[LinkerWrapper] Support relocatable linking for offloading (#80066)

Summary:
The standard GPU compilation process embeds each intermediate object
file into the host file at the `.llvm.offloading` section so it can be
linked later. We also use a special section called something like
`omp_offloading_entries` to store all the globals that need to be
registered by the runtime. The linker-wrapper's job is to link the
embedded device code stored at this section and then emit code to
register the linked image and the kernels and globals in the offloading
entry section.

One downside to RDC linking is that it can become quite big for very
large projects that wish to make use of static linking. This patch
changes the support for relocatable linking via `-r` to support a kind
of "partial" RDC compilation for offloading languages.

This primarily requires manually editing the embedded data in the
output object file for the relocatable link. We need to rename the
output section to make it distinct from the input sections that will be
merged. We then delete the old embedded object code so it won't be
linked further. We then need to rename the old offloading section so
that it is private to the module. A runtime solution could also be done
to defer entries that don't belong to the given GPU executable, but this
is easier. Note that this does not work with COFF linking, only the ELF
method for handling offloading entries, that could be made to work
similarly.

Given this support, the following compilation path should produce two
distinct images for OpenMP offloading.
```
$ clang foo.c -fopenmp --offload-arch=native -c
$ clang foo.c -lomptarget.devicertl --offload-link -r -o merged.o
$ clang main.c merged.o -fopenmp --offload-arch=native
$ ./a.out
```

Or similarly for HIP to effectively perform non-RDC mode compilation for
a subset of files.

```
$ clang -x hip foo.c --offload-arch=native --offload-new-driver -fgpu-rdc -c
$ clang -x hip foo.c -lomptarget.devicertl --offload-link -r -o merged.o
$ clang -x hip main.c merged.o --offload-arch=native --offload-new-driver -fgpu-rdc
$ ./a.out
```

One question is whether or not this should be the default behavior of
`-r` when run through the linker-wrapper or a special option. Standard
`-r` behavior is still possible if used without invoking the
linker-wrapper and it guaranteed to be correct.

clang/docs/ClangLinkerWrapper.rst

@@ -54,12 +54,30 @@ only for the linker wrapper will be forwarded to the wrapped linker job.
     --pass-remarks=<value> Pass remarks for LTO
     --print-wrapped-module Print the wrapped module's IR for testing
     --ptxas-arg=<value>    Argument to pass to the 'ptxas' invocation
+    --relocatable           Link device code to create a relocatable offloading application
     --save-temps           Save intermediate results
     --sysroot<value>       Set the system root
     --verbose              Verbose output from tools
     --v                    Display the version number and exit
     --                     The separator for the wrapped linker arguments
 
+Relocatable Linking
+===================
+
+The ``clang-linker-wrapper`` handles linking embedded device code and then 
+registering it with the appropriate runtime. Normally, this is only done when 
+the executable is created so other files containing device code can be linked 
+together. This can be somewhat problematic for users who wish to ship static 
+libraries that contain offloading code to users without a compatible offloading 
+toolchain.
+
+When using a relocatable link with ``-r``, the ``clang-linker-wrapper`` will 
+perform the device linking and registration eagerly. This will remove the 
+embedded device code and register it correctly with the runtime. Semantically, 
+this is similar to creating a shared library object. If standard relocatable 
+linking is desired, simply do not run the binaries through the 
+``clang-linker-wrapper``. This will simply append the embedded device code so 
+that it can be linked later.
 
