[Flang][OMPTarget] Changeset required for initial UMT functionality (llvm#1464)

Author: ronlieb

clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp

@@ -898,7 +898,7 @@ Error handleOverrideImages(
 /// be registered by the runtime.
 Expected<SmallVector<StringRef>> linkAndWrapDeviceFiles(
     SmallVectorImpl<SmallVector<OffloadFile>> &LinkerInputFiles,
-    const InputArgList &Args, char **Argv, int Argc) {
+    InputArgList &Args, char **Argv, int Argc) {
   llvm::TimeTraceScope TimeScope("Handle all device input");
 
   std::mutex ImageMtx;
@@ -909,6 +909,9 @@ Expected<SmallVector<StringRef>> linkAndWrapDeviceFiles(
     if (Error Err = handleOverrideImages(Args, Images))
       return std::move(Err);
 
+  bool ExcludeNVPTX = Args.hasArg(OPT_no_nvptx_whole_archive);
+  bool ExcludeAMDGPU = Args.hasArg(OPT_no_amdgpu_whole_archive);
+
   auto Err = parallelForEachError(LinkerInputFiles, [&](auto &Input) -> Error {
     llvm::TimeTraceScope TimeScope("Link device input");
 
@@ -923,6 +926,13 @@ Expected<SmallVector<StringRef>> linkAndWrapDeviceFiles(
         });
     auto LinkerArgs = getLinkerArgs(Input, BaseArgs);
 
+    const llvm::Triple Triple(LinkerArgs.getLastArgValue(OPT_triple_EQ));
+    if (Triple.isNVPTX() && ExcludeNVPTX)
+      return Error::success();
+
+    if (Triple.isAMDGPU() && ExcludeAMDGPU)
+      return Error::success();
+
     DenseSet<OffloadKind> ActiveOffloadKinds;
     for (const auto &File : Input)
       if (File.getBinary()->getOffloadKind() != OFK_None)
@@ -975,6 +985,13 @@ Expected<SmallVector<StringRef>> linkAndWrapDeviceFiles(
   if (Err)
     return std::move(Err);
 
+  // This option is specific to this link phase and the preceding link tools
+  // do not understand this option so we remove it now that we're done with it.
+  if (ExcludeNVPTX)
+    Args.eraseArg(OPT_no_nvptx_whole_archive);
+  if (ExcludeAMDGPU)
+    Args.eraseArg(OPT_no_amdgpu_whole_archive);
+
   // Create a binary image of each offloading image and embed it into a new
   // object file.
   SmallVector<StringRef> WrappedOutput;

clang/tools/clang-linker-wrapper/LinkerWrapperOpts.td

@@ -119,6 +119,9 @@ def version : Flag<["--", "-"], "version">, Flags<[HelpHidden]>, Alias<v>;
 def whole_archive : Flag<["--", "-"], "whole-archive">, Flags<[HelpHidden]>;
 def no_whole_archive : Flag<["--", "-"], "no-whole-archive">, Flags<[HelpHidden]>;
 
+def no_nvptx_whole_archive : Flag<["--", "-"], "no-nvptx-whole-archive">, Flags<[HelpHidden]>;
+def no_amdgpu_whole_archive : Flag<["--", "-"], "no-amdgpu-whole-archive">, Flags<[HelpHidden]>;
+
 def relocatable : Flag<["--", "-"], "relocatable">, 
     HelpText<"Link device code to create a relocatable offloading application">;
 def r : Flag<["-"], "r">, Alias<relocatable>;

flang/lib/Lower/OpenMP/ClauseProcessor.cpp

@@ -970,7 +970,7 @@ void ClauseProcessor::processMapObjects(
     std::map<Object, OmpMapParentAndMemberData> &parentMemberIndices,
     llvm::SmallVectorImpl<mlir::Value> &mapVars,
     llvm::SmallVectorImpl<const semantics::Symbol *> &mapSyms,
-    llvm::StringRef mapperIdNameRef) const {
+    llvm::StringRef mapperIdNameRef, bool isMotionModifier) const {
   fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
 
   auto getDefaultMapperID = [&](const omp::Object &object,
@@ -1021,7 +1021,7 @@ void ClauseProcessor::processMapObjects(
             treatIndexAsSection);
 
