[SYCL] Add clang-linker-wrapper changes to call clang-sycl-linker for SYCL offloads

asudarsa · asudarsa · commit 597fda606efb · 2025-04-16T07:33:51.000-07:00
Device code linking happens inside clang-linker-wrapper. In the current
implementation, clang-linker-wrapper does the following:

1. Extracts device code. Input_1, Input_2,.....
2. Group device code according to target devices
Inputs[triple_1] = ....
Inputs[triple_2] = ....
3. For each group, i.e. Inputs[triple_i],
   a. Gather all the offload kinds found inside those inputs in
      ActiveOffloadKinds
   b. Link all images inside Inputs[triple_i] by calling
      clang --target=triple_i ....
   c. Create a copy of that linked image for each offload kind and add it to
      Output[Kind] list.

In SYCL compilation flow, there is a deviation in Step 3b. We call device code
splitting inside the 'clang --target=triple_i ....' call and the output is now
a 'packaged' file containing multiple device images. This deviation requires us
to capture the OffloadKind during the linking stage and pass it along to the
linking function (clang), so that clang can be called with a unique option
'--sycl-link' that will help us to call 'clang-sycl-linker' under the hood
(clang-sycl-linker will do SYCL specific linking).

Our current objective is to implement an end-to-end SYCL offloading flow and
get it working. We will eventually merge our approach with the community flow.

Signed-off-by: Arvind Sudarsanam &lt;arvind.sudarsanam@intel.com&gt;
diff --git a/clang/docs/ClangOffloadPackager.rst b/clang/docs/ClangOffloadPackager.rst
@@ -112,6 +112,8 @@ the following values for the :ref:`offload kind<table-offload_kind>` and the
     +------------+-------+---------------------------------------+
     | OFK_HIP    | 0x03  | The producer was HIP                  |
     +------------+-------+---------------------------------------+
+    | OFK_SYCL   | 0x04  | The producer was SYCL                 |
+    +------------+-------+---------------------------------------+
 
 The flags are used to signify certain conditions, such as the presence of
 debugging information or whether or not LTO was used. The string entry table is
diff --git a/clang/test/Driver/linker-wrapper.c b/clang/test/Driver/linker-wrapper.c
@@ -2,13 +2,15 @@
 // REQUIRES: x86-registered-target
 // REQUIRES: nvptx-registered-target
 // REQUIRES: amdgpu-registered-target
+// REQUIRES: spirv-registered-target
 
 // An externally visible variable so static libraries extract.
 __attribute__((visibility("protected"), used)) int x;
 
 // RUN: %clang -cc1 %s -triple x86_64-unknown-linux-gnu -emit-obj -o %t.elf.o
 // RUN: %clang -cc1 %s -triple nvptx64-nvidia-cuda -emit-llvm-bc -o %t.nvptx.bc
 // RUN: %clang -cc1 %s -triple amdgcn-amd-amdhsa -emit-llvm-bc -o %t.amdgpu.bc
+// RUN: %clang -cc1 %s -triple spirv64-unknown-unknown -emit-llvm-bc -o %t.spirv.bc
 
 // RUN: clang-offload-packager -o %t.out \
 // RUN:   --image=file=%t.elf.o,kind=openmp,triple=nvptx64-nvidia-cuda,arch=sm_70 \
@@ -49,6 +51,14 @@ __attribute__((visibility("protected"), used)) int x;
 
 // AMDGPU-LTO-TEMPS: clang{{.*}} --target=amdgcn-amd-amdhsa -mcpu=gfx1030 -flto {{.*}}-save-temps
 
