Subzero: Initial implementation of multithreaded translation.

Makefile.standalone

@@ -47,6 +47,8 @@ endif
 
 # The list of CXX defines that are dependent on build parameters.
 CXX_DEFINES =
+CXX_EXTRA =
+LD_EXTRA =
 
 ifdef MINIMAL
   NOASSERT = 1
@@ -65,6 +67,12 @@ else
   OBJDIR := $(OBJDIR)+Asserts
 endif
 
+ifdef TSAN
+  OBJDIR := $(OBJDIR)+TSan
+  CXX_EXTRA += -fsanitize=thread
+  LD_EXTRA += -fsanitize=thread
+endif
+
 $(info -----------------------------------------------)
 $(info Using LLVM_SRC_PATH = $(LLVM_SRC_PATH))
 $(info Using LLVM_BIN_PATH = $(LLVM_BIN_PATH))
@@ -85,8 +93,9 @@ CXX := CCACHE_CPP2=yes $(CCACHE) $(CLANG_PATH)/clang++
 CXXFLAGS := $(LLVM_CXXFLAGS) -std=c++11 -Wall -Wextra -Werror -fno-rtti \
 	-fno-exceptions $(OPTLEVEL) $(ASSERTIONS) $(CXX_DEFINES) -g \
 	$(HOST_FLAGS) -Wno-error=unused-parameter \
-	-I$(LIBCXX_INSTALL_PATH)/include/c++/v1
-LDFLAGS := $(HOST_FLAGS) -L$(LIBCXX_INSTALL_PATH)/lib -Wl,--gc-sections
+	-I$(LIBCXX_INSTALL_PATH)/include/c++/v1 $(CXX_EXTRA)
+LDFLAGS := $(HOST_FLAGS) -L$(LIBCXX_INSTALL_PATH)/lib -Wl,--gc-sections \
+	$(LD_EXTRA)
 
 SRCS = \
 	assembler.cpp \

src/IceCfg.cpp

@@ -32,10 +32,11 @@ ArenaAllocator<> *getCurrentCfgAllocator() {
 }
 
 Cfg::Cfg(GlobalContext *Ctx)
-    : Ctx(Ctx), FunctionName(""), ReturnType(IceType_void),
-      IsInternalLinkage(false), HasError(false), FocusedTiming(false),
-      ErrorMessage(""), Entry(nullptr), NextInstNumber(Inst::NumberInitial),
-      Allocator(new ArenaAllocator<>()), Live(nullptr),
+    : Ctx(Ctx), VMask(Ctx->getVerbose()), FunctionName(""),
+      ReturnType(IceType_void), IsInternalLinkage(false), HasError(false),
+      FocusedTiming(false), ErrorMessage(""), Entry(nullptr),
+      NextInstNumber(Inst::NumberInitial), Allocator(new ArenaAllocator<>()),
+      Live(nullptr),
       Target(TargetLowering::createLowering(Ctx->getTargetArch(), this)),
       VMetadata(new VariablesMetadata(this)),
       TargetAssembler(
@@ -46,18 +47,14 @@ Cfg::Cfg(GlobalContext *Ctx)
 }
 
 Cfg::~Cfg() {
-  // TODO(stichnot,kschimpf): Set CurrentCfg=nullptr in the dtor for
-  // safety.  This can't be done currently because the translator
-  // manages the Cfg by creating a new Cfg (which sets CurrentCfg to
-  // the new value), then deleting the old Cfg (which would then reset
-  // CurrentCfg to nullptr).
+  assert(ICE_TLS_GET_FIELD(CurrentCfg) == this);
+  // Reset the thread-local CurrentCfg pointer.
+  ICE_TLS_SET_FIELD(CurrentCfg, nullptr);
 }
 
 void Cfg::setError(const IceString &Message) {
   HasError = true;
   ErrorMessage = Message;
-  OstreamLocker L(Ctx);
-  Ctx->getStrDump() << "ICE translation error: " << ErrorMessage << "\n";
 }
 
 CfgNode *Cfg::makeNode() {
@@ -478,7 +475,7 @@ void Cfg::emitIAS() {
 void Cfg::dump(const IceString &Message) {
   if (!ALLOW_DUMP)
     return;
-  if (!Ctx->isVerbose())
+  if (!isVerbose())
     return;
   OstreamLocker L(Ctx);
   Ostream &Str = Ctx->getStrDump();
@@ -486,7 +483,7 @@ void Cfg::dump(const IceString &Message) {
     Str << "================ " << Message << " ================\n";
   setCurrentNode(getEntryNode());
   // Print function name+args
-  if (getContext()->isVerbose(IceV_Instructions)) {
+  if (isVerbose(IceV_Instructions)) {
     Str << "define ";
     if (getInternal() && !Ctx->getFlags().DisableInternal)
       Str << "internal ";
@@ -500,7 +497,7 @@ void Cfg::dump(const IceString &Message) {
     Str << ") {\n";
   }
   resetCurrentNode();
-  if (getContext()->isVerbose(IceV_Liveness)) {
+  if (isVerbose(IceV_Liveness)) {
     // Print summary info about variables
     for (Variable *Var : Variables) {
       Str << "// multiblock=";
@@ -516,7 +513,7 @@ void Cfg::dump(const IceString &Message) {
   // Print each basic block
   for (CfgNode *Node : Nodes)
     Node->dump(this);
-  if (getContext()->isVerbose(IceV_Instructions))
+  if (isVerbose(IceV_Instructions))
     Str << "}\n";
 }
 

src/IceCfg.h

@@ -15,8 +15,6 @@
 #ifndef SUBZERO_SRC_ICECFG_H
 #define SUBZERO_SRC_ICECFG_H
 
-#include <memory>
-
 #include "assembler.h"
 #include "IceClFlags.h"
 #include "IceDefs.h"
@@ -42,6 +40,7 @@ public:
   }
   // Gets a pointer to the current thread's Cfg.
   static const Cfg *getCurrentCfg() { return ICE_TLS_GET_FIELD(CurrentCfg); }
+  void updateTLS() const { ICE_TLS_SET_FIELD(CurrentCfg, this); }
   // Gets a pointer to the current thread's Cfg's allocator.
   static ArenaAllocator<> *getCurrentCfgAllocator() {
     assert(ICE_TLS_GET_FIELD(CurrentCfg));
@@ -50,6 +49,12 @@ public:
 
   GlobalContext *getContext() const { return Ctx; }
 
+  // Returns true if any of the specified options in the verbose mask
+  // are set.  If the argument is omitted, it checks if any verbose
+  // options at all are set.
+  bool isVerbose(VerboseMask Mask = IceV_All) const { return VMask & Mask; }
+  void setVerbose(VerboseMask Mask) { VMask = Mask; }
+
   // Manage the name and return type of the function being translated.
   void setFunctionName(const IceString &Name) { FunctionName = Name; }
   IceString getFunctionName() const { return FunctionName; }
@@ -184,6 +189,7 @@ private:
   Cfg(GlobalContext *Ctx);
 
   GlobalContext *Ctx;
+  VerboseMask VMask;
   IceString FunctionName;
   Type ReturnType;
   bool IsInternalLinkage;