     mlir::Value baseOp = info.rawInput;
-    if (object.sym()->owner().IsDerivedType()) {
+    if (object.sym()->owner().IsDerivedType() && !isMotionModifier) {
       omp::ObjectList objectList = gatherObjectsOf(object, semaCtx);
       assert(!objectList.empty() &&
              "could not find parent objects of derived type member");
@@ -1180,7 +1180,8 @@ bool ClauseProcessor::processMotionClauses(lower::StatementContext &stmtCtx,
     if (expectation && *expectation == omp::clause::To::Expectation::Present)
       mapTypeBits |= llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_PRESENT;
     processMapObjects(stmtCtx, clauseLocation, objects, mapTypeBits,
-                      parentMemberIndices, result.mapVars, mapSymbols);
+                      parentMemberIndices, result.mapVars, mapSymbols, "",
+                      true);
   };
 
   bool clauseFound = findRepeatableClause<omp::clause::To>(callbackFn);

flang/lib/Lower/OpenMP/ClauseProcessor.h

@@ -175,7 +175,8 @@ class ClauseProcessor {
       std::map<Object, OmpMapParentAndMemberData> &parentMemberIndices,
       llvm::SmallVectorImpl<mlir::Value> &mapVars,
       llvm::SmallVectorImpl<const semantics::Symbol *> &mapSyms,
-      llvm::StringRef mapperIdNameRef = "") const;
+      llvm::StringRef mapperIdNameRef = "",
+      bool isMotionModifier = false) const;
 
   lower::AbstractConverter &converter;
   semantics::SemanticsContext &semaCtx;

flang/lib/Lower/OpenMP/Utils.cpp

@@ -551,12 +551,12 @@ void insertChildMapInfoIntoParent(
       mapOp.setMembersIndexAttr(firOpBuilder.create2DI64ArrayAttr(
           indices.second.memberPlacementIndices));
     } else {
-      // NOTE: We take the map type of the first child, this may not
-      // be the correct thing to do, however, we shall see. For the moment
-      // it allows this to work with enter and exit without causing MLIR
-      // verification issues. The more appropriate thing may be to take
-      // the "main" map type clause from the directive being used.
-      uint64_t mapType = indices.second.memberMap[0].getMapType();
+      // NOTE: We do not assign default mapped parents a map type, as
+      // selecting a childs can result in the incorrect map type being
+      // applied to the parent and data being incorrectly moved to or
+      // from device.
+      uint64_t mapType = llvm::to_underlying(
+          llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_NONE);
 
       llvm::SmallVector<mlir::Value> members;
       members.reserve(indices.second.memberMap.size());

flang/lib/Optimizer/OpenMP/MapInfoFinalization.cpp

@@ -181,7 +181,11 @@ class MapInfoFinalizationPass
     return builder.create<mlir::omp::MapInfoOp>(
         loc, baseAddrAddr.getType(), descriptor,
         mlir::TypeAttr::get(underlyingVarType),
-        builder.getIntegerAttr(builder.getIntegerType(64, false), mapType),
+        builder.getIntegerAttr(
+            builder.getIntegerType(64, false),
+            llvm::to_underlying(llvm::omp::OpenMPOffloadMappingFlags(mapType) |
+                                llvm::omp::OpenMPOffloadMappingFlags::
+                                    OMP_MAP_DESCRIPTOR_BASE_ADDR)),
         builder.getAttr<mlir::omp::VariableCaptureKindAttr>(
             mlir::omp::VariableCaptureKind::ByRef),
         baseAddrAddr, /*members=*/mlir::SmallVector<mlir::Value>{},
@@ -246,7 +250,10 @@ class MapInfoFinalizationPass
     using mapFlags = llvm::omp::OpenMPOffloadMappingFlags;
     if (llvm::isa_and_nonnull<mlir::omp::TargetExitDataOp,
                               mlir::omp::TargetUpdateOp>(target))
-      return mapTypeFlag;
+      return llvm::to_underlying(
+          mapFlags(mapTypeFlag) |
+          llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_DESCRIPTOR);
+
     mapFlags flags = mapFlags::OMP_MAP_TO |
                      (mapFlags(mapTypeFlag) & mapFlags::OMP_MAP_IMPLICIT);
     // Descriptors for objects will always be copied. This is because the