+// RUN: clang-offload-packager -o %t.out \
+// RUN:   --image=file=%t.spirv.bc,kind=sycl,triple=spirv64-unknown-unknown,arch=generic
+// RUN: %clang -cc1 %s -triple x86_64-unknown-linux-gnu -emit-obj -o %t.o -fembed-offload-object=%t.out
+// RUN: clang-linker-wrapper --host-triple=x86_64-unknown-linux-gnu --dry-run \
+// RUN:   --linker-path=/usr/bin/ld %t.o -o a.out 2>&1 | FileCheck %s --check-prefix=SPIRV-LINK
+
+// SPIRV-LINK: clang{{.*}} -o {{.*}}.img --target=spirv64-unknown-unknown {{.*}}.o --sycl-link -Xlinker -triple=spirv64-unknown-unknown -Xlinker -arch=
+
 // RUN: clang-offload-packager -o %t.out \
 // RUN:   --image=file=%t.elf.o,kind=openmp,triple=x86_64-unknown-linux-gnu \
 // RUN:   --image=file=%t.elf.o,kind=openmp,triple=x86_64-unknown-linux-gnu
diff --git a/clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp b/clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp
@@ -464,7 +464,8 @@ fatbinary(ArrayRef<std::pair<StringRef, StringRef>> InputFiles,
 } // namespace amdgcn
 
 namespace generic {
-Expected<StringRef> clang(ArrayRef<StringRef> InputFiles, const ArgList &Args) {
+Expected<StringRef> clang(ArrayRef<StringRef> InputFiles, const ArgList &Args,
+                          bool HasSYCLOffloadKind = false) {
   llvm::TimeTraceScope TimeScope("Clang");
   // Use `clang` to invoke the appropriate device tools.
   Expected<std::string> ClangPath =
@@ -554,6 +555,17 @@ Expected<StringRef> clang(ArrayRef<StringRef> InputFiles, const ArgList &Args) {
   if (Args.hasArg(OPT_embed_bitcode))
     CmdArgs.push_back("-Wl,--lto-emit-llvm");
 
+  // For linking device code with the SYCL offload kind, special handling is
+  // required. Passing --sycl-link to clang results in a call to
+  // clang-sycl-linker. Additional linker flags required by clang-sycl-linker
+  // will be communicated via the -Xlinker option.
+  if (HasSYCLOffloadKind) {
+    CmdArgs.push_back("--sycl-link");
+    CmdArgs.append(
+        {"-Xlinker", Args.MakeArgString("-triple=" + Triple.getTriple())});
+    CmdArgs.append({"-Xlinker", Args.MakeArgString("-arch=" + Arch)});
+  }
+
   for (StringRef Arg : Args.getAllArgValues(OPT_linker_arg_EQ))
     CmdArgs.append({"-Xlinker", Args.MakeArgString(Arg)});
   for (StringRef Arg : Args.getAllArgValues(OPT_compiler_arg_EQ))
@@ -567,7 +579,8 @@ Expected<StringRef> clang(ArrayRef<StringRef> InputFiles, const ArgList &Args) {
 } // namespace generic
 
 Expected<StringRef> linkDevice(ArrayRef<StringRef> InputFiles,
-                               const ArgList &Args) {
+                               const ArgList &Args,
+                               bool HasSYCLOffloadKind = false) {
   const llvm::Triple Triple(Args.getLastArgValue(OPT_triple_EQ));
   switch (Triple.getArch()) {
   case Triple::nvptx:
@@ -582,7 +595,7 @@ Expected<StringRef> linkDevice(ArrayRef<StringRef> InputFiles,
   case Triple::spirv64:
   case Triple::systemz:
   case Triple::loongarch64:
-    return generic::clang(InputFiles, Args);
+    return generic::clang(InputFiles, Args, HasSYCLOffloadKind);
   default:
     return createStringError(Triple.getArchName() +
                              " linking is not supported");
@@ -936,6 +949,38 @@ Expected<SmallVector<StringRef>> linkAndWrapDeviceFiles(
       InputFiles.emplace_back(*FileNameOrErr);
     }
 