src/IceCfgNode.cpp

@@ -917,11 +917,11 @@ void CfgNode::dump(Cfg *Func) const {
   Func->setCurrentNode(this);
   Ostream &Str = Func->getContext()->getStrDump();
   Liveness *Liveness = Func->getLiveness();
-  if (Func->getContext()->isVerbose(IceV_Instructions)) {
+  if (Func->isVerbose(IceV_Instructions)) {
     Str << getName() << ":\n";
   }
   // Dump list of predecessor nodes.
-  if (Func->getContext()->isVerbose(IceV_Preds) && !InEdges.empty()) {
+  if (Func->isVerbose(IceV_Preds) && !InEdges.empty()) {
     Str << "    // preds = ";
     bool First = true;
     for (CfgNode *I : InEdges) {
@@ -936,13 +936,13 @@ void CfgNode::dump(Cfg *Func) const {
   LivenessBV LiveIn;
   if (Liveness)
     LiveIn = Liveness->getLiveIn(this);
-  if (Func->getContext()->isVerbose(IceV_Liveness) && !LiveIn.empty()) {
+  if (Func->isVerbose(IceV_Liveness) && !LiveIn.empty()) {
     Str << "    // LiveIn:";
     for (SizeT i = 0; i < LiveIn.size(); ++i) {
       if (LiveIn[i]) {
         Variable *Var = Liveness->getVariable(i, this);
         Str << " %" << Var->getName(Func);
-        if (Func->getContext()->isVerbose(IceV_RegOrigins) && Var->hasReg()) {
+        if (Func->isVerbose(IceV_RegOrigins) && Var->hasReg()) {
           Str << ":" << Func->getTarget()->getRegName(Var->getRegNum(),
                                                       Var->getType());
         }
@@ -951,7 +951,7 @@ void CfgNode::dump(Cfg *Func) const {
     Str << "\n";
   }
   // Dump each instruction.
-  if (Func->getContext()->isVerbose(IceV_Instructions)) {
+  if (Func->isVerbose(IceV_Instructions)) {
     for (const Inst &I : Phis)
       I.dumpDecorated(Func);
     for (const Inst &I : Insts)
@@ -961,13 +961,13 @@ void CfgNode::dump(Cfg *Func) const {
   LivenessBV LiveOut;
   if (Liveness)
     LiveOut = Liveness->getLiveOut(this);
-  if (Func->getContext()->isVerbose(IceV_Liveness) && !LiveOut.empty()) {
+  if (Func->isVerbose(IceV_Liveness) && !LiveOut.empty()) {
     Str << "    // LiveOut:";
     for (SizeT i = 0; i < LiveOut.size(); ++i) {
       if (LiveOut[i]) {
         Variable *Var = Liveness->getVariable(i, this);
         Str << " %" << Var->getName(Func);
-        if (Func->getContext()->isVerbose(IceV_RegOrigins) && Var->hasReg()) {
+        if (Func->isVerbose(IceV_RegOrigins) && Var->hasReg()) {
           Str << ":" << Func->getTarget()->getRegName(Var->getRegNum(),
                                                       Var->getType());
         }
@@ -976,7 +976,7 @@ void CfgNode::dump(Cfg *Func) const {
     Str << "\n";
   }
   // Dump list of successor nodes.
-  if (Func->getContext()->isVerbose(IceV_Succs)) {
+  if (Func->isVerbose(IceV_Succs)) {
     Str << "    // succs = ";
     bool First = true;
     for (CfgNode *I : OutEdges) {

src/IceClFlags.h

@@ -28,8 +28,9 @@ public:
         DumpStats(false), AllowUninitializedGlobals(false),
         TimeEachFunction(false), DisableIRGeneration(false),
         AllowErrorRecovery(false), GenerateUnitTestMessages(false),
-        DefaultGlobalPrefix(""), DefaultFunctionPrefix(""), TimingFocusOn(""),
-        VerboseFocusOn(""), TranslateOnly("") {}
+        NumTranslationThreads(0), DefaultGlobalPrefix(""),
+        DefaultFunctionPrefix(""), TimingFocusOn(""), VerboseFocusOn(""),
+        TranslateOnly("") {}
   bool DisableInternal;
   bool SubzeroTimingEnabled;
   bool DisableTranslation;
@@ -46,6 +47,7 @@ public:
   bool DisableIRGeneration;
   bool AllowErrorRecovery;
   bool GenerateUnitTestMessages;
+  size_t NumTranslationThreads; // 0 means completely sequential
   IceString DefaultGlobalPrefix;
   IceString DefaultFunctionPrefix;
   IceString TimingFocusOn;

src/IceConverter.cpp

@@ -891,8 +891,6 @@ void Converter::convertFunctions() {
     if (ALLOW_DUMP && Ctx->getFlags().TimeEachFunction)
       Ctx->popTimer(TimerID, StackID);
   }
-
-  emitConstants();
 }
 
 } // end of namespace Ice

src/IceDefs.h

@@ -23,7 +23,9 @@
 #include <limits>
 #include <list>
 #include <map>
+#include <mutex>
 #include <string>
+#include <system_error>
 #include <vector>
 #include "llvm/ADT/ArrayRef.h"
 #include "llvm/ADT/BitVector.h"
@@ -58,6 +60,7 @@ class InstTarget;
 class LiveRange;
 class Liveness;
 class Operand;
+class TargetGlobalLowering;
 class TargetLowering;
 class Variable;
 class VariableDeclaration;
@@ -120,6 +123,17 @@ typedef llvm::BitVector LivenessBV;
 typedef uint32_t TimerStackIdT;
 typedef uint32_t TimerIdT;
 
+// Use alignas(MaxCacheLineSize) to isolate variables/fields that
+// might be contended while multithreading.  Assumes the maximum cache
+// line size is 64.
+enum {
+  MaxCacheLineSize = 64
+};
+// Use ICE_CACHELINE_BOUNDARY to force the next field in a declaration
+// list to be aligned to the next cache line.
+#define ICE_CACHELINE_BOUNDARY                                                 \
+  alignas(MaxCacheLineSize) struct {}
+
 // PNaCl is ILP32, so theoretically we should only need 32-bit offsets.
 typedef int32_t RelocOffsetT;
 enum { RelocAddrSize = 4 };
@@ -163,6 +177,37 @@ typedef uint32_t VerboseMask;
 typedef llvm::raw_ostream Ostream;
 typedef llvm::raw_fd_ostream Fdstream;
 
+typedef std::mutex GlobalLockType;
+
+enum ErrorCodes {
+  EC_None = 0,
+  EC_Args,
+  EC_Bitcode,
+  EC_Translation
+};
+
+// Wrapper around std::error_code for allowing multiple errors to be
+// folded into one.  The current implementation keeps track of the
+// first error, which is likely to be the most useful one, and this
+// could be extended to e.g. collect a vector of errors.
+class ErrorCode : public std::error_code {
+  ErrorCode(const ErrorCode &) = delete;
+  ErrorCode &operator=(const ErrorCode &) = delete;
+
+public:
+  ErrorCode() : HasError(false) {}
+  void assign(ErrorCodes Code) {
+    if (!HasError) {
+      HasError = true;
+      std::error_code::assign(Code, std::generic_category());
+    }
+  }
+  void assign(int Code) { assign(static_cast<ErrorCodes>(Code)); }
+
+private:
+  bool HasError;
+};
+
 // Reverse range adaptors written in terms of llvm::make_range().
 template <typename T>
 llvm::iterator_range<typename T::const_reverse_iterator>

src/IceGlobalContext.cpp

@@ -129,16 +129,17 @@ GlobalContext::GlobalContext(Ostream *OsDump, Ostream *OsEmit,
                              ELFStreamer *ELFStr, VerboseMask Mask,
                              TargetArch Arch, OptLevel Opt,
                              IceString TestPrefix, const ClFlags &Flags)
-    : StrDump(OsDump), StrEmit(OsEmit), VMask(Mask),
-      ConstPool(new ConstantPool()), Arch(Arch), Opt(Opt),
-      TestPrefix(TestPrefix), Flags(Flags), RNG(""), ObjectWriter() {
+    : ConstPool(new ConstantPool()), ErrorStatus(), StrDump(OsDump),
+      StrEmit(OsEmit), VMask(Mask), Arch(Arch), Opt(Opt),
+      TestPrefix(TestPrefix), Flags(Flags), RNG(""), ObjectWriter(),
+      CfgQ(/*MaxSize=*/Flags.NumTranslationThreads,
+           /*Sequential=*/(Flags.NumTranslationThreads == 0)) {
   // Make sure thread_local fields are properly initialized before any
   // accesses are made.  Do this here instead of at the start of
   // main() so that all clients (e.g. unit tests) can benefit for
   // free.
   GlobalContext::TlsInit();
   Cfg::TlsInit();
-
   // Create a new ThreadContext for the current thread.  No need to
   // lock AllThreadContexts at this point since no other threads have
   // access yet to this GlobalContext object.
@@ -156,6 +157,43 @@ GlobalContext::GlobalContext(Ostream *OsDump, Ostream *OsEmit,
   }
 }
 
