Merge branch 'upc-ir'

CMakeLists.txt

@@ -177,6 +177,21 @@ set(UPC_PTS "packed" CACHE STRING "UPC shared pointer representation (packed|str
 set(UPC_PACKED_BITS "20,10,34" CACHE STRING "UPC shared pointer bits per field (phase,thread,addr) default 20,10,34")
 set(UPC_PTS_VADDR_ORDER "first" CACHE STRING "UPC shared pointer field order default first")
 
+# UPC IR Remote Pointer Representation
+set(UPC_IR_REMOTE_POINTER_BITS "20,44" CACHE STRING "UPC remote pointer bits per field (thread,addr) default 20,44")
+string(REPLACE "," ";" UPC_IR_RP_LIST ${UPC_IR_REMOTE_POINTER_BITS})
+list(GET UPC_IR_RP_LIST 0 UPC_RP_THREAD)
+list(GET UPC_IR_RP_LIST 1 UPC_RP_ADDR)
+math(EXPR UPC_RP_SIZE "${UPC_RP_THREAD} + ${UPC_RP_ADDR}")
+if (NOT UPC_RP_SIZE EQUAL 64)
+  message(FATAL_ERROR "UPC Remote Pointer fields must add to 64 bits (UPC_IR_REMOTE_POINTER_BITS).") 
+endif()
+set(UPC_IR_RP_THREAD ${UPC_RP_THREAD})
+set(UPC_IR_RP_ADDR ${UPC_RP_ADDR})
+# UPC IR Remote Pointer address space
+set(UPC_IR_RP_ADDRSPACE "16" CACHE STRING "UPC remote pointer address space")
+set(UPC_IR_RP_ADDRSPACE ${UPC_IR_RP_ADDRSPACE})
+
 # Check for Portals4 and SLURM libraries if necessary
 set(LIBUPC_RUNTIME_MODEL smp CACHE STRING "specify the runtime implementation model for UPC, where MODEL may be: '
 smp' (Symmetric Multiprocessing) or 'portals4' (Infiniband with Portals4 library)")

include/clang/Basic/LangOptions.def

@@ -156,6 +156,7 @@ LANGOPT(UPCPtsRep, 1, 1, "UPC pointer representation (1 packed, 0 struct)")
 LANGOPT(UPCVaddrFirst, 1, 0, "UPC pointer field order")
 LANGOPT(UPCInlineLib, 1, 0, "inline the UPC runtime")
 LANGOPT(UPCPreInclude, 1, 0, "pre-include UPC runtime header files")
+LANGOPT(UPCGenIr, 1, 0, "generate LLVM IR for UPC shared accesses")
 LANGOPT(UPCTLDEnable, 1, 0, "enable support for Thread Local Data")
 
 LANGOPT(MRTD , 1, 0, "-mrtd calling convention")

include/clang/Config/config.h.cmake

@@ -22,6 +22,15 @@
 /* the field order */
 #cmakedefine UPC_PTS_VADDR_ORDER "${UPC_PTS_VADDR_ORDER}"
 
+/* UPC Remote Pointer thread field size */
+#cmakedefine UPC_IR_RP_THREAD ${UPC_IR_RP_THREAD}
+
+/* UPC Remote Pointer address field size */
+#cmakedefine UPC_IR_RP_ADDR ${UPC_IR_RP_ADDR}
+
+/* UPC Remote Pointer address space */
+#cmakedefine UPC_IR_RP_ADDRSPACE ${UPC_IR_RP_ADDRSPACE}
+
 /* UPC runtime NUMA enable. */
 #cmakedefine LIBUPC_ENABLE_NUMA 1
 

include/clang/Config/config.h.in

@@ -27,9 +27,18 @@
 /* the number of bits in each field. */                                         
 #undef UPC_PACKED_BITS
 
-/* the field order */                                                           
+/* the field order. */
 #undef UPC_PTS_VADDR_ORDER
 
+/* UPC Remote Pointer thread field size. */
+#undef UPC_IR_RP_THREAD
+
+/* UPC Remote Pointer address field size. */
+#undef UPC_IR_RP_ADDR
+
+/* UPC Remote Pointer address space. */
+#undef UPC_IR_RP_ADDRSPACE
+
 /* Define to 1 if UPC runtime NUMA is supported. */
 #undef LIBUPC_ENABLE_NUMA
 

include/clang/Driver/Options.td

@@ -800,6 +800,10 @@ def fno_upc_pre_include : Flag<["-"], "fno-upc-pre-include">, Group<f_Group>, Fl
 def fupc_debug : Flag<["-"], "fupc-debug">, Group<f_Group>, Flags<[CC1Option]>,
   HelpText<"Generate UPC runtime calls that include debugging information">;
 def fno_upc_debug : Flag<["-"], "fno-upc-debug">, Group<f_Group>;
+def fupc_ir : Flag<["-"], "fupc-ir">,
+                      Group<f_Group>, Flags<[CC1Option]>;
+def fno_upc_ir : Flag<["-"], "fno-upc-ir">,
+                         Group<f_Group>, Flags<[CC1Option]>;
 def foperator_arrow_depth_EQ : Joined<["-"], "foperator-arrow-depth=">,
                                Group<f_Group>;
 def ftest_coverage : Flag<["-"], "ftest-coverage">, Group<f_Group>;