+    if (ActiveOffloadKindMask & OFK_SYCL) {
+      // Link the remaining device files using the device linker.
+      auto OutputOrErr = linkDevice(InputFiles, LinkerArgs, HasSYCLOffloadKind);
+      if (!OutputOrErr)
+        return OutputOrErr.takeError();
+      // Output is a packaged object of device images. Unpackage the images and
+      // copy them to Images[Kind]
+      ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
+          MemoryBuffer::getFileOrSTDIN(*OutputOrErr);
+      if (std::error_code EC = BufferOrErr.getError())
+        return createFileError(*OutputOrErr, EC);
+
+      MemoryBufferRef Buffer = **BufferOrErr;
+      SmallVector<OffloadFile> Binaries;
+      if (Error Err = extractOffloadBinaries(Buffer, Binaries))
+        return std::move(Err);
+      for (auto &OffloadFile : Binaries) {
+        auto TheBinary = OffloadFile.getBinary();
+        OffloadingImage TheImage{};
+        TheImage.TheImageKind = TheBinary->getImageKind();
+        TheImage.TheOffloadKind = TheBinary->getOffloadKind();
+        TheImage.StringData["triple"] = TheBinary->getTriple();
+        TheImage.StringData["arch"] = TheBinary->getArch();
+        TheImage.Image = MemoryBuffer::getMemBufferCopy(TheBinary->getImage());
+        Images[OFK_SYCL].emplace_back(std::move(TheImage));
+      }
+    }
+
+    // Exit early if no other offload kind found (other than OFK_SYCL).
+    if ((ActiveOffloadKindMask ^ OFK_SYCL) == 0) {
+      return Error::success();
+
     // Link the remaining device files using the device linker.
     auto OutputOrErr = linkDevice(InputFiles, LinkerArgs);
     if (!OutputOrErr)
@@ -944,7 +989,7 @@ Expected<SmallVector<StringRef>> linkAndWrapDeviceFiles(
     // Store the offloading image for each linked output file.
     for (OffloadKind Kind = OFK_OpenMP; Kind != OFK_LAST;
          Kind = static_cast<OffloadKind>((uint16_t)(Kind) << 1)) {
-      if ((ActiveOffloadKindMask & Kind) == 0)
+      if (((ActiveOffloadKindMask & Kind) == 0) || (Kind == OFK_SYCL))
         continue;
       llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> FileOrErr =
           llvm::MemoryBuffer::getFileOrSTDIN(*OutputOrErr);
@@ -988,6 +1033,11 @@ Expected<SmallVector<StringRef>> linkAndWrapDeviceFiles(
              A.StringData["arch"] > B.StringData["arch"] ||
              A.TheOffloadKind < B.TheOffloadKind;
     });
+    if (Kind == OFK_SYCL) {
+      // TODO: Update once SYCL offload wrapping logic is available.
+      reportError(
+          createStringError("SYCL offload wrapping logic is not available"));
+    }
     auto BundledImagesOrErr = bundleLinkedOutput(Input, Args, Kind);
     if (!BundledImagesOrErr)
       return BundledImagesOrErr.takeError();
diff --git a/clang/tools/clang-sycl-linker/ClangSYCLLinker.cpp b/clang/tools/clang-sycl-linker/ClangSYCLLinker.cpp
@@ -70,6 +70,8 @@ static StringRef OutputFile;
 /// Directory to dump SPIR-V IR if requested by user.
 static SmallString<128> SPIRVDumpDir;
 
+using OffloadingImage = OffloadBinary::OffloadingImage;
+
 static void printVersion(raw_ostream &OS) {
   OS << clang::getClangToolFullVersion("clang-sycl-linker") << '\n';
 }
@@ -168,10 +170,10 @@ Expected<SmallVector<std::string>> getInput(const ArgList &Args) {
 /// are LLVM IR bitcode files.
 // TODO: Support SPIR-V IR files.
 Expected<std::unique_ptr<Module>> getBitcodeModule(StringRef File,
-                                                   LLVMContext &C) {
+                                                   LLVMContext &Ctx) {
   SMDiagnostic Err;
 