 Example
 =======

clang/docs/OffloadingDesign.rst

@@ -470,3 +470,37 @@ We can see the steps created by clang to generate the offloading code using the
     # "nvptx64-nvidia-cuda" - "NVPTX::Assembler", inputs: ["/tmp/zaxpy-07f434.s"], output: "/tmp/zaxpy-0af7b7.o"
     # "x86_64-unknown-linux-gnu" - "clang", inputs: ["/tmp/zaxpy-e6a41b.bc", "/tmp/zaxpy-0af7b7.o"], output: "/tmp/zaxpy-416cad.o"
     # "x86_64-unknown-linux-gnu" - "Offload::Linker", inputs: ["/tmp/zaxpy-416cad.o"], output: "a.out"
+
+Relocatable Linking
+-------------------
+
+The offloading compilation pipeline normally will defer the final device linking 
+and runtime registration until the ``clang-linker-wrapper`` is run to create the 
+executable. This is the standard behaviour when compiling for OpenMP offloading 
+or CUDA and HIP in ``-fgpu-rdc`` mode. However, there are some cases where the 
+user may wish to perform this device handling prematurely. This is described in 
+the :doc:`linker wrapper documentation<ClangLinkerWrapper>`.
+
+Effectively, this allows the user to handle offloading specific linking ahead of 
+time when shipping objects or static libraries. This can be thought of as 
+performing a standard ``-fno-gpu-rdc`` compilation on a subset of object files. 
+This can be useful to reduce link time, prevent users from interacting with the 
+library's device code, or for shipping libraries to incompatible compilers.
+
+Normally, if a relocatable link is done using ``clang -r`` it will simply merge 
+the ``.llvm.offloading`` sections which will then be linked later when the 
+executable is created. However, if the ``-r`` flag is used with the offloading 
+toolchain, it will perform the device linking and registration phases and then 
+merge the registration code into the final relocatable object file.
+
+The following example shows how using the relocatable link with the offloading 
+pipeline can create a static library with offloading code that can be 
+redistributed without requiring any additional handling.
+
+.. code-block:: console
+
+    $ clang++ -fopenmp -fopenmp-targets=nvptx64 foo.cpp -c
+    $ clang++ -lomptarget.devicertl --offload-link -r foo.o -o merged.o
+    $ llvm-ar rcs libfoo.a merged.o
+    # g++ app.cpp -L. -lfoo
+

clang/test/Driver/linker-wrapper-image.c

@@ -9,6 +9,8 @@
 // RUN:   -fembed-offload-object=%t.out
 // RUN: clang-linker-wrapper --print-wrapped-module --dry-run --host-triple=x86_64-unknown-linux-gnu \
 // RUN:   --linker-path=/usr/bin/ld -- %t.o -o a.out 2>&1 | FileCheck %s --check-prefixes=OPENMP,OPENMP-ELF
+// RUN: clang-linker-wrapper --print-wrapped-module --dry-run -r --host-triple=x86_64-unknown-linux-gnu \
+// RUN:   --linker-path=/usr/bin/ld -- %t.o -o a.out 2>&1 | FileCheck %s --check-prefixes=OPENMP-ELF,OPENMP-REL
 // RUN: clang-linker-wrapper --print-wrapped-module --dry-run --host-triple=x86_64-unknown-windows-gnu \
 // RUN:   --linker-path=/usr/bin/ld -- %t.o -o a.out 2>&1 | FileCheck %s --check-prefixes=OPENMP,OPENMP-COFF
 
@@ -19,6 +21,8 @@
 //      OPENMP-COFF: @__start_omp_offloading_entries = weak_odr hidden constant [0 x %struct.__tgt_offload_entry] zeroinitializer, section "omp_offloading_entries$OA"
 // OPENMP-COFF-NEXT: @__stop_omp_offloading_entries = weak_odr hidden constant [0 x %struct.__tgt_offload_entry] zeroinitializer, section "omp_offloading_entries$OZ"
 
