in register)
+///   0x2, Direct, Reg + Offset          (frame index)
+///   0x3, Indirect, [Reg + Offset]      (spilled value)
+///   0x4, Constant, Offset              (small constant)
+///   0x5, ConstIndex, Constants[Offset] (large constant)
+///
+void StackMaps::serializeToStackMapSection() {
+  // Bail out if there's no stack map data.
+  if (CSInfos.empty())
+    return;
+
+  MCContext &OutContext = AP.OutStreamer.getContext();
+  const TargetRegisterInfo *TRI = AP.TM.getRegisterInfo();
+
+  // Create the section.
+  const MCSection *StackMapSection =
+    OutContext.getMachOSection("__LLVM_STACKMAPS", "__llvm_stackmaps", 0,
+                               SectionKind::getMetadata());
+  AP.OutStreamer.SwitchSection(StackMapSection);
+
+  // Emit a dummy symbol to force section inclusion.
+  AP.OutStreamer.EmitLabel(
+    OutContext.GetOrCreateSymbol(Twine("__LLVM_StackMaps")));
+
+  // Serialize data.
+  const char *WSMP = "Stack Maps: ";
+  const MCRegisterInfo &MCRI = *OutContext.getRegisterInfo();
+
+  DEBUG(dbgs() << "********** Stack Map Output **********\n");
+
+  // Header.
+  AP.OutStreamer.EmitIntValue(0, 4);
+
+  // Num constants.
+  AP.OutStreamer.EmitIntValue(ConstPool.getNumConstants(), 4);
+
+  // Constant pool entries.
+  for (unsigned i = 0; i < ConstPool.getNumConstants(); ++i)
+    AP.OutStreamer.EmitIntValue(ConstPool.getConstant(i), 8);
+
+  DEBUG(dbgs() << WSMP << "#callsites = " << CSInfos.size() << "\n");
+  AP.OutStreamer.EmitIntValue(CSInfos.size(), 4);
+
+  for (CallsiteInfoList::const_iterator CSII = CSInfos.begin(),
+                                        CSIE = CSInfos.end();
+       CSII != CSIE; ++CSII) {
+
+    unsigned CallsiteID = CSII->ID;
+    const LocationVec &CSLocs = CSII->Locations;
+
+    DEBUG(dbgs() << WSMP << "callsite " << CallsiteID << "\n");
+
+    // Verify stack map entry. It's better to communicate a problem to the
+    // runtime than crash in case of in-process compilation. Currently, we do
+    // simple overflow checks, but we may eventually communicate other
+    // compilation errors this way.
+    if (CSLocs.size() > UINT16_MAX) {
+      AP.OutStreamer.EmitIntValue(UINT32_MAX, 4); // Invalid ID.
+      AP.OutStreamer.EmitValue(CSII->CSOffsetExpr, 4);
+      AP.OutStreamer.EmitIntValue(0, 2); // Reserved.
+      AP.OutStreamer.EmitIntValue(0, 2); // 0 locations.
+      continue;
+    }
+
+    AP.OutStreamer.EmitIntValue(CallsiteID, 4);
+    AP.OutStreamer.EmitValue(CSII->CSOffsetExpr, 4);
+
+    // Reserved for flags.
+    AP.OutStreamer.EmitIntValue(0, 2);
+
+    DEBUG(dbgs() << WSMP << "  has " << CSLocs.size() << " locations\n");
+
+    AP.OutStreamer.EmitIntValue(CSLocs.size(), 2);
+
+    unsigned operIdx = 0;
+    for (LocationVec::const_iterator LocI = CSLocs.begin(), LocE = CSLocs.end();
+         LocI != LocE; ++LocI, ++operIdx) {
+      const Location &Loc = *LocI;
+      DEBUG(
+        dbgs() << WSMP << "  Loc " << operIdx << ": ";
+        switch (Loc.LocType) {
+        case Location::Unprocessed:
+          dbgs() << "<Unprocessed operand>";
+          break;
+        case Location::Register:
+          dbgs() << "Register " << MCRI.getName(Loc.Reg);
+          break;
+        case Location::Direct:
+          dbgs() << "Direct " << MCRI.getName(Loc.Reg);
+          if (Loc.Offset)
+            dbgs() << " + " << Loc.Offset;
+          break;
+        case Location::Indirect:
+          dbgs() << "Indirect " << MCRI.getName(Loc.Reg)
+                 << " + " << Loc.Offset;
+          break;
+        case Location::Constant:
+          dbgs() << "Constant " << Loc.Offset;
+          break;
+        case Location::ConstantIndex:
+          dbgs() << "Constant Index " << Loc.Offset;
+          break;
+        }
+        dbgs() << "\n";
+      );
+
+      unsigned RegNo = 0;
+      if(Loc.Reg) {
+        RegNo = MCRI.getDwarfRegNum(Loc.Reg, false);
+        for (MCSuperRegIterator SR(Loc.Reg, TRI);
+             SR.isValid() && (int)RegNo < 0; ++SR) {
+          RegNo = TRI->getDwarfRegNum(*SR, false);
+        }
+      }
+      else {
+        assert((Loc.LocType != Location::Register
+                && Loc.LocType != Location::Register) &&
+               "Missing location register");
+      }
+      AP.OutStreamer.EmitIntValue(Loc.LocType, 1);
+      AP.OutStreamer.EmitIntValue(0, 1); // Reserved location flags.
+      AP.OutStreamer.EmitIntValue(RegNo, 2);
+      AP.OutStreamer.EmitIntValue(Loc.Offset, 4);
+    }
+  }
+
+  AP.OutStreamer.AddBlankLine();
+
+  CSInfos.clear();
+}
diff --git a/lib/Target/X86/X86AsmPrinter.cpp b/lib/Target/X86/X86AsmPrinter.cpp
index 84f633fa84..1a81ba8f41 100644
--- a/lib/Target/X86/X86AsmPrinter.cpp
+++ b/lib/Target/X86/X86AsmPrinter.cpp
@@ -626,6 +626,8 @@ void X86AsmPrinter::EmitEndOfAsmFile(Module &M) {
       OutStreamer.AddBlankLine();
     }
 