+void GlobalContext::translateFunctions() {
+  while (Cfg *Func = cfgQueueBlockingPop()) {
+    // Reset per-function stats being accumulated in TLS.
+    resetStats();
+    // Install Func in TLS for Cfg-specific container allocators.
+    Func->updateTLS();
+    // Set verbose level to none if the current function does NOT
+    // match the -verbose-focus command-line option.
+    if (!matchSymbolName(Func->getFunctionName(), getFlags().VerboseFocusOn))
+      Func->setVerbose(IceV_None);
+    // Disable translation if -notranslate is specified, or if the
+    // current function matches the -translate-only option.  If
+    // translation is disabled, just dump the high-level IR and
+    // continue.
+    if (getFlags().DisableTranslation ||
+        !matchSymbolName(Func->getFunctionName(), getFlags().TranslateOnly)) {
+      Func->dump();
+    } else {
+      Func->translate();
+      if (Func->hasError()) {
+        getErrorStatus()->assign(EC_Translation);
+        OstreamLocker L(this);
+        getStrDump() << "ICE translation error: " << Func->getError() << "\n";
+      } else {
+        if (getFlags().UseIntegratedAssembler)
+          Func->emitIAS();
+        else
+          Func->emit();
+        // TODO(stichnot): actually add to emit queue
+      }
+      // TODO(stichnot): fix multithreaded stats dumping.
+      dumpStats(Func->getFunctionName());
+    }
+    delete Func;
+  }
+}
+
 // Scan a string for S[0-9A-Z]*_ patterns and replace them with
 // S<num>_ where <num> is the next base-36 value.  If a type name
 // legitimately contains that pattern, then the substitution will be
@@ -471,12 +509,18 @@ TimerIdT GlobalContext::getTimerID(TimerStackIdT StackID,
 }
 
 void GlobalContext::pushTimer(TimerIdT ID, TimerStackIdT StackID) {
+  // TODO(stichnot): Timers are completely broken for multithreading; fix.
+  if (getFlags().NumTranslationThreads)
+    llvm::report_fatal_error("Timers and multithreading are currently broken");
   auto Timers = getTimers();
   assert(StackID < Timers->size());
   Timers->at(StackID).push(ID);
 }
 
 void GlobalContext::popTimer(TimerIdT ID, TimerStackIdT StackID) {
+  // TODO(stichnot): Timers are completely broken for multithreading; fix.
+  if (getFlags().NumTranslationThreads)
+    llvm::report_fatal_error("Timers and multithreading are currently broken");
   auto Timers = getTimers();
   assert(StackID < Timers->size());
   Timers->at(StackID).pop(ID);

src/IceGlobalContext.h

@@ -15,8 +15,8 @@
 #ifndef SUBZERO_SRC_ICEGLOBALCONTEXT_H
 #define SUBZERO_SRC_ICEGLOBALCONTEXT_H
 
-#include <memory>
 #include <mutex>
+#include <thread>
 
 #include "IceDefs.h"
 #include "IceClFlags.h"
@@ -24,6 +24,7 @@
 #include "IceRNG.h"
 #include "IceTimerTree.h"
 #include "IceTypes.h"
+#include "IceUtils.h"
 
 namespace Ice {
 
@@ -31,8 +32,6 @@ class ClFlags;
 class ConstantPool;
 class FuncSigType;
 
-typedef std::mutex GlobalLockType;
-
 // LockedPtr is a way to provide automatically locked access to some object.
 template <typename T> class LockedPtr {
   LockedPtr() = delete;
@@ -102,14 +101,7 @@ public:
                 IceString TestPrefix, const ClFlags &Flags);
   ~GlobalContext();
 
-  // Returns true if any of the specified options in the verbose mask
-  // are set.  If the argument is omitted, it checks if any verbose
-  // options at all are set.
   VerboseMask getVerbose() const { return VMask; }
-  bool isVerbose(VerboseMask Mask = IceV_All) const { return VMask & Mask; }
-  void setVerbose(VerboseMask Mask) { VMask = Mask; }
-  void addVerbose(VerboseMask Mask) { VMask |= Mask; }
-  void subVerbose(VerboseMask Mask) { VMask &= ~Mask; }
 
   // The dump and emit streams need to be used by only one thread at a
   // time.  This is done by exclusively reserving the streams via
@@ -129,6 +121,9 @@ public:
 
   TargetArch getTargetArch() const { return Arch; }
   OptLevel getOptLevel() const { return Opt; }
+  LockedPtr<ErrorCode> getErrorStatus() {
+    return LockedPtr<ErrorCode>(&ErrorStatus, &ErrorStatusLock);
+  }
 
   // When emitting assembly, we allow a string to be prepended to
   // names of translated functions.  This makes it easier to create an
@@ -229,34 +224,107 @@ public:
   void dumpTimers(TimerStackIdT StackID = TSK_Default,
                   bool DumpCumulative = true);
 
+  // Adds a newly parsed and constructed function to the Cfg work
+  // queue.  Notifies any idle workers that a new function is
+  // available for translating.  May block if the work queue is too
+  // large, in order to control memory footprint.
+  void cfgQueueBlockingPush(Cfg *Func) { CfgQ.blockingPush(Func); }
+  // Takes a Cfg from the work queue for translating.  May block if
+  // the work queue is currently empty.  Returns nullptr if there is
+  // no more work - the queue is empty and either end() has been
+  // called or the Sequential flag was set.
+  Cfg *cfgQueueBlockingPop() { return CfgQ.blockingPop(); }
+  // Notifies that no more work will be added to the work queue.
+  void cfgQueueNotifyEnd() { CfgQ.notifyEnd(); }
+
+  void startWorkerThreads() {
+    size_t NumWorkers = getFlags().NumTranslationThreads;
+    for (size_t i = 0; i < NumWorkers; ++i) {
+      ThreadContext *WorkerTLS = new ThreadContext();
+      AllThreadContexts.push_back(WorkerTLS);
+      TranslationThreads.push_back(std::thread(
+          &GlobalContext::translateFunctionsWrapper, this, WorkerTLS));
+    }
+    if (NumWorkers) {
+      // TODO(stichnot): start a new thread for the emitter queue worker.
+    }
+  }
+
+  void waitForWorkerThreads() {
+    cfgQueueNotifyEnd();
+    // TODO(stichnot): call end() on the emitter work queue.
+    for (std::thread &Worker : TranslationThreads) {
+      Worker.join();
+    }
+    TranslationThreads.clear();
+    // TODO(stichnot): join the emitter thread.
+  }
+
+  // Translation thread startup routine.
+  void translateFunctionsWrapper(ThreadContext *MyTLS) {
+    ICE_TLS_SET_FIELD(TLS, MyTLS);
+    translateFunctions();
+  }
+  // Translate functions from the Cfg queue until the queue is empty.
+  void translateFunctions();
+
+  // Utility function to match a symbol name against a match string.
+  // This is used in a few cases where we want to take some action on
+  // a particular function or symbol based on a command-line argument,
+  // such as changing the verbose level for a particular function.  An
+  // empty Match argument means match everything.  Returns true if
+  // there is a match.
+  static bool matchSymbolName(const IceString &SymbolName,
+                              const IceString &Match) {
+    return Match.empty() || Match == SymbolName;
+  }
+
 private:
-  // Try to make sure the mutexes are allocated on separate cache
-  // lines, assuming the maximum cache line size is 64.
-  const static size_t MaxCacheLineSize = 64;
-  alignas(MaxCacheLineSize) GlobalLockType AllocLock;
-  alignas(MaxCacheLineSize) GlobalLockType ConstPoolLock;
-  alignas(MaxCacheLineSize) GlobalLockType StatsLock;
-  alignas(MaxCacheLineSize) GlobalLockType TimerLock;
+  // Try to ensure mutexes are allocated on separate cache lines.
+
+  ICE_CACHELINE_BOUNDARY;
+  // Managed by getAllocator()
+  GlobalLockType AllocLock;
+  ArenaAllocator<> Allocator;
 
+  ICE_CACHELINE_BOUNDARY;
+  // Managed by getConstantPool()
+  GlobalLockType ConstPoolLock;
+  std::unique_ptr<ConstantPool> ConstPool;
+
+  ICE_CACHELINE_BOUNDARY;
+  // Managed by getErrorStatus()
+  GlobalLockType ErrorStatusLock;
+  ErrorCode ErrorStatus;
+
+  ICE_CACHELINE_BOUNDARY;
+  // Managed by getStatsCumulative()
+  GlobalLockType StatsLock;
+  CodeStats StatsCumulative;
+
+  ICE_CACHELINE_BOUNDARY;
+  // Managed by getTimers()
+  GlobalLockType TimerLock;
+  std::vector<TimerStack> Timers;
+
+  ICE_CACHELINE_BOUNDARY;
   // StrLock is a global lock on the dump and emit output streams.
   typedef std::mutex StrLockType;
   StrLockType StrLock;
-
   Ostream *StrDump; // Stream for dumping / diagnostics
   Ostream *StrEmit; // Stream for code emission
 
-  ArenaAllocator<> Allocator;
-  VerboseMask VMask;
-  std::unique_ptr<ConstantPool> ConstPool;
+  ICE_CACHELINE_BOUNDARY;
+
+  const VerboseMask VMask;
   Intrinsics IntrinsicsInfo;
   const TargetArch Arch;
   const OptLevel Opt;
   const IceString TestPrefix;
   const ClFlags &Flags;
-  RandomNumberGenerator RNG;
+  RandomNumberGenerator RNG; // TODO(stichnot): Move into Cfg.
   std::unique_ptr<ELFObjectWriter> ObjectWriter;
-  CodeStats StatsCumulative;
-  std::vector<TimerStack> Timers;
+  BoundedProducerConsumerQueue<Cfg> CfgQ;
 
