[XCore] Move lowering of thread local storage to a separate pass.

Thread local storage is not supported by the XMOS linker so we handle
thread local variables by lowering the variable to an array of n elements
(where n is the number of hardware threads per core, currently 8
for all XMOS devices) indexed by the the current thread ID.

Previously this lowering was spread across the XCoreISelLowering and the
XCoreAsmPrinter classes. Moving this to a separate pass should be much
cleaner.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181124 91177308-0d34-0410-b5e6-96231b3b80d8

6 files changed, 158 insertions, 55 deletions

diff --git a/lib/Target/XCore/CMakeLists.txt b/lib/Target/XCore/CMakeLists.txt
index 099ad390d2..d5bfddc23e 100644
--- a/lib/Target/XCore/CMakeLists.txt
+++ b/lib/Target/XCore/CMakeLists.txt
@@ -15,6 +15,7 @@ add_llvm_target(XCoreCodeGen
   XCoreInstrInfo.cpp
   XCoreISelDAGToDAG.cpp
   XCoreISelLowering.cpp
+  XCoreLowerThreadLocal.cpp
   XCoreMachineFunctionInfo.cpp
   XCoreMCInstLower.cpp
   XCoreRegisterInfo.cpp
diff --git a/lib/Target/XCore/XCore.h b/lib/Target/XCore/XCore.h
index 08f091e5b8..2f375fc952 100644
--- a/lib/Target/XCore/XCore.h
+++ b/lib/Target/XCore/XCore.h
@@ -20,12 +20,16 @@
 
 namespace llvm {
   class FunctionPass;
+  class ModulePass;
   class TargetMachine;
   class XCoreTargetMachine;
   class formatted_raw_ostream;
 
+  void initializeXCoreLowerThreadLocalPass(PassRegistry &p);
+
   FunctionPass *createXCoreISelDag(XCoreTargetMachine &TM,
                                    CodeGenOpt::Level OptLevel);
+  ModulePass *createXCoreLowerThreadLocalPass();
 
 } // end namespace llvm;
 
diff --git a/lib/Target/XCore/XCoreAsmPrinter.cpp b/lib/Target/XCore/XCoreAsmPrinter.cpp
index 0d146ba4d9..e177ad300c 100644
--- a/lib/Target/XCore/XCoreAsmPrinter.cpp
+++ b/lib/Target/XCore/XCoreAsmPrinter.cpp
@@ -36,7 +36,6 @@
 #include "llvm/MC/MCInst.h"
 #include "llvm/MC/MCStreamer.h"
 #include "llvm/MC/MCSymbol.h"
-#include "llvm/Support/CommandLine.h"
 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/Support/TargetRegistry.h"
 #include "llvm/Support/raw_ostream.h"
@@ -46,12 +45,6 @@
 #include <cctype>
 using namespace llvm;
 
-static cl::opt<unsigned> MaxThreads("xcore-max-threads", cl::Optional,
-  cl::desc("Maximum number of threads (for emulation thread-local storage)"),
-  cl::Hidden,
-  cl::value_desc("number"),
-  cl::init(8));
-
 namespace {
   class XCoreAsmPrinter : public AsmPrinter {
     const XCoreSubtarget &Subtarget;
@@ -152,10 +145,10 @@ void XCoreAsmPrinter::EmitGlobalVariable(const GlobalVariable *GV) {
 
   EmitAlignment(Align > 2 ? Align : 2, GV);
   
-  unsigned Size = TD->getTypeAllocSize(C->getType());
   if (GV->isThreadLocal()) {
-    Size *= MaxThreads;
+    report_fatal_error("TLS is not supported by this target!");
   }
+  unsigned Size = TD->getTypeAllocSize(C->getType());
   if (MAI->hasDotTypeDotSizeDirective()) {
     OutStreamer.EmitSymbolAttribute(GVSym, MCSA_ELF_TypeObject);
     OutStreamer.EmitRawText("\t.size " + Twine(GVSym->getName()) + "," +
@@ -164,10 +157,6 @@ void XCoreAsmPrinter::EmitGlobalVariable(const GlobalVariable *GV) {
   OutStreamer.EmitLabel(GVSym);
   
   EmitGlobalConstant(C);
-  if (GV->isThreadLocal()) {
-    for (unsigned i = 1; i < MaxThreads; ++i)
-      EmitGlobalConstant(C);
-  }
   // The ABI requires that unsigned scalar types smaller than 32 bits
   // are padded to 32 bits.
   if (Size < 4)
diff --git a/lib/Target/XCore/XCoreISelLowering.cpp b/lib/Target/XCore/XCoreISelLowering.cpp
index f0346f6715..28472ea339 100644
--- a/lib/Target/XCore/XCoreISelLowering.cpp
+++ b/lib/Target/XCore/XCoreISelLowering.cpp
@@ -120,9 +120,6 @@ XCoreTargetLowering::XCoreTargetLowering(XCoreTargetMachine &XTM)
   setOperationAction(ISD::GlobalAddress, MVT::i32,   Custom);
   setOperationAction(ISD::BlockAddress, MVT::i32 , Custom);
 