+// OPENMP-REL: @.omp_offloading.device_image = internal unnamed_addr constant [[[SIZE:[0-9]+]] x i8] c"\10\FF\10\AD{{.*}}", section ".llvm.offloading.relocatable", align 8
+
 //      OPENMP: @.omp_offloading.device_image = internal unnamed_addr constant [[[SIZE:[0-9]+]] x i8] c"\10\FF\10\AD{{.*}}", section ".llvm.offloading", align 8
 // OPENMP-NEXT: @.omp_offloading.device_images = internal unnamed_addr constant [1 x %__tgt_device_image] [%__tgt_device_image { ptr getelementptr inbounds ([[[BEGIN:[0-9]+]] x i8], ptr @.omp_offloading.device_image, i64 1, i64 0), ptr getelementptr inbounds ([[[END:[0-9]+]] x i8], ptr @.omp_offloading.device_image, i64 1, i64 0), ptr @__start_omp_offloading_entries, ptr @__stop_omp_offloading_entries }]
 // OPENMP-NEXT: @.omp_offloading.descriptor = internal constant %__tgt_bin_desc { i32 1, ptr @.omp_offloading.device_images, ptr @__start_omp_offloading_entries, ptr @__stop_omp_offloading_entries }
@@ -42,6 +46,8 @@
 // RUN:   -fembed-offload-object=%t.out
 // RUN: clang-linker-wrapper --print-wrapped-module --dry-run --host-triple=x86_64-unknown-linux-gnu \
 // RUN:   --linker-path=/usr/bin/ld -- %t.o -o a.out 2>&1 | FileCheck %s --check-prefixes=CUDA,CUDA-ELF
+// RUN: clang-linker-wrapper --print-wrapped-module --dry-run -r --host-triple=x86_64-unknown-linux-gnu \
+// RUN:   --linker-path=/usr/bin/ld -- %t.o -o a.out 2>&1 | FileCheck %s --check-prefixes=CUDA,CUDA-ELF
 // RUN: clang-linker-wrapper --print-wrapped-module --dry-run --host-triple=x86_64-unknown-windows-gnu \
 // RUN:   --linker-path=/usr/bin/ld -- %t.o -o a.out 2>&1 | FileCheck %s --check-prefixes=CUDA,CUDA-COFF
 
@@ -140,6 +146,8 @@
 // RUN:   -fembed-offload-object=%t.out
 // RUN: clang-linker-wrapper --print-wrapped-module --dry-run --host-triple=x86_64-unknown-linux-gnu \
 // RUN:   --linker-path=/usr/bin/ld -- %t.o -o a.out 2>&1 | FileCheck %s --check-prefixes=HIP,HIP-ELF
+// RUN: clang-linker-wrapper --print-wrapped-module --dry-run --host-triple=x86_64-unknown-linux-gnu -r \
+// RUN:   --linker-path=/usr/bin/ld -- %t.o -o a.out 2>&1 | FileCheck %s --check-prefixes=HIP,HIP-ELF
 // RUN: clang-linker-wrapper --print-wrapped-module --dry-run --host-triple=x86_64-unknown-windows-gnu \
 // RUN:   --linker-path=/usr/bin/ld -- %t.o -o a.out 2>&1 | FileCheck %s --check-prefixes=HIP,HIP-COFF
 

clang/test/Driver/linker-wrapper.c

@@ -176,10 +176,36 @@ __attribute__((visibility("protected"), used)) int x;
 // 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
 // RUN: %clang -cc1 %s -triple x86_64-unknown-linux-gnu -emit-obj -o %t.o -fembed-offload-object=%t.out
-// RUN: llvm-ar rcs %t.a %t.o
 // RUN: clang-linker-wrapper --host-triple=x86_64-unknown-linux-gnu --dry-run \
-// RUN:   --linker-path=/usr/bin/ld.lld -- -r --whole-archive %t.a --no-whole-archive \
+// RUN:   --linker-path=/usr/bin/ld.lld -- -r %t.o \
 // RUN:   %t.o -o a.out 2>&1 | FileCheck %s --check-prefix=RELOCATABLE-LINK
 