+    SM.serializeToStackMapSection();
+
     // Funny Darwin hack: This flag tells the linker that no global symbols
     // contain code that falls through to other global symbols (e.g. the obvious
     // implementation of multiple entry points).  If this doesn't occur, the
diff --git a/lib/Target/X86/X86AsmPrinter.h b/lib/Target/X86/X86AsmPrinter.h
index 6eed5ce08c..050c78df6a 100644
--- a/lib/Target/X86/X86AsmPrinter.h
+++ b/lib/Target/X86/X86AsmPrinter.h
@@ -16,6 +16,7 @@
 #include "llvm/CodeGen/AsmPrinter.h"
 #include "llvm/CodeGen/MachineModuleInfo.h"
 #include "llvm/CodeGen/ValueTypes.h"
+#include "llvm/CodeGen/StackMaps.h"
 #include "llvm/Support/Compiler.h"
 
 namespace llvm {
@@ -24,9 +25,20 @@ class MCStreamer;
 
 class LLVM_LIBRARY_VISIBILITY X86AsmPrinter : public AsmPrinter {
   const X86Subtarget *Subtarget;
+  StackMaps SM;
+
+  // Parses operands of PATCHPOINT and STACKMAP to produce stack map Location
+  // structures. Returns a result location and an iterator to the operand
+  // immediately following the operands consumed.
+  //
+  // This method is implemented in X86MCInstLower.cpp.
+  static std::pair<StackMaps::Location, MachineInstr::const_mop_iterator>
+    stackmapOperandParser(MachineInstr::const_mop_iterator MOI,
+                          MachineInstr::const_mop_iterator MOE);
+
  public:
   explicit X86AsmPrinter(TargetMachine &TM, MCStreamer &Streamer)
-    : AsmPrinter(TM, Streamer) {
+    : AsmPrinter(TM, Streamer), SM(*this, stackmapOperandParser) {
     Subtarget = &TM.getSubtarget<X86Subtarget>();
   }
 
diff --git a/lib/Target/X86/X86InstrInfo.cpp b/lib/Target/X86/X86InstrInfo.cpp
index 7f66c6ef11..369b031113 100644
--- a/lib/Target/X86/X86InstrInfo.cpp
+++ b/lib/Target/X86/X86InstrInfo.cpp
@@ -24,6 +24,7 @@
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/StackMaps.h"
 #include "llvm/IR/DerivedTypes.h"
 #include "llvm/IR/LLVMContext.h"
 #include "llvm/MC/MCAsmInfo.h"
@@ -4192,10 +4193,44 @@ breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
   MI->addRegisterKilled(Reg, TRI, true);
 }
 
-MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
-                                                  MachineInstr *MI,
-                                           const SmallVectorImpl<unsigned> &Ops,
-                                                  int FrameIndex) const {
+static MachineInstr* foldPatchpoint(MachineFunction &MF,
+                                    MachineInstr *MI,
+                                    const SmallVectorImpl<unsigned> &Ops,
+                                    int FrameIndex,
+                                    const TargetInstrInfo &TII) {
+  MachineInstr *NewMI =
+    MF.CreateMachineInstr(TII.get(MI->getOpcode()), MI->getDebugLoc(), true);
+  MachineInstrBuilder MIB(MF, NewMI);
+
+  bool isPatchPoint = MI->getOpcode() == TargetOpcode::PATCHPOINT;
+  unsigned StartIdx = isPatchPoint ? MI->getOperand(3).getImm() + 4 : 2;
+
+  // No need to fold the meta data and function arguments
+  for (unsigned i = 0; i < StartIdx; ++i)
+    MIB.addOperand(MI->getOperand(i));
+
+  for (unsigned i = StartIdx; i < MI->getNumOperands(); ++i) {
+    MachineOperand &MO = MI->getOperand(i);
+    if (std::find(Ops.begin(), Ops.end(), i) != Ops.end()) {
+      MIB.addOperand(MachineOperand::CreateImm(StackMaps::IndirectMemRefOp));
+      MIB.addOperand(MachineOperand::CreateFI(FrameIndex));
+      addOffset(MIB, 0);
+    }
+    else
+      MIB.addOperand(MO);
+  }
+  return NewMI;
+}
+
+MachineInstr*
+X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF, MachineInstr *MI,
+                                    const SmallVectorImpl<unsigned> &Ops,
+                                    int FrameIndex) const {
+  // Special case stack map and patch point intrinsics.
+  if (MI->getOpcode() == TargetOpcode::STACKMAP
+      || MI->getOpcode() == TargetOpcode::PATCHPOINT) {
+    return foldPatchpoint(MF, MI, Ops, FrameIndex, *this);
+  }
   // Check switch flag
   if (NoFusing) return NULL;
 
diff --git a/lib/Target/X86/X86MCInstLower.cpp b/lib/Target/X86/X86MCInstLower.cpp
index bd2be3e6c9..04d1ad57a7 100644
--- a/lib/Target/X86/X86MCInstLower.cpp
+++ b/lib/Target/X86/X86MCInstLower.cpp
@@ -17,6 +17,7 @@
 #include "X86COFFMachineModuleInfo.h"
 #include "llvm/ADT/SmallString.h"
 #include "llvm/CodeGen/MachineModuleInfoImpls.h"
+#include "llvm/CodeGen/StackMaps.h"
 #include "llvm/IR/Type.h"
 #include "llvm/MC/MCAsmInfo.h"
 #include "llvm/MC/MCContext.h"
@@ -686,6 +687,123 @@ static void LowerTlsAddr(MCStreamer &OutStreamer,
     .addExpr(tlsRef));
 }
 