   LockedPtr<ArenaAllocator<>> getAllocator() {
     return LockedPtr<ArenaAllocator<>>(&Allocator, &AllocLock);
@@ -272,6 +340,7 @@ private:
   }
 
   std::vector<ThreadContext *> AllThreadContexts;
+  std::vector<std::thread> TranslationThreads;
   // Each thread has its own TLS pointer which is also held in
   // AllThreadContexts.
   ICE_TLS_DECLARE_FIELD(ThreadContext *, TLS);

src/IceInst.cpp

@@ -464,10 +464,9 @@ void Inst::dumpDecorated(const Cfg *Func) const {
   if (!ALLOW_DUMP)
     return;
   Ostream &Str = Func->getContext()->getStrDump();
-  if (!Func->getContext()->isVerbose(IceV_Deleted) &&
-      (isDeleted() || isRedundantAssign()))
+  if (!Func->isVerbose(IceV_Deleted) && (isDeleted() || isRedundantAssign()))
     return;
-  if (Func->getContext()->isVerbose(IceV_InstNumbers)) {
+  if (Func->isVerbose(IceV_InstNumbers)) {
     char buf[30];
     InstNumberT Number = getNumber();
     if (Number == NumberDeleted)
@@ -500,7 +499,7 @@ void Inst::dumpExtras(const Cfg *Func) const {
   bool First = true;
   // Print "LIVEEND={a,b,c}" for all source operands whose live ranges
   // are known to end at this instruction.
-  if (Func->getContext()->isVerbose(IceV_Liveness)) {
+  if (Func->isVerbose(IceV_Liveness)) {
     for (SizeT I = 0; I < getSrcSize(); ++I) {
       Operand *Src = getSrc(I);
       SizeT NumVars = Src->getNumVars();

src/IceOperand.cpp

@@ -399,15 +399,15 @@ void Variable::dump(const Cfg *Func, Ostream &Str) const {
     Str << "%" << getName(Func);
     return;
   }
-  if (Func->getContext()->isVerbose(IceV_RegOrigins) ||
+  if (Func->isVerbose(IceV_RegOrigins) ||
       (!hasReg() && !Func->getTarget()->hasComputedFrame()))
     Str << "%" << getName(Func);
   if (hasReg()) {
-    if (Func->getContext()->isVerbose(IceV_RegOrigins))
+    if (Func->isVerbose(IceV_RegOrigins))
       Str << ":";
     Str << Func->getTarget()->getRegName(RegNum, getType());
   } else if (Func->getTarget()->hasComputedFrame()) {
-    if (Func->getContext()->isVerbose(IceV_RegOrigins))
+    if (Func->isVerbose(IceV_RegOrigins))
       Str << ":";
     Str << "[" << Func->getTarget()->getRegName(
                       Func->getTarget()->getFrameOrStackReg(), IceType_i32);

src/IceRegAlloc.cpp

@@ -52,7 +52,7 @@ void dumpDisableOverlap(const Cfg *Func, const Variable *Var,
                         const char *Reason) {
   if (!ALLOW_DUMP)
     return;
-  if (Func->getContext()->isVerbose(IceV_LinearScan)) {
+  if (Func->isVerbose(IceV_LinearScan)) {
     VariablesMetadata *VMetadata = Func->getVMetadata();
     Ostream &Str = Func->getContext()->getStrDump();
     Str << "Disabling Overlap due to " << Reason << " " << *Var
@@ -265,8 +265,7 @@ void LinearScan::scan(const llvm::SmallBitVector &RegMaskFull,
   TimerMarker T(TimerStack::TT_linearScan, Func);
   assert(RegMaskFull.any()); // Sanity check
   GlobalContext *Ctx = Func->getContext();
-  const bool Verbose =
-      ALLOW_DUMP && Ctx->isVerbose(IceV_LinearScan);
+  const bool Verbose = ALLOW_DUMP && Func->isVerbose(IceV_LinearScan);
   if (Verbose)
     Ctx->lockStr();
   Func->resetCurrentNode();
@@ -738,7 +737,7 @@ void LinearScan::scan(const llvm::SmallBitVector &RegMaskFull,
 void LinearScan::dump(Cfg *Func) const {
   if (!ALLOW_DUMP)
     return;
-  if (!Func->getContext()->isVerbose(IceV_LinearScan))
+  if (!Func->isVerbose(IceV_LinearScan))
     return;
   Ostream &Str = Func->getContext()->getStrDump();
   Func->resetCurrentNode();

src/IceTargetLowering.cpp

@@ -252,25 +252,24 @@ void TargetLowering::regAlloc(RegAllocKind Kind) {
   LinearScan.scan(RegMask, RandomizeRegisterAllocation);
 }
 
-TargetGlobalInitLowering *
-TargetGlobalInitLowering::createLowering(TargetArch Target,
-                                         GlobalContext *Ctx) {
+TargetGlobalLowering *TargetGlobalLowering::createLowering(GlobalContext *Ctx) {
   // These statements can be #ifdef'd to specialize the code generator
   // to a subset of the available targets.  TODO: use CRTP.
+  TargetArch Target = Ctx->getTargetArch();
   if (Target == Target_X8632)
-    return TargetGlobalInitX8632::create(Ctx);
+    return TargetGlobalX8632::create(Ctx);
 #if 0
   if (Target == Target_X8664)
-    return IceTargetGlobalInitX8664::create(Ctx);
+    return TargetGlobalX8664::create(Ctx);
   if (Target == Target_ARM32)
-    return IceTargetGlobalInitARM32::create(Ctx);
+    return TargetGlobalARM32::create(Ctx);
   if (Target == Target_ARM64)
-    return IceTargetGlobalInitARM64::create(Ctx);
+    return TargetGlobalARM64::create(Ctx);
 #endif
   llvm_unreachable("Unsupported target");
   return nullptr;
 }
 
-TargetGlobalInitLowering::~TargetGlobalInitLowering() {}
+TargetGlobalLowering::~TargetGlobalLowering() {}
 
 } // end of namespace Ice

src/IceTargetLowering.h

@@ -199,8 +199,6 @@ public:
   virtual void addProlog(CfgNode *Node) = 0;
   virtual void addEpilog(CfgNode *Node) = 0;
 
-  virtual void emitConstants() const = 0;
-
   virtual ~TargetLowering() {}
 
 protected:
@@ -242,23 +240,23 @@ protected:
   LoweringContext Context;
 };
 
-// TargetGlobalInitLowering is used for "lowering" global
-// initializers.  It is separated out from TargetLowering because it
-// does not require a Cfg.
-class TargetGlobalInitLowering {
-  TargetGlobalInitLowering(const TargetGlobalInitLowering &) = delete;
-  TargetGlobalInitLowering &operator=(const TargetGlobalInitLowering &) =
-      delete;
+// TargetGlobalLowering is used for "lowering" global initializers,
+// including the internal constant pool.  It is separated out from
+// TargetLowering because it does not require a Cfg.
+class TargetGlobalLowering {
+  TargetGlobalLowering() = delete;
+  TargetGlobalLowering(const TargetGlobalLowering &) = delete;
+  TargetGlobalLowering &operator=(const TargetGlobalLowering &) = delete;
 
 public:
-  static TargetGlobalInitLowering *createLowering(TargetArch Target,
-                                                  GlobalContext *Ctx);
-  virtual ~TargetGlobalInitLowering();
+  static TargetGlobalLowering *createLowering(GlobalContext *Ctx);
+  virtual ~TargetGlobalLowering();
 
-  virtual void lower(const VariableDeclaration &Var) = 0;
+  virtual void lowerInit(const VariableDeclaration &Var) const = 0;
+  virtual void lowerConstants(GlobalContext *Ctx) const = 0;
 
 protected:
-  TargetGlobalInitLowering(GlobalContext *Ctx) : Ctx(Ctx) {}
+  TargetGlobalLowering(GlobalContext *Ctx) : Ctx(Ctx) {}
   GlobalContext *Ctx;
 };
 

src/IceTargetLoweringX8632.cpp

@@ -897,7 +897,7 @@ void TargetX8632::addProlog(CfgNode *Node) {
     Var->setStackOffset(Linked->getStackOffset());
   }
 
-  if (ALLOW_DUMP && Func->getContext()->isVerbose(IceV_Frame)) {
+  if (ALLOW_DUMP && Func->isVerbose(IceV_Frame)) {
     OstreamLocker L(Func->getContext());
     Ostream &Str = Func->getContext()->getStrDump();
 