-// RELOCATABLE-LINK-NOT: clang{{.*}} -o {{.*}}.img --target=x86_64-unknown-linux-gnu
+// RELOCATABLE-LINK: clang{{.*}} -o {{.*}}.img --target=x86_64-unknown-linux-gnu
 // RELOCATABLE-LINK: /usr/bin/ld.lld{{.*}}-r
+// RELOCATABLE-LINK: llvm-objcopy{{.*}}a.out --remove-section .llvm.offloading
+
+// RUN: clang-offload-packager -o %t.out \
+// RUN:   --image=file=%t.elf.o,kind=hip,triple=amdgcn-amd-amdhsa,arch=gfx90a \
+// RUN:   --image=file=%t.elf.o,kind=hip,triple=amdgcn-amd-amdhsa,arch=gfx90a
+// 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.lld -- -r %t.o \
+// RUN:   %t.o -o a.out 2>&1 | FileCheck %s --check-prefix=RELOCATABLE-LINK-HIP
+
+// RELOCATABLE-LINK-HIP: clang{{.*}} -o {{.*}}.img --target=amdgcn-amd-amdhsa
+// RELOCATABLE-LINK-HIP: clang-offload-bundler{{.*}} -type=o -bundle-align=4096 -targets=host-x86_64-unknown-linux,hipv4-amdgcn-amd-amdhsa--gfx90a -input=/dev/null -input={{.*}} -output={{.*}}
+// RELOCATABLE-LINK-HIP: /usr/bin/ld.lld{{.*}}-r
+// RELOCATABLE-LINK-HIP: llvm-objcopy{{.*}}a.out --remove-section .llvm.offloading
+
+// RUN: clang-offload-packager -o %t.out \
+// RUN:   --image=file=%t.elf.o,kind=cuda,triple=nvptx64-nvidia-cuda,arch=sm_89 \
+// RUN:   --image=file=%t.elf.o,kind=cuda,triple=nvptx64-nvidia-cuda,arch=sm_89
+// 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.lld -- -r %t.o \
+// RUN:   %t.o -o a.out 2>&1 | FileCheck %s --check-prefix=RELOCATABLE-LINK-CUDA
+
+// RELOCATABLE-LINK-CUDA: clang{{.*}} -o {{.*}}.img --target=nvptx64-nvidia-cuda
+// RELOCATABLE-LINK-CUDA: fatbinary{{.*}} -64 --create {{.*}}.fatbin --image=profile=sm_89,file={{.*}}.img
+// RELOCATABLE-LINK-CUDA: /usr/bin/ld.lld{{.*}}-r
+// RELOCATABLE-LINK-CUDA: llvm-objcopy{{.*}}a.out --remove-section .llvm.offloading

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

@@ -241,6 +241,70 @@ Expected<std::string> findProgram(StringRef Name, ArrayRef<StringRef> Paths) {
   return *Path;
 }
 
