Concurrency support for PNaCl ABI

lib/Transforms/NaCl/RewriteAtomics.cpp

Index: lib/Transforms/NaCl/RewriteAtomics.cpp
diff --git a/lib/Transforms/NaCl/RewriteAtomics.cpp b/lib/Transforms/NaCl/RewriteAtomics.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..2e86dd09f8f7b7cfbca8f31a0f56f9af5a6c7f1d
--- /dev/null
+++ b/lib/Transforms/NaCl/RewriteAtomics.cpp
@@ -0,0 +1,259 @@
+//===- RewriteAtomics.cpp - Stabilize instructions used for concurrency ---===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This pass encodes atomics, volatiles and fences using NaCl intrinsics
+// instead of LLVM's regular IR instructions.
+//
+// All of the above are transformed into one of the
+// @llvm.nacl.atomic.* intrinsics.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/ADT/Twine.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/Intrinsics.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/NaClIntrinsics.h"
+#include "llvm/InstVisitor.h"
+#include "llvm/Pass.h"
+#include "llvm/Support/Compiler.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Transforms/NaCl.h"
+
+using namespace llvm;
+
+namespace {
+class RewriteAtomics : public ModulePass {
+public:
+  static char ID; // Pass identification, replacement for typeid
+  RewriteAtomics() : ModulePass(ID) {
+    // This is a module pass because it may have to introduce
+    // intrinsic declarations into the module and modify a global function.
+    initializeRewriteAtomicsPass(*PassRegistry::getPassRegistry());
+  }
+
+  virtual bool runOnModule(Module &M);
+};
+
+template <class T> Twine ToTwine(const T &V) {
+  std::string S;
+  raw_string_ostream OS(S);
+  OS << const_cast<T &>(V);
+  return OS.str();
+}
+
+class AtomicVisitor : public InstVisitor<AtomicVisitor> {
+  Module &M;
+  LLVMContext &C;
+  NaCl::AtomicIntrinsics AI;
+  bool ModifiedModule;
+
+  AtomicVisitor() LLVM_DELETED_FUNCTION;
+  AtomicVisitor(const AtomicVisitor &) LLVM_DELETED_FUNCTION;
+  AtomicVisitor &operator=(const AtomicVisitor &) LLVM_DELETED_FUNCTION;
+
+  template <class Instruction>
+  ConstantInt *freezeMemoryOrder(const Instruction &I) const;
+  void checkSizeMatchesType(const Instruction &I, unsigned S,
+                            const Type *T) const;
+  void checkAlignment(const Instruction &I, unsigned Alignment,
+                      unsigned Size) const;
+  void replaceInstructionWithIntrinsicCall(Instruction &I, Intrinsic::ID ID,
+                                           Type *OverloadedType,
+                                           ArrayRef<Value *> Args);
+
+  // Most atomics instructions deal with at least one pointer, this
+  // struct automates some of this and has generic sanity checks.
+  template <class Instruction> struct PointerHelper {
+    Value *P;
+    Type *PET;
+    unsigned Size;
+    PointerHelper(const AtomicVisitor &AV, Instruction &I)
+        : P(I.getPointerOperand()) {
+      if (I.getPointerAddressSpace() != 0)
+        report_fatal_error("unhandled pointer address space " +
+                           Twine(I.getPointerAddressSpace()) + " for atomic: " +
+                           ToTwine(I));
+      assert(P->getType()->isPointerTy() && "expected a pointer");
+      PET = P->getType()->getPointerElementType();
+      Size = PET->getIntegerBitWidth();
+      AV.checkSizeMatchesType(I, Size, PET);
+    }
+  };
+
+public:
+  AtomicVisitor(Module &M)

[jvoung (off chromium), 2013/07/03 20:14:20]

nit: Might be more useful to have public API be listed first instead of last.

[JF, 2013/07/03 22:28:30]

Done.