@@ -966,75 +966,6 @@ void TargetX8632::addEpilog(CfgNode *Node) {
   }
 }
 
-template <typename T> struct PoolTypeConverter {};
-
-template <> struct PoolTypeConverter<float> {
-  typedef uint32_t PrimitiveIntType;
-  typedef ConstantFloat IceType;
-  static const Type Ty = IceType_f32;
-  static const char *TypeName;
-  static const char *AsmTag;
-  static const char *PrintfString;
-};
-const char *PoolTypeConverter<float>::TypeName = "float";
-const char *PoolTypeConverter<float>::AsmTag = ".long";
-const char *PoolTypeConverter<float>::PrintfString = "0x%x";
-
-template <> struct PoolTypeConverter<double> {
-  typedef uint64_t PrimitiveIntType;
-  typedef ConstantDouble IceType;
-  static const Type Ty = IceType_f64;
-  static const char *TypeName;
-  static const char *AsmTag;
-  static const char *PrintfString;
-};
-const char *PoolTypeConverter<double>::TypeName = "double";
-const char *PoolTypeConverter<double>::AsmTag = ".quad";
-const char *PoolTypeConverter<double>::PrintfString = "0x%llx";
-
-template <typename T> void TargetX8632::emitConstantPool() const {
-  // Note: Still used by emit IAS.
-  Ostream &Str = Ctx->getStrEmit();
-  Type Ty = T::Ty;
-  SizeT Align = typeAlignInBytes(Ty);
-  ConstantList Pool = Ctx->getConstantPool(Ty);
-
-  Str << "\t.section\t.rodata.cst" << Align << ",\"aM\",@progbits," << Align
-      << "\n";
-  Str << "\t.align\t" << Align << "\n";
-  for (Constant *C : Pool) {
-    typename T::IceType *Const = llvm::cast<typename T::IceType>(C);
-    typename T::IceType::PrimType Value = Const->getValue();
-    // Use memcpy() to copy bits from Value into RawValue in a way
-    // that avoids breaking strict-aliasing rules.
-    typename T::PrimitiveIntType RawValue;
-    memcpy(&RawValue, &Value, sizeof(Value));
-    char buf[30];
-    int CharsPrinted =
-        snprintf(buf, llvm::array_lengthof(buf), T::PrintfString, RawValue);
-    assert(CharsPrinted >= 0 &&
-           (size_t)CharsPrinted < llvm::array_lengthof(buf));
-    (void)CharsPrinted; // avoid warnings if asserts are disabled
-    Const->emitPoolLabel(Str);
-    Str << ":\n\t" << T::AsmTag << "\t" << buf << "\t# " << T::TypeName << " "
-        << Value << "\n";
-  }
-}
-
-void TargetX8632::emitConstants() const {
-  // No need to emit constants from the int pool since (for x86) they
-  // are embedded as immediates in the instructions, just emit float/double.
-  if (Ctx->getFlags().UseELFWriter) {
-    ELFObjectWriter *Writer = Ctx->getObjectWriter();
-    Writer->writeConstantPool<ConstantFloat>(IceType_f32);
-    Writer->writeConstantPool<ConstantDouble>(IceType_f64);
-  } else {
-    OstreamLocker L(Ctx);
-    emitConstantPool<PoolTypeConverter<float>>();
-    emitConstantPool<PoolTypeConverter<double>>();
-  }
-}
-
 void TargetX8632::split64(Variable *Var) {
   switch (Var->getType()) {
   default:
@@ -3567,7 +3498,7 @@ void dumpAddressOpt(const Cfg *Func, const Variable *Base,
                     const Inst *Reason) {
   if (!ALLOW_DUMP)
     return;
-  if (!Func->getContext()->isVerbose(IceV_AddrOpt))
+  if (!Func->isVerbose(IceV_AddrOpt))
     return;
   OstreamLocker L(Func->getContext());
   Ostream &Str = Func->getContext()->getStrDump();
@@ -3740,7 +3671,7 @@ bool matchOffsetBase(const VariablesMetadata *VMetadata, Variable *&Base,
 void computeAddressOpt(Cfg *Func, const Inst *Instr, Variable *&Base,
                        Variable *&Index, uint16_t &Shift, int32_t &Offset) {
   Func->resetCurrentNode();
-  if (Func->getContext()->isVerbose(IceV_AddrOpt)) {
+  if (Func->isVerbose(IceV_AddrOpt)) {
     OstreamLocker L(Func->getContext());
     Ostream &Str = Func->getContext()->getStrDump();
     Str << "\nStarting computeAddressOpt for instruction:\n  ";
@@ -4582,7 +4513,7 @@ void TargetX8632::makeRandomRegisterPermutation(
 
   assert(NumShuffled + NumPreserved == RegX8632::Reg_NUM);
 
-  if (Func->getContext()->isVerbose(IceV_Random)) {
+  if (Func->isVerbose(IceV_Random)) {
     OstreamLocker L(Func->getContext());
     Ostream &Str = Func->getContext()->getStrDump();
     Str << "Register equivalence classes:\n";
@@ -4630,10 +4561,10 @@ void ConstantUndef::emit(GlobalContext *) const {
   llvm_unreachable("undef value encountered by emitter.");
 }
 
-TargetGlobalInitX8632::TargetGlobalInitX8632(GlobalContext *Ctx)
-    : TargetGlobalInitLowering(Ctx) {}
+TargetGlobalX8632::TargetGlobalX8632(GlobalContext *Ctx)
+    : TargetGlobalLowering(Ctx) {}
 
-void TargetGlobalInitX8632::lower(const VariableDeclaration &Var) {
+void TargetGlobalX8632::lowerInit(const VariableDeclaration &Var) const {
   // TODO(jvoung): handle this without text.
   if (Ctx->getFlags().UseELFWriter)
     return;
@@ -4714,4 +4645,76 @@ void TargetGlobalInitX8632::lower(const VariableDeclaration &Var) {
   Str << "\t.size\t" << MangledName << ", " << Size << "\n";
 }
 
+template <typename T> struct PoolTypeConverter {};
+
+template <> struct PoolTypeConverter<float> {
+  typedef uint32_t PrimitiveIntType;
+  typedef ConstantFloat IceType;
+  static const Type Ty = IceType_f32;
+  static const char *TypeName;
+  static const char *AsmTag;
+  static const char *PrintfString;
+};
+const char *PoolTypeConverter<float>::TypeName = "float";
+const char *PoolTypeConverter<float>::AsmTag = ".long";
+const char *PoolTypeConverter<float>::PrintfString = "0x%x";
+
+template <> struct PoolTypeConverter<double> {
+  typedef uint64_t PrimitiveIntType;
+  typedef ConstantDouble IceType;
+  static const Type Ty = IceType_f64;
+  static const char *TypeName;
+  static const char *AsmTag;
+  static const char *PrintfString;
+};
+const char *PoolTypeConverter<double>::TypeName = "double";
+const char *PoolTypeConverter<double>::AsmTag = ".quad";
+const char *PoolTypeConverter<double>::PrintfString = "0x%llx";
+
+template <typename T>
+void TargetGlobalX8632::emitConstantPool(GlobalContext *Ctx) {
+  // Note: Still used by emit IAS.
+  Ostream &Str = Ctx->getStrEmit();
+  Type Ty = T::Ty;
+  SizeT Align = typeAlignInBytes(Ty);
+  ConstantList Pool = Ctx->getConstantPool(Ty);
+
+  Str << "\t.section\t.rodata.cst" << Align << ",\"aM\",@progbits," << Align
+      << "\n";
+  Str << "\t.align\t" << Align << "\n";
+  for (Constant *C : Pool) {
+    typename T::IceType *Const = llvm::cast<typename T::IceType>(C);
+    typename T::IceType::PrimType Value = Const->getValue();
+    // Use memcpy() to copy bits from Value into RawValue in a way
+    // that avoids breaking strict-aliasing rules.
+    typename T::PrimitiveIntType RawValue;
+    memcpy(&RawValue, &Value, sizeof(Value));
+    char buf[30];
+    int CharsPrinted =
+        snprintf(buf, llvm::array_lengthof(buf), T::PrintfString, RawValue);
+    assert(CharsPrinted >= 0 &&
+           (size_t)CharsPrinted < llvm::array_lengthof(buf));
+    (void)CharsPrinted; // avoid warnings if asserts are disabled
+    Const->emitPoolLabel(Str);
+    Str << ":\n\t" << T::AsmTag << "\t" << buf << "\t# " << T::TypeName << " "
+        << Value << "\n";
+  }
+}
+
+void TargetGlobalX8632::lowerConstants(GlobalContext *Ctx) const {
+  if (Ctx->getFlags().DisableTranslation)
+    return;
+  // No need to emit constants from the int pool since (for x86) they
+  // are embedded as immediates in the instructions, just emit float/double.
+  if (Ctx->getFlags().UseELFWriter) {
+    ELFObjectWriter *Writer = Ctx->getObjectWriter();
+    Writer->writeConstantPool<ConstantFloat>(IceType_f32);
+    Writer->writeConstantPool<ConstantDouble>(IceType_f64);
+  } else {
+    OstreamLocker L(Ctx);
+    emitConstantPool<PoolTypeConverter<float>>(Ctx);
+    emitConstantPool<PoolTypeConverter<double>>(Ctx);
+  }
+}
+
 } // end of namespace Ice

src/IceTargetLoweringX8632.h

@@ -56,7 +56,6 @@ public:
   void lowerArguments() override;
   void addProlog(CfgNode *Node) override;
   void addEpilog(CfgNode *Node) override;
-  void emitConstants() const override;
   SizeT makeNextLabelNumber() { return NextLabelNumber++; }
   // Ensure that a 64-bit Variable has been split into 2 32-bit
   // Variables, creating them if necessary.  This is needed for all
@@ -487,25 +486,27 @@ protected:
 
 private:
   ~TargetX8632() override {}
-  template <typename T> void emitConstantPool() const;
 };
 