+static std::pair<StackMaps::Location, MachineInstr::const_mop_iterator>
+parseMemoryOperand(StackMaps::Location::LocationType LocTy,
+                   MachineInstr::const_mop_iterator MOI,
+                   MachineInstr::const_mop_iterator MOE) {
+
+  typedef StackMaps::Location Location;
+
+  assert(std::distance(MOI, MOE) >= 5 && "Too few operands to encode mem op.");
+
+  const MachineOperand &Base = *MOI;
+  const MachineOperand &Scale = *(++MOI);
+  const MachineOperand &Index = *(++MOI);
+  const MachineOperand &Disp = *(++MOI);
+  const MachineOperand &ZeroReg = *(++MOI);
+
+  // Sanity check for supported operand format.
+  assert(Base.isReg() &&
+         Scale.isImm() && Scale.getImm() == 1 &&
+         Index.isReg() && Index.getReg() == 0 &&
+         Disp.isImm() && ZeroReg.isReg() && (ZeroReg.getReg() == 0) &&
+         "Unsupported x86 memory operand sequence.");
+
+  return std::make_pair(
+           Location(LocTy, Base.getReg(), Disp.getImm()), ++MOI);
+}
+
+std::pair<StackMaps::Location, MachineInstr::const_mop_iterator>
+X86AsmPrinter::stackmapOperandParser(MachineInstr::const_mop_iterator MOI,
+                                     MachineInstr::const_mop_iterator MOE) {
+
+  typedef StackMaps::Location Location;
+
+  const MachineOperand &MOP = *MOI;
+  assert(!MOP.isRegMask() && (!MOP.isReg() || !MOP.isImplicit()) &&
+         "Register mask and implicit operands should not be processed.");
+
+  if (MOP.isImm()) {
+    switch (MOP.getImm()) {
+    default: llvm_unreachable("Unrecognized operand type.");
+    case StackMaps::DirectMemRefOp:
+      return parseMemoryOperand(StackMaps::Location::Direct,
+                                llvm::next(MOI), MOE);
+    case StackMaps::IndirectMemRefOp:
+      return parseMemoryOperand(StackMaps::Location::Indirect,
+                                llvm::next(MOI), MOE);
+    case StackMaps::ConstantOp: {
+      ++MOI;
+      assert(MOI->isImm() && "Expected constant operand.");
+      int64_t Imm = MOI->getImm();
+      return std::make_pair(Location(Location::Constant, 0, Imm), ++MOI);
+    }
+    }
+  }
+
+  // Otherwise this is a reg operand.
+  assert(MOP.isReg() && "Expected register operand here.");
+  assert(TargetRegisterInfo::isPhysicalRegister(MOP.getReg()) &&
+         "Virtreg operands should have been rewritten before now.");
+  return std::make_pair(Location(Location::Register, MOP.getReg(), 0), ++MOI);
+}
+
+static MachineInstr::const_mop_iterator
+getStackMapEndMOP(MachineInstr::const_mop_iterator MOI,
+                  MachineInstr::const_mop_iterator MOE) {
+  for (; MOI != MOE; ++MOI)
+    if (MOI->isRegMask() || (MOI->isReg() && MOI->isImplicit()))
+      break;
+
+  return MOI;
+}
+
+static void LowerSTACKMAP(MCStreamer &OutStreamer,
+                          X86MCInstLower &MCInstLowering,
+                          StackMaps &SM,
+                          const MachineInstr &MI)
+{
+  int64_t ID = MI.getOperand(0).getImm();
+  unsigned NumNOPBytes = MI.getOperand(1).getImm();
+
+  assert((int32_t)ID == ID && "Stack maps hold 32-bit IDs");
+  SM.recordStackMap(MI, ID, llvm::next(MI.operands_begin(), 2),
+                    getStackMapEndMOP(MI.operands_begin(), MI.operands_end()));
+  // Emit padding.
+  for (unsigned i = 0; i < NumNOPBytes; ++i)
+    OutStreamer.EmitInstruction(MCInstBuilder(X86::NOOP));
+}
+
+static void LowerPATCHPOINT(MCStreamer &OutStreamer,
+                            X86MCInstLower &MCInstLowering,
+                            StackMaps &SM,
+                            const MachineInstr &MI)
+{
+  int64_t ID = MI.getOperand(0).getImm();
+  assert((int32_t)ID == ID && "Stack maps hold 32-bit IDs");
+
+  // Get the number of arguments participating in the call. This number was
+  // adjusted during call lowering by subtracting stack args.
+  int64_t StackMapIdx = MI.getOperand(3).getImm() + 4;
+  assert(StackMapIdx <= MI.getNumOperands() && "Patchpoint dropped args.");
+
+  SM.recordStackMap(MI, ID, llvm::next(MI.operands_begin(), StackMapIdx),
+                     getStackMapEndMOP(MI.operands_begin(), MI.operands_end()));
+
+  // Emit call. We need to know how many bytes we encoded here.
+  unsigned EncodedBytes = 2;
+  OutStreamer.EmitInstruction(MCInstBuilder(X86::CALL64r)
+                              .addReg(MI.getOperand(2).getReg()));
+
+  // Emit padding.
+  unsigned NumNOPBytes = MI.getOperand(1).getImm();
+  assert(NumNOPBytes >= EncodedBytes &&
+         "Patchpoint can't request size less than the length of a call.");
+
+  for (unsigned i = EncodedBytes; i < NumNOPBytes; ++i)
+    OutStreamer.EmitInstruction(MCInstBuilder(X86::NOOP));
+}
+
 void X86AsmPrinter::EmitInstruction(const MachineInstr *MI) {
   X86MCInstLower MCInstLowering(*MF, *this);
   switch (MI->getOpcode()) {
@@ -775,6 +893,12 @@ void X86AsmPrinter::EmitInstruction(const MachineInstr *MI) {
       .addExpr(DotExpr));
     return;
   }
+
+  case TargetOpcode::STACKMAP:
+    return LowerSTACKMAP(OutStreamer, MCInstLowering, SM, *MI);
+
+  case TargetOpcode::PATCHPOINT:
+    return LowerPATCHPOINT(OutStreamer, MCInstLowering, SM, *MI);
   }
 
   MCInst TmpInst;