-  // Thread Local Storage
-  setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
-
   // Conversion of i64 -> double produces constantpool nodes
   setOperationAction(ISD::ConstantPool, MVT::i32,   Custom);
 
@@ -172,7 +169,6 @@ LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   switch (Op.getOpcode())
   {
   case ISD::GlobalAddress:      return LowerGlobalAddress(Op, DAG);
-  case ISD::GlobalTLSAddress:   return LowerGlobalTLSAddress(Op, DAG);
   case ISD::BlockAddress:       return LowerBlockAddress(Op, DAG);
   case ISD::ConstantPool:       return LowerConstantPool(Op, DAG);
   case ISD::BR_JT:              return LowerBR_JT(Op, DAG);
@@ -255,44 +251,6 @@ static inline SDValue BuildGetId(SelectionDAG &DAG, DebugLoc dl) {
                      DAG.getConstant(Intrinsic::xcore_getid, MVT::i32));
 }
 
-static inline bool isZeroLengthArray(Type *Ty) {
-  ArrayType *AT = dyn_cast_or_null<ArrayType>(Ty);
-  return AT && (AT->getNumElements() == 0);
-}
-
-SDValue XCoreTargetLowering::
-LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
-{
-  // FIXME there isn't really debug info here
-  DebugLoc dl = Op.getDebugLoc();
-  // transform to label + getid() * size
-  const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
-  SDValue GA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32);
-  const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
-  if (!GVar) {
-    // If GV is an alias then use the aliasee to determine size
-    if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
-      GVar = dyn_cast_or_null<GlobalVariable>(GA->resolveAliasedGlobal());
-  }
-  if (!GVar) {
-    llvm_unreachable("Thread local object not a GlobalVariable?");
-  }
-  Type *Ty = cast<PointerType>(GV->getType())->getElementType();
-  if (!Ty->isSized() || isZeroLengthArray(Ty)) {
-#ifndef NDEBUG
-    errs() << "Size of thread local object " << GVar->getName()
-           << " is unknown\n";
-#endif
-    llvm_unreachable(0);
-  }
-  SDValue base = getGlobalAddressWrapper(GA, GV, DAG);
-  const DataLayout *TD = TM.getDataLayout();
-  unsigned Size = TD->getTypeAllocSize(Ty);
-  SDValue offset = DAG.getNode(ISD::MUL, dl, MVT::i32, BuildGetId(DAG, dl),
-                       DAG.getConstant(Size, MVT::i32));
-  return DAG.getNode(ISD::ADD, dl, MVT::i32, base, offset);
-}
-
 SDValue XCoreTargetLowering::
 LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const
 {
diff --git a/lib/Target/XCore/XCoreLowerThreadLocal.cpp b/lib/Target/XCore/XCoreLowerThreadLocal.cpp
new file mode 100644
index 0000000000..2e328b4e34
--- /dev/null
+++ b/lib/Target/XCore/XCoreLowerThreadLocal.cpp
@@ -0,0 +1,145 @@
+//===-- XCoreLowerThreadLocal - Lower thread local variables --------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+///
+/// \file
+/// \brief This file contains a pass that lowers thread local variables on the
+///        XCore.
+///
+//===----------------------------------------------------------------------===//
+
+#include "XCore.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/DerivedTypes.h"
+#include "llvm/IR/GlobalVariable.h"
+#include "llvm/IR/Intrinsics.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/Module.h"
+#include "llvm/Pass.h"
+#include "llvm/Support/CommandLine.h"
+
+#define DEBUG_TYPE "xcore-lower-thread-local"
+
+using namespace llvm;
+
+static cl::opt<unsigned> MaxThreads(
+  "xcore-max-threads", cl::Optional,
+  cl::desc("Maximum number of threads (for emulation thread-local storage)"),
+  cl::Hidden, cl::value_desc("number"), cl::init(8));
+
+namespace {
+  /// Lowers thread local variables on the XCore. Each thread local variable is
+  /// expanded to an array of n elements indexed by the thread ID where n is the
+  /// fixed number hardware threads supported by the device.
+  struct XCoreLowerThreadLocal : public ModulePass {
+    static char ID;
+
+    XCoreLowerThreadLocal() : ModulePass(ID) {
+      initializeXCoreLowerThreadLocalPass(*PassRegistry::getPassRegistry());
+    }
+
+    bool lowerGlobal(GlobalVariable *GV);
+
+    bool runOnModule(Module &M);