+/// Returns the hashed value for a constant string.
+std::string getHash(StringRef Str) {
+  llvm::MD5 Hasher;
+  llvm::MD5::MD5Result Hash;
+  Hasher.update(Str);
+  Hasher.final(Hash);
+  return llvm::utohexstr(Hash.low(), /*LowerCase=*/true);
+}
+
+/// Renames offloading entry sections in a relocatable link so they do not
+/// conflict with a later link job.
+Error relocateOffloadSection(const ArgList &Args, StringRef Output) {
+  llvm::Triple Triple(
+      Args.getLastArgValue(OPT_host_triple_EQ, sys::getDefaultTargetTriple()));
+  if (Triple.isOSWindows())
+    return createStringError(
+        inconvertibleErrorCode(),
+        "Relocatable linking is not supported on COFF targets");
+
+  Expected<std::string> ObjcopyPath =
+      findProgram("llvm-objcopy", {getMainExecutable("llvm-objcopy")});
+  if (!ObjcopyPath)
+    return ObjcopyPath.takeError();
+
+  // Use the linker output file to get a unique hash. This creates a unique
+  // identifier to rename the sections to that is deterministic to the contents.
+  auto BufferOrErr = DryRun ? MemoryBuffer::getMemBuffer("")
+                            : MemoryBuffer::getFileOrSTDIN(Output);
+  if (!BufferOrErr)
+    return createStringError(inconvertibleErrorCode(), "Failed to open %s",
+                             Output.str().c_str());
+  std::string Suffix = "_" + getHash((*BufferOrErr)->getBuffer());
+
+  SmallVector<StringRef> ObjcopyArgs = {
+      *ObjcopyPath,
+      Output,
+  };
+
+  // Remove the old .llvm.offloading section to prevent further linking.
+  ObjcopyArgs.emplace_back("--remove-section");
+  ObjcopyArgs.emplace_back(".llvm.offloading");
+  for (StringRef Prefix : {"omp", "cuda", "hip"}) {
+    auto Section = (Prefix + "_offloading_entries").str();
+    // Rename the offloading entires to make them private to this link unit.
+    ObjcopyArgs.emplace_back("--rename-section");
+    ObjcopyArgs.emplace_back(
+        Args.MakeArgString(Section + "=" + Section + Suffix));
+
+    // Rename the __start_ / __stop_ symbols appropriately to iterate over the
+    // newly renamed section containing the offloading entries.
+    ObjcopyArgs.emplace_back("--redefine-sym");
+    ObjcopyArgs.emplace_back(Args.MakeArgString("__start_" + Section + "=" +
+                                                "__start_" + Section + Suffix));
+    ObjcopyArgs.emplace_back("--redefine-sym");
+    ObjcopyArgs.emplace_back(Args.MakeArgString("__stop_" + Section + "=" +
+                                                "__stop_" + Section + Suffix));
+  }
+
+  if (Error Err = executeCommands(*ObjcopyPath, ObjcopyArgs))
+    return Err;
+
+  return Error::success();
+}
+
 /// Runs the wrapped linker job with the newly created input.
 Error runLinker(ArrayRef<StringRef> Files, const ArgList &Args) {
   llvm::TimeTraceScope TimeScope("Execute host linker");
@@ -265,6 +329,10 @@ Error runLinker(ArrayRef<StringRef> Files, const ArgList &Args) {
     LinkerArgs.push_back(Arg);
   if (Error Err = executeCommands(LinkerPath, LinkerArgs))
     return Err;
+
+  if (Args.hasArg(OPT_relocatable))
+    return relocateOffloadSection(Args, ExecutableName);
+
   return Error::success();
 }
 
@@ -910,7 +978,8 @@ wrapDeviceImages(ArrayRef<std::unique_ptr<MemoryBuffer>> Buffers,
   case OFK_OpenMP:
     if (Error Err = offloading::wrapOpenMPBinaries(
             M, BuffersToWrap,
-            offloading::getOffloadEntryArray(M, "omp_offloading_entries")))
+            offloading::getOffloadEntryArray(M, "omp_offloading_entries"),
+            /*Suffix=*/"", /*Relocatable=*/Args.hasArg(OPT_relocatable)))
       return std::move(Err);
     break;
   case OFK_Cuda:
@@ -1356,12 +1425,6 @@ Expected<SmallVector<SmallVector<OffloadFile>>>
 getDeviceInput(const ArgList &Args) {
   llvm::TimeTraceScope TimeScope("ExtractDeviceCode");
 
-  // If the user is requesting a reloctable link we ignore the device code. The
-  // actual linker will merge the embedded device code sections so they can be
-  // linked when the executable is finally created.
-  if (Args.hasArg(OPT_relocatable))
-    return SmallVector<SmallVector<OffloadFile>>{};
-
   StringRef Root = Args.getLastArgValue(OPT_sysroot_EQ);
   SmallVector<StringRef> LibraryPaths;
   for (const opt::Arg *Arg : Args.filtered(OPT_library_path, OPT_libpath))