-class TargetGlobalInitX8632 : public TargetGlobalInitLowering {
-  TargetGlobalInitX8632(const TargetGlobalInitX8632 &) = delete;
-  TargetGlobalInitX8632 &operator=(const TargetGlobalInitX8632 &) = delete;
+class TargetGlobalX8632 : public TargetGlobalLowering {
+  TargetGlobalX8632() = delete;
+  TargetGlobalX8632(const TargetGlobalX8632 &) = delete;
+  TargetGlobalX8632 &operator=(const TargetGlobalX8632 &) = delete;
 
 public:
-  static TargetGlobalInitLowering *create(GlobalContext *Ctx) {
-    return new TargetGlobalInitX8632(Ctx);
+  static TargetGlobalLowering *create(GlobalContext *Ctx) {
+    return new TargetGlobalX8632(Ctx);
   }
 
-  virtual void lower(const VariableDeclaration &Var) final;
+  virtual void lowerInit(const VariableDeclaration &Var) const final;
+  virtual void lowerConstants(GlobalContext *Ctx) const final;
 
 protected:
-  TargetGlobalInitX8632(GlobalContext *Ctx);
+  TargetGlobalX8632(GlobalContext *Ctx);
 
 private:
-  ~TargetGlobalInitX8632() override {}
+  ~TargetGlobalX8632() override {}
+  template <typename T> static void emitConstantPool(GlobalContext *Ctx);
 };
 
 template <> void ConstantInteger32::emit(GlobalContext *Ctx) const;

src/IceTranslator.cpp

@@ -12,13 +12,6 @@
 //
 //===----------------------------------------------------------------------===//
 
-#include <iostream>
-#include <memory>
-
-#include "llvm/IR/Constant.h"
-#include "llvm/IR/Constants.h"
-#include "llvm/IR/Module.h"
-
 #include "IceCfg.h"
 #include "IceClFlags.h"
 #include "IceDefs.h"
@@ -28,16 +21,11 @@
 
 using namespace Ice;
 
-namespace {
-
-// Match a symbol name against a match string.  An empty match string
-// means match everything.  Returns true if there is a match.
-bool matchSymbolName(const IceString &SymbolName, const IceString &Match) {
-  return Match.empty() || Match == SymbolName;
+Translator::Translator(GlobalContext *Ctx, const ClFlags &Flags)
+    : Ctx(Ctx), Flags(Flags),
+      GlobalLowering(TargetGlobalLowering::createLowering(Ctx)), ErrorStatus() {
 }
 
-} // end of anonymous namespace
-
 Translator::~Translator() {}
 
 IceString Translator::createUnnamedName(const IceString &Prefix, SizeT Index) {
@@ -66,49 +54,28 @@ bool Translator::checkIfUnnamedNameSafe(const IceString &Name, const char *Kind,
   return false;
 }
 
-void Translator::translateFcn(Cfg *Fcn) {
-  Ctx->resetStats();
-  Func.reset(Fcn);
-  VerboseMask OldVerboseMask = Ctx->getVerbose();
-  if (!matchSymbolName(Func->getFunctionName(), Ctx->getFlags().VerboseFocusOn))
-    Ctx->setVerbose(IceV_None);
-
-  if (Ctx->getFlags().DisableTranslation ||
-      !matchSymbolName(Func->getFunctionName(),
-                       Ctx->getFlags().TranslateOnly)) {
-    Func->dump();
-  } else {
-    Func->translate();
-    if (Func->hasError()) {
-      std::cerr << "ICE translation error: " << Func->getError() << "\n";
-      ErrorStatus = true;
-    }
-
-    if (!ErrorStatus) {
-      if (Ctx->getFlags().UseIntegratedAssembler) {
-        Func->emitIAS();
-      } else {
-        Func->emit();
-      }
-    }
-    Ctx->dumpStats(Func->getFunctionName());
+void Translator::translateFcn(Cfg *Func) {
+  Ctx->cfgQueueBlockingPush(Func);
+  if (Ctx->getFlags().NumTranslationThreads == 0) {
+    Ctx->translateFunctions();
   }
-
-  Ctx->setVerbose(OldVerboseMask);
 }
 
 void Translator::emitConstants() {
-  if (!Ctx->getFlags().DisableTranslation && Func)
-    Func->getTarget()->emitConstants();
+  if (!getErrorStatus())
+    GlobalLowering->lowerConstants(Ctx);
+}
+
+void Translator::transferErrorCode() const {
+  if (getErrorStatus())
+    Ctx->getErrorStatus()->assign(getErrorStatus().value());
 }
 
 void Translator::lowerGlobals(
     const VariableDeclarationListType &VariableDeclarations) {
-  std::unique_ptr<TargetGlobalInitLowering> GlobalLowering(
-      TargetGlobalInitLowering::createLowering(Ctx->getTargetArch(), Ctx));
   bool DisableTranslation = Ctx->getFlags().DisableTranslation;
   const bool DumpGlobalVariables =
-      ALLOW_DUMP && Ctx->isVerbose() && Ctx->getFlags().VerboseFocusOn.empty();
+      ALLOW_DUMP && Ctx->getVerbose() && Ctx->getFlags().VerboseFocusOn.empty();
   OstreamLocker L(Ctx);
   Ostream &Stream = Ctx->getStrDump();
   const IceString &TranslateOnly = Ctx->getFlags().TranslateOnly;
@@ -116,8 +83,7 @@ void Translator::lowerGlobals(
     if (DumpGlobalVariables)
       Global->dump(getContext(), Stream);
     if (!DisableTranslation &&
-        matchSymbolName(Global->getName(), TranslateOnly))
-      GlobalLowering->lower(*Global);
+        GlobalContext::matchSymbolName(Global->getName(), TranslateOnly))
+      GlobalLowering->lowerInit(*Global);
   }
-  GlobalLowering.reset();
 }

src/IceTranslator.h

@@ -15,8 +15,6 @@
 #ifndef SUBZERO_SRC_ICETRANSLATOR_H
 #define SUBZERO_SRC_ICETRANSLATOR_H
 
-#include <memory>
-
 namespace llvm {
 class Module;
 }
@@ -38,24 +36,25 @@ class Translator {
 public:
   typedef std::vector<VariableDeclaration *> VariableDeclarationListType;
 
-  Translator(GlobalContext *Ctx, const ClFlags &Flags)
-      : Ctx(Ctx), Flags(Flags), ErrorStatus(0) {}
-
+  Translator(GlobalContext *Ctx, const ClFlags &Flags);
   ~Translator();
-  bool getErrorStatus() const { return ErrorStatus; }
+  const ErrorCode &getErrorStatus() const { return ErrorStatus; }
 
   GlobalContext *getContext() const { return Ctx; }
 
   const ClFlags &getFlags() const { return Flags; }
 
   /// Translates the constructed ICE function Fcn to machine code.
-  /// Takes ownership of Fcn. Note: As a side effect, Field Func is
-  /// set to Fcn.
-  void translateFcn(Cfg *Fcn);
+  /// Takes ownership of Func.
+  void translateFcn(Cfg *Func);
 
   /// Emits the constant pool.
   void emitConstants();
 