+  };
+}
+
+char XCoreLowerThreadLocal::ID = 0;
+
+INITIALIZE_PASS(XCoreLowerThreadLocal, "xcore-lower-thread-local",
+                "Lower thread local variables", false, false)
+
+ModulePass *llvm::createXCoreLowerThreadLocalPass() {
+  return new XCoreLowerThreadLocal();
+}
+
+static ArrayType *createLoweredType(Type *OriginalType) {
+  return ArrayType::get(OriginalType, MaxThreads);
+}
+
+static Constant *
+createLoweredInitializer(ArrayType *NewType, Constant *OriginalInitializer) {
+  SmallVector<Constant *, 8> Elements(MaxThreads);
+  for (unsigned i = 0; i != MaxThreads; ++i) {
+    Elements[i] = OriginalInitializer;
+  }
+  return ConstantArray::get(NewType, Elements);
+}
+
+static bool hasNonInstructionUse(GlobalVariable *GV) {
+  for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
+       ++UI)
+    if (!isa<Instruction>(*UI))
+      return true;
+
+  return false;
+}
+
+static bool isZeroLengthArray(Type *Ty) {
+  ArrayType *AT = dyn_cast<ArrayType>(Ty);
+  return AT && (AT->getNumElements() == 0);
+}
+
+bool XCoreLowerThreadLocal::lowerGlobal(GlobalVariable *GV) {
+  Module *M = GV->getParent();
+  LLVMContext &Ctx = M->getContext();
+  if (!GV->isThreadLocal())
+    return false;
+
+  // Skip globals that we can't lower and leave it for the backend to error.
+  if (hasNonInstructionUse(GV) ||
+      !GV->getType()->isSized() || isZeroLengthArray(GV->getType()))
+    return false;
+
+  // Create replacement global.
+  ArrayType *NewType = createLoweredType(GV->getType()->getElementType());
+  Constant *NewInitializer = createLoweredInitializer(NewType,
+                                                      GV->getInitializer());
+  GlobalVariable *NewGV =
+    new GlobalVariable(*M, NewType, GV->isConstant(), GV->getLinkage(),
+                       NewInitializer, "", 0, GlobalVariable::NotThreadLocal,
+                       GV->getType()->getAddressSpace(),
+                       GV->isExternallyInitialized());
+
+  // Update uses.
+  SmallVector<User *, 16> Users(GV->use_begin(), GV->use_end());
+  for (unsigned I = 0, E = Users.size(); I != E; ++I) {
+    User *U = Users[I];
+    Instruction *Inst = cast<Instruction>(U);
+    IRBuilder<> Builder(Inst);
+    Function *GetID = Intrinsic::getDeclaration(GV->getParent(),
+                                                Intrinsic::xcore_getid);
+    Value *ThreadID = Builder.CreateCall(GetID);
+    SmallVector<Value *, 2> Indices;
+    Indices.push_back(Constant::getNullValue(Type::getInt64Ty(Ctx)));
+    Indices.push_back(ThreadID);
+    Value *Addr = Builder.CreateInBoundsGEP(NewGV, Indices);
+    U->replaceUsesOfWith(GV, Addr);
+  }
+
+  // Remove old global.
+  NewGV->takeName(GV);
+  GV->eraseFromParent();
+  return true;
+}
+
+bool XCoreLowerThreadLocal::runOnModule(Module &M) {
+  // Find thread local globals.
+  bool MadeChange = false;
+  SmallVector<GlobalVariable *, 16> ThreadLocalGlobals;
+  for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
+       GVI != E; ++GVI) {
+    GlobalVariable *GV = GVI;
+    if (GV->isThreadLocal())
+      ThreadLocalGlobals.push_back(GV);
+  }
+  for (unsigned I = 0, E = ThreadLocalGlobals.size(); I != E; ++I) {
+    MadeChange |= lowerGlobal(ThreadLocalGlobals[I]);
+  }
+  return MadeChange;
+}
diff --git a/lib/Target/XCore/XCoreTargetMachine.cpp b/lib/Target/XCore/XCoreTargetMachine.cpp
index 28c3d12c05..07e5fff141 100644
--- a/lib/Target/XCore/XCoreTargetMachine.cpp
+++ b/lib/Target/XCore/XCoreTargetMachine.cpp
@@ -46,6 +46,7 @@ public:
     return getTM<XCoreTargetMachine>();
   }
 
+  virtual bool addPreISel();
   virtual bool addInstSelector();
 };
 } // namespace
@@ -54,6 +55,11 @@ TargetPassConfig *XCoreTargetMachine::createPassConfig(PassManagerBase &PM) {
   return new XCorePassConfig(this, PM);
 }
 
+bool XCorePassConfig::addPreISel() {
+  addPass(createXCoreLowerThreadLocalPass());
+  return false;
+}
+
 bool XCorePassConfig::addInstSelector() {
   addPass(createXCoreISelDag(getXCoreTargetMachine(), getOptLevel()));
   return false;