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

@@ -127,8 +127,9 @@ 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 relocatable : Flag<["--", "-"], "relocatable">, Flags<[HelpHidden]>;
-def r : Flag<["-"], "r">, Alias<relocatable>, Flags<[HelpHidden]>;
+def relocatable : Flag<["--", "-"], "relocatable">, 
+    HelpText<"Link device code to create a relocatable offloading application">;
+def r : Flag<["-"], "r">, Alias<relocatable>;
 
 // link.exe-style linker options.
 def out : Joined<["/", "-", "/?", "-?"], "out:">, Flags<[HelpHidden]>;

llvm/include/llvm/Frontend/Offloading/OffloadWrapper.h

@@ -20,10 +20,13 @@ using EntryArrayTy = std::pair<GlobalVariable *, GlobalVariable *>;
 /// \param EntryArray Optional pair pointing to the `__start` and `__stop`
 /// symbols holding the `__tgt_offload_entry` array.
 /// \param Suffix An optional suffix appended to the emitted symbols.
+/// \param Relocatable Indicate if we need to change the offloading section to
+/// create a relocatable object.
 llvm::Error wrapOpenMPBinaries(llvm::Module &M,
                                llvm::ArrayRef<llvm::ArrayRef<char>> Images,
                                EntryArrayTy EntryArray,
-                               llvm::StringRef Suffix = "");
+                               llvm::StringRef Suffix = "",
+                               bool Relocatable = false);
 
 /// Wraps the input fatbinary image into the module \p M as global symbols and
 /// registers the images with the CUDA runtime.

llvm/lib/Frontend/Offloading/OffloadWrapper.cpp

@@ -112,7 +112,8 @@ PointerType *getBinDescPtrTy(Module &M) {
 ///
 /// Global variable that represents BinDesc is returned.
 GlobalVariable *createBinDesc(Module &M, ArrayRef<ArrayRef<char>> Bufs,
-                              EntryArrayTy EntryArray, StringRef Suffix) {
+                              EntryArrayTy EntryArray, StringRef Suffix,
+                              bool Relocatable) {
   LLVMContext &C = M.getContext();
   auto [EntriesB, EntriesE] = EntryArray;
 
@@ -129,7 +130,8 @@ GlobalVariable *createBinDesc(Module &M, ArrayRef<ArrayRef<char>> Bufs,
                                      GlobalVariable::InternalLinkage, Data,
                                      ".omp_offloading.device_image" + Suffix);
     Image->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
-    Image->setSection(".llvm.offloading");
+    Image->setSection(Relocatable ? ".llvm.offloading.relocatable"
+                                  : ".llvm.offloading");
     Image->setAlignment(Align(object::OffloadBinary::getAlignment()));
 
     StringRef Binary(Buf.data(), Buf.size());
@@ -582,8 +584,9 @@ void createRegisterFatbinFunction(Module &M, GlobalVariable *FatbinDesc,
 
 Error offloading::wrapOpenMPBinaries(Module &M, ArrayRef<ArrayRef<char>> Images,
                                      EntryArrayTy EntryArray,
-                                     llvm::StringRef Suffix) {
-  GlobalVariable *Desc = createBinDesc(M, Images, EntryArray, Suffix);
+                                     llvm::StringRef Suffix, bool Relocatable) {
+  GlobalVariable *Desc =
+      createBinDesc(M, Images, EntryArray, Suffix, Relocatable);
   if (!Desc)
     return createStringError(inconvertibleErrorCode(),
                              "No binary descriptors created.");

llvm/lib/Object/OffloadBinary.cpp

@@ -83,7 +83,7 @@ Error extractFromObject(const ObjectFile &Obj,
       if (!NameOrErr)
         return NameOrErr.takeError();
 
-      if (!NameOrErr->equals(".llvm.offloading"))
+      if (!NameOrErr->starts_with(".llvm.offloading"))
         continue;
     }