+      : M(M), C(M.getContext()), AI(C), ModifiedModule(false) {}
+  ~AtomicVisitor() {}
+  bool modifiedModule() const { return ModifiedModule; }
+
+  void visitLoadInst(LoadInst &I);
+  void visitStoreInst(StoreInst &I);
+  void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I);
+  void visitAtomicRMWInst(AtomicRMWInst &I);
+  void visitFenceInst(FenceInst &I);
+};
+}
+
+char RewriteAtomics::ID = 0;
+INITIALIZE_PASS(RewriteAtomics, "nacl-rewrite-atomics",
+                "rewrite atomics, volatiles and fences into stable "
+                "@llvm.nacl.atomics.* intrinsics",
+                false, false)
+
+bool RewriteAtomics::runOnModule(Module &M) {
+  AtomicVisitor AV(M);
+  AV.visit(M);
+  return AV.modifiedModule();
+}
+
+template <class Instruction>
+ConstantInt *AtomicVisitor::freezeMemoryOrder(const Instruction &I) const {
+  NaCl::MemoryOrder AO = NaCl::MemoryOrderInvalid;
+
+  // TODO Volatile load/store are promoted to sequentially consistent
+  //      for now. We could do something weaker.
+  if (const LoadInst *L = dyn_cast<LoadInst>(&I)) {
+    if (L->isVolatile())
+      AO = NaCl::MemoryOrderSequentiallyConsistent;
+  } else if (const StoreInst *S = dyn_cast<StoreInst>(&I)) {
+    if (S->isVolatile())
+      AO = NaCl::MemoryOrderSequentiallyConsistent;
+  }
+
+  if (AO == NaCl::MemoryOrderInvalid)

[jvoung (off chromium), 2013/07/03 20:14:20]

Put curlies around the if's then-statement since it's more than one line?

[eliben, 2013/07/03 20:24:21]

Inspired by Steve's stories from yesterday, JF ;-) ?

[JF, 2013/07/03 22:28:30]

Done.

+    switch (I.getOrdering()) {
+    default:
+    case NotAtomic: llvm_unreachable("unexpected memory order");
+    // Monotonic is a strict superset of Unordered. Both can therefore
+    // map to Relaxed ordering, which is in the C11/C++11 standard.
+    case Unordered: AO = NaCl::MemoryOrderRelaxed; break;
+    case Monotonic: AO = NaCl::MemoryOrderRelaxed; break;
+    // TODO Consume is currently unspecified by LLVM's internal IR.
+    case Acquire: AO = NaCl::MemoryOrderAcquire; break;
+    case Release: AO = NaCl::MemoryOrderRelease; break;
+    case AcquireRelease: AO = NaCl::MemoryOrderAcquireRelease; break;
+    case SequentiallyConsistent:
+      AO = NaCl::MemoryOrderSequentiallyConsistent; break;
+    }
+
+  // TODO For now only sequential consistency is allowed.
+  AO = NaCl::MemoryOrderSequentiallyConsistent;
+
+  return ConstantInt::get(Type::getInt32Ty(C), AO);
+}
+
+void AtomicVisitor::checkSizeMatchesType(const Instruction &I, unsigned S,
+                                         const Type *T) const {
+  Type *IntType = Type::getIntNTy(C, S);
+  if (IntType && T == IntType)
+    return;
+  report_fatal_error("unsupported atomic type " + ToTwine(*T) + " of size " +
+                     Twine(S) + " in: " + ToTwine(I));
+}
+
+void AtomicVisitor::checkAlignment(const Instruction &I, unsigned Alignment,

[jvoung (off chromium), 2013/07/03 17:50:26]

nit: Perhaps name variables AlignmentBits, SizeBits or something to be clear
about the units. 

[JF, 2013/07/03 22:28:30]

Done.