lib/CodeGen/BackendUtil.cpp

@@ -36,6 +36,7 @@
 #include "llvm/Transforms/Instrumentation.h"
 #include "llvm/Transforms/ObjCARC.h"
 #include "llvm/Transforms/Scalar.h"
+#include "llvm/Transforms/UPC.h"
 using namespace clang;
 using namespace llvm;
 
@@ -222,6 +223,11 @@ static void addDataFlowSanitizerPass(const PassManagerBuilder &Builder,
   PM.add(createDataFlowSanitizerPass(CGOpts.SanitizerBlacklistFile));
 }
 
+static void addLowerUPCPointersPass(const PassManagerBuilder &Builder,
+                                    PassManagerBase &PM) {
+  PM.add(llvm::createLowerUPCPointersPass());
+}
+
 void EmitAssemblyHelper::CreatePasses(TargetMachine *TM) {
   unsigned OptLevel = CodeGenOpts.OptimizationLevel;
   CodeGenOptions::InliningMethod Inlining = CodeGenOpts.getInlining();
@@ -536,6 +542,17 @@ bool EmitAssemblyHelper::AddEmitPasses(BackendAction Action,
       CodeGenOpts.OptimizationLevel > 0)
     PM->add(createObjCARCContractPass());
 
+  // Add LLVM UPC lowering pass unless we're generating LLVM output
+  if (LangOpts.UPCGenIr) {
+    switch(Action) {
+    case Backend_EmitLL:
+    case Backend_EmitBC:
+      break;
+    default:
+      PM->add(llvm::createLowerUPCPointersPass());
+    }
+  }
+
   if (TM->addPassesToEmitFile(*PM, OS, CGFT,
                               /*DisableVerify=*/!CodeGenOpts.VerifyModule)) {
     Diags.Report(diag::err_fe_unable_to_interface_with_target);

lib/CodeGen/CGExprUPC.cpp

@@ -17,6 +17,7 @@
 #include "clang/Basic/TargetInfo.h"
 #include "llvm/IR/DataLayout.h"
 #include "llvm/ADT/SmallVector.h"
+#include "clang/Config/config.h" // for UPC_IR_RP_THREAD/ADDR
 using namespace clang;
 using namespace CodeGen;
 
@@ -180,11 +181,41 @@ llvm::Value *CodeGenFunction::EmitUPCLoad(llvm::Value *Addr,
   return EmitFromMemory(EmitUPCLoad(Addr, isStrict, DestLTy, Align, Loc), Ty);
 }
 
+static const int UPCAddrSpace = UPC_IR_RP_ADDRSPACE;
+
+static llvm::Value *ConvertPTStoLLVMPtr(CodeGenFunction& CGF,
+                                        llvm::Value *Ptr, llvm::Type *LTy) {
+  llvm::Value *Addr = CGF.EmitUPCPointerGetAddr(Ptr);
+  llvm::Value *Thread = CGF.EmitUPCPointerGetThread(Ptr);
+
+  const LangOptions& LangOpts = CGF.getContext().getLangOpts();
+  if(!LangOpts.UPCPtsRep) {
+    llvm::Value *SectionStart = CGF.CGM.getModule().getOrInsertGlobal("__upc_shared_start", CGF.Int8Ty);
+    llvm::Value *StartInt = CGF.Builder.CreatePtrToInt(SectionStart, CGF.PtrDiffTy, "sect.cast");
+    Addr = CGF.Builder.CreateSub(Addr, StartInt, "addr.sub");
+  }
+
+  llvm::Value *IntVal = CGF.Builder.CreateOr(Thread, CGF.Builder.CreateShl(Addr, UPC_IR_RP_THREAD));
+
+  llvm::Type *LLPtsTy = LTy->getPointerTo(UPCAddrSpace);
+  return CGF.Builder.CreateIntToPtr(IntVal, LLPtsTy);
+}
+
 llvm::Value *CodeGenFunction::EmitUPCLoad(llvm::Value *Addr,
                                           bool isStrict,
                                           llvm::Type *LTy,
                                           CharUnits Align,
                                           SourceLocation Loc) {
+  const LangOptions& Opts = getContext().getLangOpts();
+  if (Opts.UPCGenIr) {
+    llvm::Value *InternalAddr = ConvertPTStoLLVMPtr(*this, Addr, LTy);
+    llvm::LoadInst * Result = Builder.CreateLoad(InternalAddr);
+    if(isStrict) {
+      Result->setOrdering(llvm::SequentiallyConsistent);
+    }
+    Result->setAlignment(Align.getQuantity());
+    return Result;
+  }
   const ASTContext& Context = getContext();
   const llvm::DataLayout &Target = CGM.getDataLayout();
   uint64_t Size = Target.getTypeSizeInBits(LTy);
@@ -251,7 +282,16 @@ void CodeGenFunction::EmitUPCStore(llvm::Value *Value,
                                    bool isStrict,
                                    CharUnits Align,
                                    SourceLocation Loc) {
-
+  const LangOptions& Opts = getContext().getLangOpts();
+  if (Opts.UPCGenIr) {
+    llvm::Value *InternalAddr = ConvertPTStoLLVMPtr(*this, Addr, Value->getType());
+    llvm::StoreInst * Result = Builder.CreateStore(Value, InternalAddr);
+    if(isStrict) {
+      Result->setOrdering(llvm::SequentiallyConsistent);
+    }
+    Result->setAlignment(Align.getQuantity());
+    return;
+  }
   const ASTContext& Context = getContext();
   const llvm::DataLayout &Target = CGM.getDataLayout();
   uint64_t Size = Target.getTypeSizeInBits(Value->getType());