-  auto M = getLazyIRFileModule(File, Err, C);
+  auto M = getLazyIRFileModule(File, Err, Ctx);
   if (M)
     return std::move(M);
   return createStringError(Err.getMessage());
@@ -211,16 +213,16 @@ Expected<SmallVector<std::string>> getSYCLDeviceLibs(const ArgList &Args) {
 /// 3. Link all the images gathered in Step 2 with the output of Step 1 using
 /// linkInModule API. LinkOnlyNeeded flag is used.
 Expected<StringRef> linkDeviceCode(ArrayRef<std::string> InputFiles,
-                                   const ArgList &Args, LLVMContext &C) {
+                                   const ArgList &Args, LLVMContext &Ctx) {
   llvm::TimeTraceScope TimeScope("SYCL link device code");
 
   assert(InputFiles.size() && "No inputs to link");
 
-  auto LinkerOutput = std::make_unique<Module>("sycl-device-link", C);
+  auto LinkerOutput = std::make_unique<Module>("sycl-device-link", Ctx);
   Linker L(*LinkerOutput);
   // Link SYCL device input files.
   for (auto &File : InputFiles) {
-    auto ModOrErr = getBitcodeModule(File, C);
+    auto ModOrErr = getBitcodeModule(File, Ctx);
     if (!ModOrErr)
       return ModOrErr.takeError();
     if (L.linkInModule(std::move(*ModOrErr)))
@@ -235,7 +237,7 @@ Expected<StringRef> linkDeviceCode(ArrayRef<std::string> InputFiles,
   // Link in SYCL device library files.
   const llvm::Triple Triple(Args.getLastArgValue(OPT_triple_EQ));
   for (auto &File : *SYCLDeviceLibFiles) {
-    auto LibMod = getBitcodeModule(File, C);
+    auto LibMod = getBitcodeModule(File, Ctx);
     if (!LibMod)
       return LibMod.takeError();
     if ((*LibMod)->getTargetTriple() == Triple) {
@@ -278,18 +280,18 @@ Expected<StringRef> linkDeviceCode(ArrayRef<std::string> InputFiles,
 /// Converts 'File' from LLVM bitcode to SPIR-V format using SPIR-V backend.
 /// 'Args' encompasses all arguments required for linking device code and will
 /// be parsed to generate options required to be passed into the backend.
-static Expected<StringRef> runSPIRVCodeGen(StringRef File, const ArgList &Args,
-                                           LLVMContext &C) {
+static Error runSPIRVCodeGen(StringRef File, const ArgList &Args,
+                             StringRef OutputFile, LLVMContext &Ctx) {
   llvm::TimeTraceScope TimeScope("SPIR-V code generation");
 
   // Parse input module.
-  SMDiagnostic Err;
-  std::unique_ptr<Module> M = parseIRFile(File, Err, C);
+  SMDiagnostic E;
+  std::unique_ptr<Module> M = parseIRFile(File, E, Ctx);
   if (!M)
-    return createStringError(Err.getMessage());
+    return createStringError(E.getMessage());
 
   if (Error Err = M->materializeAll())
-    return std::move(Err);
+    return Err;
 
   Triple TargetTriple(Args.getLastArgValue(OPT_triple_EQ));
   M->setTargetTriple(TargetTriple);
@@ -333,7 +335,7 @@ static Expected<StringRef> runSPIRVCodeGen(StringRef File, const ArgList &Args,
     errs() << formatv("SPIR-V Backend: input: {0}, output: {1}\n", File,
                       OutputFile);
 
-  return OutputFile;
+  return Error::success();
 }
 
 /// Performs the following steps:
@@ -342,17 +344,61 @@ static Expected<StringRef> runSPIRVCodeGen(StringRef File, const ArgList &Args,
 Error runSYCLLink(ArrayRef<std::string> Files, const ArgList &Args) {
   llvm::TimeTraceScope TimeScope("SYCL device linking");
 