+                                   unsigned Size) const {
+  if (Alignment < Size)
+    report_fatal_error("atomic load/store must be at least naturally aligned, "
+                       "got " + Twine(Alignment) + ", expected at least " +
+                       Twine(Size) + ", in: " + ToTwine(I));
+}
+
+void AtomicVisitor::replaceInstructionWithIntrinsicCall(
+    Instruction &I, Intrinsic::ID ID, Type *OverloadedType,
+    ArrayRef<Value *> Args) {
+  Function *F = AI.find(ID, OverloadedType)->getDeclaration(&M);
+  CallInst *Call = CallInst::Create(F, Args, "", &I);
+  Call->setDebugLoc(I.getDebugLoc());
+  I.replaceAllUsesWith(Call);
+  I.eraseFromParent();
+  ModifiedModule = true;
+}
+
+// %res = load {atomic|volatile} T* %ptr memory_order, align sizeof(T)
+// becomes:
+// %res = call T @llvm.nacl.atomic.load.i<size>(%ptr, memory_order)
+void AtomicVisitor::visitLoadInst(LoadInst &I) {
+  if (I.isSimple())
+    return;
+  PointerHelper<LoadInst> PH(*this, I);
+  checkAlignment(I, I.getAlignment() * 8, PH.Size);
+  Value *Args[] = { PH.P, freezeMemoryOrder(I) };
+  replaceInstructionWithIntrinsicCall(I, Intrinsic::nacl_atomic_load, PH.PET,
+                                      Args);
+}
+
+// store {atomic|volatile} T %val, T* %ptr memory_order, align sizeof(T)
+// becomes:
+// call void @llvm.nacl.atomic.store.i<size>(%val, %ptr, memory_order)
+void AtomicVisitor::visitStoreInst(StoreInst &I) {
+  if (I.isSimple())
+    return;
+  PointerHelper<StoreInst> PH(*this, I);
+  checkAlignment(I, I.getAlignment() * 8, PH.Size);
+  checkSizeMatchesType(I, PH.Size, I.getValueOperand()->getType());
+  Value *Args[] = { I.getValueOperand(), PH.P, freezeMemoryOrder(I) };
+  replaceInstructionWithIntrinsicCall(I, Intrinsic::nacl_atomic_store, PH.PET,
+                                      Args);
+}
+
+// %res = atomicrmw OP T* %ptr, T %val memory_order
+// becomes:
+// %res = call T @llvm.nacl.atomic.rmw.i<size>(OP, %ptr, %val, memory_order)
+void AtomicVisitor::visitAtomicRMWInst(AtomicRMWInst &I) {
+  NaCl::AtomicRMWOperation Op;
+  switch (I.getOperation()) {
+  default:
+    report_fatal_error("unsupported atomicrmw operation: " + ToTwine(I));
+  case AtomicRMWInst::Add: Op = NaCl::AtomicAdd; break;
+  case AtomicRMWInst::Sub: Op = NaCl::AtomicSub; break;
+  case AtomicRMWInst::And: Op = NaCl::AtomicAnd; break;
+  case AtomicRMWInst::Or:  Op = NaCl::AtomicOr;  break;
+  case AtomicRMWInst::Xor: Op = NaCl::AtomicXor; break;
+  case AtomicRMWInst::Xchg: Op = NaCl::AtomicExchange; break;
+  }
+  PointerHelper<AtomicRMWInst> PH(*this, I);
+  checkSizeMatchesType(I, PH.Size, I.getValOperand()->getType());
+  Value *Args[] = { ConstantInt::get(Type::getInt32Ty(C), Op), PH.P,
+                    I.getValOperand(), freezeMemoryOrder(I) };
+  replaceInstructionWithIntrinsicCall(I, Intrinsic::nacl_atomic_rmw, PH.PET,
+                                      Args);
+}
+
+// %res = cmpxchg T* %ptr, T %old, T %new memory_order
+// becomes:
+// %res = call T @llvm.nacl.atomic.cmpxchg.i<size>(
+//     %object, %expected, %desired, memory_order_success, memory_order_failure)
+void AtomicVisitor::visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
+  PointerHelper<AtomicCmpXchgInst> PH(*this, I);
+  checkSizeMatchesType(I, PH.Size, I.getCompareOperand()->getType());
+  checkSizeMatchesType(I, PH.Size, I.getNewValOperand()->getType());
+  // TODO LLVM currently doesn't support specifying separate memory
+  //      orders for compare exchange's success and failure cases: LLVM
+  //      IR implicitly drops the Release part of the specified memory
+  //      order on failure.
+  Value *Args[] = { PH.P, I.getCompareOperand(), I.getNewValOperand(),
+                    freezeMemoryOrder(I), freezeMemoryOrder(I) };
+  replaceInstructionWithIntrinsicCall(I, Intrinsic::nacl_atomic_cmpxchg, PH.PET,
+                                      Args);
+}
+
+// fence memory_order
+// becomes:
+// call void @llvm.nacl.atomic.fence(memory_order)
+void AtomicVisitor::visitFenceInst(FenceInst &I) {
+  Type *T = Type::getInt32Ty(C); // Fences aren't overloaded on type.
+  Value *Args[] = { freezeMemoryOrder(I) };
+  replaceInstructionWithIntrinsicCall(I, Intrinsic::nacl_atomic_fence, T, Args);
+}
+
+ModulePass *llvm::createRewriteAtomicsPass() { return new RewriteAtomics(); }