+  /// If there was an error during bitcode reading/parsing, copy the
+  /// error code into the GlobalContext.
+  void transferErrorCode() const;
+
   /// Lowers the given list of global addresses to target. Generates
   /// list of corresponding variable declarations.
   void lowerGlobals(const VariableDeclarationListType &VariableDeclarations);
@@ -72,18 +71,9 @@ public:
 protected:
   GlobalContext *Ctx;
   const ClFlags &Flags;
-  // The exit status of the translation. False is successful. True
-  // otherwise.
-  bool ErrorStatus;
-  // Ideally, Func would be inside the methods that converts IR to
-  // functions.  However, emitting the constant pool requires a valid
-  // Cfg object, so we need to defer deleting the last non-empty Cfg
-  // object to emit the constant pool (via emitConstants). TODO:
-  // Since all constants are globally pooled in the GlobalContext
-  // object, change all Constant related functions to use
-  // GlobalContext instead of Cfg, and then make emitConstantPool use
-  // that.
-  std::unique_ptr<Cfg> Func;
+  std::unique_ptr<TargetGlobalLowering> GlobalLowering;
+  // Exit status of the translation. False is successful. True otherwise.
+  ErrorCode ErrorStatus;
 };
 
 } // end of namespace Ice

src/IceUtils.h

@@ -54,6 +54,133 @@ public:
   }
 };
 
+// BoundedProducerConsumerQueue is a work queue that allows multiple
+// producers and multiple consumers.  A producer adds entries using
+// blockingPush(), and may block if the queue is "full".  A producer
+// uses notifyEnd() to indicate that no more entries will be added.  A
+// consumer removes an item using blockingPop(), which will return
+// nullptr if notifyEnd() has been called and the queue is empty (it
+// never returns nullptr if the queue contained any items).
+//
+// The MaxSize ctor arg controls the maximum size the queue can grow
+// to (subject to a hard limit of MaxStaticSize-1).  The Sequential
+// arg indicates purely sequential execution in which the single
+// thread should never wait().
+//
+// Two condition variables are used in the implementation.
+// GrewOrEnded signals a waiting worker that a producer has changed
+// the state of the queue.  Shrunk signals a blocked producer that a
+// consumer has changed the state of the queue.
+//
+// The methods begin with Sequential-specific code to be most clear.
+// The lock and condition variables are not used in the Sequential
+// case.
+//
+// Internally, the queue is implemented as a circular array of size
+// MaxStaticSize, where the queue boundaries are denoted by the Front
+// and Back fields.  Front==Back indicates an empty queue.
+template <typename T, size_t MaxStaticSize = 128>
+class BoundedProducerConsumerQueue {
+  BoundedProducerConsumerQueue() = delete;
+  BoundedProducerConsumerQueue(const BoundedProducerConsumerQueue &) = delete;
+  BoundedProducerConsumerQueue &
+  operator=(const BoundedProducerConsumerQueue &) = delete;
+
+public:
+  BoundedProducerConsumerQueue(size_t MaxSize, bool Sequential)
+      : Back(0), Front(0), MaxSize(std::min(MaxSize, MaxStaticSize)),
+        Sequential(Sequential), IsEnded(false) {}
+  void blockingPush(T *Item) {
+    {
+      std::unique_lock<GlobalLockType> L(Lock);
+      // If the work queue is already "full", wait for a consumer to
+      // grab an element and shrink the queue.
+      Shrunk.wait(L, [this] { return size() < MaxSize || Sequential; });
+      push(Item);
+    }
+    GrewOrEnded.notify_one();
+  }
+  T *blockingPop() {
+    T *Item = nullptr;
+    bool ShouldNotifyProducer = false;
+    {
+      std::unique_lock<GlobalLockType> L(Lock);
+      GrewOrEnded.wait(L, [this] { return IsEnded || !empty() || Sequential; });
+      if (!empty()) {
+        Item = pop();
+        ShouldNotifyProducer = !IsEnded;
+      }
+    }
+    if (ShouldNotifyProducer)
+      Shrunk.notify_one();
+    return Item;
+  }
+  void notifyEnd() {
+    {
+      std::lock_guard<GlobalLockType> L(Lock);
+      IsEnded = true;
+    }
+    GrewOrEnded.notify_all();
+  }
+
+private:
+  const static size_t MaxStaticSizeMask = MaxStaticSize - 1;
+  static_assert(!(MaxStaticSize & (MaxStaticSize - 1)),
+                "MaxStaticSize must be a power of 2");
+
+  // WorkItems and Lock are read/written by all.
+  ICE_CACHELINE_BOUNDARY;
+  T *WorkItems[MaxStaticSize];
+  ICE_CACHELINE_BOUNDARY;
+  // Lock guards access to WorkItems, Front, Back, and IsEnded.
+  GlobalLockType Lock;
+
+  ICE_CACHELINE_BOUNDARY;
+  // GrewOrEnded is written by the producers and read by the
+  // consumers.  It is notified (by the producer) when something is
+  // added to the queue, in case consumers are waiting for a non-empty
+  // queue.
+  std::condition_variable GrewOrEnded;
+  // Back is the index into WorkItems[] of where the next element will
+  // be pushed.  (More precisely, Back&MaxStaticSize is the index.)
+  // It is written by the producers, and read by all via size() and
+  // empty().
+  size_t Back;
+
+  ICE_CACHELINE_BOUNDARY;
+  // Shrunk is notified (by the consumer) when something is removed
+  // from the queue, in case a producer is waiting for the queue to
+  // drop below maximum capacity.  It is written by the consumers and
+  // read by the producers.
+  std::condition_variable Shrunk;
+  // Front is the index into WorkItems[] of the oldest element,
+  // i.e. the next to be popped.  (More precisely Front&MaxStaticSize
+  // is the index.)  It is written by the consumers, and read by all
+  // via size() and empty().
+  size_t Front;
+
+  ICE_CACHELINE_BOUNDARY;
+
+  // MaxSize and Sequential are read by all and written by none.
+  const size_t MaxSize;
+  const bool Sequential;
+  // IsEnded is read by the consumers, and only written once by the
+  // producer.
+  bool IsEnded;
+
+  // The lock must be held when the following methods are called.
+  bool empty() const { return Front == Back; }
+  size_t size() const { return Back - Front; }
+  void push(T *Item) {
+    WorkItems[Back++ & MaxStaticSizeMask] = Item;
+    assert(size() <= MaxStaticSize);
+  }
+  T *pop() {
+    assert(!empty());
+    return WorkItems[Front++ & MaxStaticSizeMask];
+  }
+};
+
 } // end of namespace Ice
 
 #endif // SUBZERO_SRC_ICEUTILS_H

src/PNaClTranslator.cpp

@@ -32,8 +32,6 @@
 #include "IceOperand.h"
 #include "PNaClTranslator.h"
 
-#include <memory>
-
 namespace {
 using namespace llvm;
 
@@ -159,7 +157,7 @@ public:
   typedef std::vector<Ice::FunctionDeclaration *> FunctionDeclarationListType;
 
   TopLevelParser(Ice::Translator &Translator, NaClBitcodeHeader &Header,
-                 NaClBitstreamCursor &Cursor, bool &ErrorStatus)
+                 NaClBitstreamCursor &Cursor, Ice::ErrorCode &ErrorStatus)
       : NaClBitcodeParser(Cursor), Translator(Translator), Header(Header),
         ErrorStatus(ErrorStatus), NumErrors(0), NumFunctionIds(0),
         NumFunctionBlocks(0), BlockParser(nullptr) {}
@@ -362,7 +360,7 @@ private:
   // The bitcode header.
   NaClBitcodeHeader &Header;
   // The exit status that should be set to true if an error occurs.
-  bool &ErrorStatus;
+  Ice::ErrorCode &ErrorStatus;
   // The number of errors reported.
   unsigned NumErrors;
   // The types associated with each type ID.
@@ -425,7 +423,7 @@ private:
 };
 
 bool TopLevelParser::Error(const std::string &Message) {
-  ErrorStatus = true;
+  ErrorStatus.assign(Ice::EC_Bitcode);
   ++NumErrors;
   Ice::GlobalContext *Context = Translator.getContext();
   Ice::OstreamLocker L(Context);
@@ -2812,7 +2810,6 @@ private:
 
   void ExitBlock() override {
     InstallGlobalNamesAndGlobalVarInitializers();
-    getTranslator().emitConstants();
   }
 
   void ProcessRecord() override;
@@ -2945,7 +2942,7 @@ void PNaClTranslator::translate(const std::string &IRFilename) {
       MemoryBuffer::getFileOrSTDIN(IRFilename);
   if (std::error_code EC = ErrOrFile.getError()) {
     errs() << "Error reading '" << IRFilename << "': " << EC.message() << "\n";
-    ErrorStatus = true;
+    ErrorStatus.assign(EC.value());
     return;
   }
 