-  LLVMContext C;
+  LLVMContext Ctx;
 
   // Link all input bitcode files and SYCL device library files, if any.
-  auto LinkedFile = linkDeviceCode(Files, Args, C);
+  auto LinkedFile = linkDeviceCode(Files, Args, Ctx);
   if (!LinkedFile)
     reportError(LinkedFile.takeError());
 
+  // TODO: SYCL post link functionality involves device code splitting and will
+  // result in multiple bitcode codes.
+  // The following lines are placeholders to represent multiple files and will
+  // be refactored once SYCL post link support is available.
+  SmallVector<std::string> SplitModules;
+  SplitModules.emplace_back(*LinkedFile);
+
   // SPIR-V code generation step.
-  auto SPVFile = runSPIRVCodeGen(*LinkedFile, Args, C);
-  if (!SPVFile)
-    return SPVFile.takeError();
+  for (size_t I = 0, E = SplitModules.size(); I != E; ++I) {
+    auto Stem = OutputFile.rsplit('.').first;
+    std::string SPVFile(Stem);
+    SPVFile.append("_" + utostr(I) + ".spv");
+    auto Err = runSPIRVCodeGen(SplitModules[I], Args, SPVFile, Ctx);
+    if (Err)
+      return std::move(Err);
+    SplitModules[I] = SPVFile;
+  }
+
+  // Write the final output into file.
+  int FD = -1;
+  if (std::error_code EC = sys::fs::openFileForWrite(OutputFile, FD))
+    return errorCodeToError(EC);
+  llvm::raw_fd_ostream FS(FD, /*shouldClose=*/true);
+
+  for (size_t I = 0, E = SplitModules.size(); I != E; ++I) {
+    auto File = SplitModules[I];
+    llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> FileOrErr =
+        llvm::MemoryBuffer::getFileOrSTDIN(File);
+    if (std::error_code EC = FileOrErr.getError()) {
+      if (DryRun)
+        FileOrErr = MemoryBuffer::getMemBuffer("");
+      else
+        return createFileError(File, EC);
+    }
+    OffloadingImage TheImage{};
+    TheImage.TheImageKind = IMG_Object;
+    TheImage.TheOffloadKind = OFK_SYCL;
+    TheImage.StringData["triple"] =
+        Args.MakeArgString(Args.getLastArgValue(OPT_triple_EQ));
+    TheImage.StringData["arch"] =
+        Args.MakeArgString(Args.getLastArgValue(OPT_arch_EQ));
+    TheImage.Image = std::move(*FileOrErr);
+
+    llvm::SmallString<0> Buffer = OffloadBinary::write(TheImage);
+    if (Buffer.size() % OffloadBinary::getAlignment() != 0)
+      return createStringError("Offload binary has invalid size alignment");
+    FS << Buffer;
+  }
   return Error::success();
 }
 
@@ -394,7 +440,7 @@ int main(int argc, char **argv) {
   DryRun = Args.hasArg(OPT_dry_run);
   SaveTemps = Args.hasArg(OPT_save_temps);
 
-  OutputFile = "a.spv";
+  OutputFile = "a.out";
   if (Args.hasArg(OPT_o))
     OutputFile = Args.getLastArgValue(OPT_o);
 
diff --git a/llvm/lib/Object/OffloadBinary.cpp b/llvm/lib/Object/OffloadBinary.cpp
@@ -301,6 +301,7 @@ OffloadKind object::getOffloadKind(StringRef Name) {
       .Case("openmp", OFK_OpenMP)
       .Case("cuda", OFK_Cuda)
       .Case("hip", OFK_HIP)
+      .Case("sycl", OFK_SYCL)
       .Default(OFK_None);
 }
 
@@ -312,6 +313,8 @@ StringRef object::getOffloadKindName(OffloadKind Kind) {
     return "cuda";
   case OFK_HIP:
     return "hip";
+  case OFK_SYCL:
+    return "sycl";
   default:
     return "none";
   }