lib/CodeGen/CMakeLists.txt

@@ -7,6 +7,7 @@ set(LLVM_LINK_COMPONENTS
   ipo
   linker
   vectorize
+  upc
   )
 
 add_clang_library(clangCodeGen

lib/Driver/Tools.cpp

@@ -3525,6 +3525,8 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
                   options::OPT_fno_upc_inline_lib);
   Args.AddAllArgs(CmdArgs, options::OPT_fupc_pre_include,
                   options::OPT_fno_upc_pre_include);
+  Args.AddAllArgs(CmdArgs, options::OPT_fupc_ir,
+                  options::OPT_fno_upc_ir);
 
   if (Args.hasFlag(options::OPT_fupc_debug,
                    options::OPT_fno_upc_debug, false))

lib/Frontend/CompilerInvocation.cpp

@@ -1497,6 +1497,9 @@ static void ParseLangArgs(LangOptions &Opts, ArgList &Args, InputKind IK,
        Args.hasFlag(OPT_fupc_inline_lib, OPT_fno_upc_inline_lib, false)))
     Opts.UPCInlineLib = true;
 
+  if (Opts.UPC && Args.hasArg(OPT_fupc_ir))
+    Opts.UPCGenIr = true;
+
   // This is the __NO_INLINE__ define, which just depends on things like the
   // optimization level and -fno-inline, not actually whether the backend has
   // inlining enabled.

lib/Frontend/InitPreprocessor.cpp

@@ -378,6 +378,9 @@ static void InitializeStandardPredefinedMacros(const TargetInfo &TI,
     if (LangOpts.UPCInlineLib) {
       Builder.defineMacro("__UPC_INLINE_LIB__", "1");
     }
+    if (LangOpts.UPCGenIr) {
+      Builder.defineMacro("__UPC_SHARED_IR__", "1");
+    }
   }
 
   // Not "standard" per se, but available even with the -undef flag.

runtime/libupc/CMakeLists.txt

@@ -363,6 +363,7 @@ endif()
 elseif(LIBUPC_RUNTIME_MODEL STREQUAL portals4)
   set(LIBUPC_SOURCES
     portals4/gupcr_access.c
+    portals4/gupcr_llvm_access.c
     portals4/gupcr_addr.c
     portals4/gupcr_alloc.upc
     portals4/gupcr_atomic_sup.c
@@ -415,7 +416,9 @@ elseif(LIBUPC_RUNTIME_MODEL STREQUAL portals4)
 
   set(LIBUPC_SOURCES_INLINE
     ${PROJECT_SOURCE_DIR}/portals4/gupcr_access.c
+    ${PROJECT_SOURCE_DIR}/portals4/gupcr_llvm_access.c
     ${PROJECT_SOURCE_DIR}/portals4/gupcr_access.h
+    ${PROJECT_SOURCE_DIR}/portals4/gupcr_llvm_access.h
     ${PROJECT_SOURCE_DIR}/portals4/gupcr_config.h
     ${PROJECT_SOURCE_DIR}/portals4/gupcr_defs.h
     ${PROJECT_SOURCE_DIR}/portals4/gupcr_gmem.h

runtime/libupc/autoconf/configure.ac

@@ -2112,8 +2112,7 @@ if test "$upc_runtime_model" = "smp"; then
       upc_numa=yes ;;
     no)
       upc_numa=no ;;
-    *)
-      AC_MSG_ERROR([--enable-upc-numa accepts only yes or no])
+    *)      AC_MSG_ERROR([--enable-upc-numa accepts only yes or no])
     esac
   ],
   [