@@ -2958,7 +2955,7 @@ void PNaClTranslator::translateBuffer(const std::string &IRFilename,
   if (MemBuf->getBufferSize() % 4 != 0) {
     errs() << IRFilename
            << ": Bitcode stream should be a multiple of 4 bytes in length.\n";
-    ErrorStatus = true;
+    ErrorStatus.assign(EC_Bitcode);
     return;
   }
 
@@ -2969,7 +2966,7 @@ void PNaClTranslator::translateBuffer(const std::string &IRFilename,
   NaClBitcodeHeader Header;
   if (Header.Read(BufPtr, EndBufPtr) || !Header.IsSupported()) {
     errs() << "Invalid PNaCl bitcode header.\n";
-    ErrorStatus = true;
+    ErrorStatus.assign(EC_Bitcode);
     return;
   }
 
@@ -2981,7 +2978,7 @@ void PNaClTranslator::translateBuffer(const std::string &IRFilename,
   int TopLevelBlocks = 0;
   while (!InputStream.AtEndOfStream()) {
     if (Parser.Parse()) {
-      ErrorStatus = true;
+      ErrorStatus.assign(EC_Bitcode);
       return;
     }
     ++TopLevelBlocks;
@@ -2991,7 +2988,7 @@ void PNaClTranslator::translateBuffer(const std::string &IRFilename,
     errs() << IRFilename
            << ": Contains more than one module. Found: " << TopLevelBlocks
            << "\n";
-    ErrorStatus = true;
+    ErrorStatus.assign(EC_Bitcode);
   }
 }
 

src/PNaClTranslator.h

@@ -34,8 +34,8 @@ public:
       : Translator(Ctx, Flags) {}
 
   // Reads the PNaCl bitcode file and translates to ICE, which is then
-  // converted to machine code. Sets ErrorStatus to true if any
-  // errors occurred.
+  // converted to machine code. Sets ErrorStatus to 1 if any errors
+  // occurred.
   void translate(const std::string &IRFilename);
 
   // Reads MemBuf, assuming it is the PNaCl bitcode contents of IRFilename.

src/llvm2ice.cpp

@@ -203,6 +203,18 @@ GenerateBuildAtts("build-atts",
                            "this executable."),
                   cl::init(false));
 
+// Number of translation threads (in addition to the parser thread and
+// the emitter thread).  The special case of 0 means purely
+// sequential, i.e. parser, translator, and emitter all within the
+// same single thread.  (This may need a slight rework if we expand to
+// multiple parser or emitter threads.)
+static cl::opt<uint32_t>
+NumThreads("threads",
+           cl::desc("Number of translation threads (0 for purely sequential)"),
+           // TODO(stichnot): Settle on a good default.  Consider
+           // something related to std::thread::hardware_concurrency().
+           cl::init(0));
+
 static int GetReturnValue(int Val) {
   if (AlwaysExitSuccess)
     return 0;
@@ -274,12 +286,12 @@ int main(int argc, char **argv) {
 
   ValidateAndGenerateBuildAttributes(GenerateBuildAtts ? Ls.get() : nullptr);
   if (GenerateBuildAtts)
-    return GetReturnValue(0);
+    return GetReturnValue(Ice::EC_None);
 
   if (!ALLOW_DISABLE_IR_GEN && DisableIRGeneration) {
     *Ls << "Error: Build doesn't allow --no-ir-gen when not "
         << "ALLOW_DISABLE_IR_GEN!\n";
-    return GetReturnValue(1);
+    return GetReturnValue(Ice::EC_Args);
   }
 
   Ice::ClFlags Flags;
@@ -288,14 +300,15 @@ int main(int argc, char **argv) {
   Flags.DisableTranslation = DisableTranslation;
   Flags.FunctionSections = FunctionSections;
   Flags.DataSections = DataSections;
-  Flags.UseIntegratedAssembler = UseIntegratedAssembler;
   Flags.UseELFWriter = UseELFWriter;
+  Flags.UseIntegratedAssembler = UseIntegratedAssembler;
   Flags.UseSandboxing = UseSandboxing;
   Flags.PhiEdgeSplit = EnablePhiEdgeSplit;
   Flags.DecorateAsm = DecorateAsm;
   Flags.DumpStats = DumpStats;
   Flags.AllowUninitializedGlobals = AllowUninitializedGlobals;
   Flags.TimeEachFunction = TimeEachFunction;
+  Flags.NumTranslationThreads = NumThreads;
   Flags.DefaultGlobalPrefix = DefaultGlobalPrefix;
   Flags.DefaultFunctionPrefix = DefaultFunctionPrefix;
   Flags.TimingFocusOn = TimingFocusOn;
@@ -319,7 +332,7 @@ int main(int argc, char **argv) {
   if (UseELFWriter) {
     if (OutputFilename == "-") {
       *Ls << "Error: writing binary ELF to stdout is unsupported\n";
-      return GetReturnValue(1);
+      return GetReturnValue(Ice::EC_Args);
     }
     std::string ErrorInfo;
     raw_fd_ostream *FdOs =
@@ -328,7 +341,7 @@ int main(int argc, char **argv) {
     if (!ErrorInfo.empty()) {
       *Ls << "Failed to open output file: " << OutputFilename << ":\n"
           << ErrorInfo << "\n";
-      return GetReturnValue(1);
+      return GetReturnValue(Ice::EC_Args);
     }
     ELFStr.reset(new Ice::ELFStreamer(*FdOs));
   } else {
@@ -351,11 +364,14 @@ int main(int argc, char **argv) {
     Ctx.getObjectWriter()->writeInitialELFHeader();
   }
 
-  int ErrorStatus = 0;
+  Ctx.startWorkerThreads();
+
+  std::unique_ptr<Ice::Translator> Translator;
   if (BuildOnRead) {
-    Ice::PNaClTranslator Translator(&Ctx, Flags);
-    Translator.translate(IRFilename);
-    ErrorStatus = Translator.getErrorStatus();
+    std::unique_ptr<Ice::PNaClTranslator> PTranslator(
+        new Ice::PNaClTranslator(&Ctx, Flags));
+    PTranslator->translate(IRFilename);
+    Translator.reset(PTranslator.release());
   } else if (ALLOW_LLVM_IR) {
     // Parse the input LLVM IR file into a module.
     SMDiagnostic Err;
@@ -367,17 +383,23 @@ int main(int argc, char **argv) {
 
     if (!Mod) {
       Err.print(argv[0], errs());
-      return GetReturnValue(1);
+      return GetReturnValue(Ice::EC_Bitcode);
     }
 
-    Ice::Converter Converter(Mod, &Ctx, Flags);
-    Converter.convertToIce();
-    ErrorStatus = Converter.getErrorStatus();
+    std::unique_ptr<Ice::Converter> Converter(
+        new Ice::Converter(Mod, &Ctx, Flags));
+    Converter->convertToIce();
+    Translator.reset(Converter.release());
   } else {
     *Ls << "Error: Build doesn't allow LLVM IR, "
         << "--build-on-read=0 not allowed\n";
-    return GetReturnValue(1);
+    return GetReturnValue(Ice::EC_Args);
   }
+
+  Ctx.waitForWorkerThreads();
+  Translator->transferErrorCode();
+  Translator->emitConstants();
+
   if (UseELFWriter) {
     Ice::TimerMarker T1(Ice::TimerStack::TT_emit, &Ctx);
     Ctx.getObjectWriter()->writeNonUserSections();
@@ -390,5 +412,5 @@ int main(int argc, char **argv) {
   }
   const bool FinalStats = true;
   Ctx.dumpStats("_FINAL_", FinalStats);
-  return GetReturnValue(ErrorStatus);
+  return GetReturnValue(Ctx.getErrorStatus()->value());
 }

tests_lit/llvm2ice_tests/nacl-atomic-errors.ll

 ; Test that some errors trigger when the usage of NaCl atomic
 ; intrinsics does not match the required ABI.
 
-; RUN: not %p2i -i %s --args --verbose none 2>&1 | FileCheck %s
+; RUN: %p2i -i %s --args --verbose none --exit-success 2>&1 | FileCheck %s
 
 declare i8 @llvm.nacl.atomic.load.i8(i8*, i32)
 declare i16 @llvm.nacl.atomic.load.i16(i16*, i32)

unittest/BitcodeMunge.cpp

@@ -34,7 +34,7 @@ bool IceTest::SubzeroBitcodeMunger::runTest(
   Translator.translateBuffer(TestName, MungedInput.get());
 
   cleanupTest();
-  return Translator.getErrorStatus() == 0;
+  return Translator.getErrorStatus().value() == 0;
 }
 
 } // end of namespace IceTest