Rollup merge of rust-lang#149170 - ZuseZ4:automate-offload-packager, r=oli-obk

automate gpu offloading - part 1

Automates step 1 from the rustc-dev-guide offload section:
https://rustc-dev-guide.rust-lang.org/offload/usage.html#compile-instructions
`"clang-offload-packager" "-o" "host.out" "--image=file=device.bc,triple=amdgcn-amd-amdhsa,arch=gfx90a,kind=openmp"`

Verified on an MI 250X

cc ``@jhuber6,`` ``@kevinsala,`` ``@jdoerfert,`` ``@Sa4dUs``

r? oli-obk

Author: matthiaskrgr

compiler/rustc_codegen_llvm/src/back/write.rs

@@ -765,6 +765,13 @@ pub(crate) unsafe fn llvm_optimize(
             llvm_plugins.len(),
         )
     };
+
+    if cgcx.target_is_like_gpu && config.offload.contains(&config::Offload::Enable) {
+        unsafe {
+            llvm::LLVMRustBundleImages(module.module_llvm.llmod(), module.module_llvm.tm.raw());
+        }
+    }
+
     result.into_result().unwrap_or_else(|()| llvm_err(dcx, LlvmError::RunLlvmPasses))
 }
 

compiler/rustc_codegen_llvm/src/llvm/ffi.rs

@@ -1641,6 +1641,9 @@ unsafe extern "C" {
         Name: *const c_char,
     ) -> &'a Value;
 
+    /// Processes the module and writes it in an offload compatible way into a "host.out" file.
+    pub(crate) fn LLVMRustBundleImages<'a>(M: &'a Module, TM: &'a TargetMachine) -> bool;
+
     /// Writes a module to the specified path. Returns 0 on success.
     pub(crate) fn LLVMWriteBitcodeToFile(M: &Module, Path: *const c_char) -> c_int;
 

compiler/rustc_llvm/llvm-wrapper/RustWrapper.cpp

@@ -25,6 +25,7 @@
 #include "llvm/IR/Module.h"
 #include "llvm/IR/Value.h"
 #include "llvm/Object/COFFImportFile.h"
+#include "llvm/Object/OffloadBinary.h"
 #include "llvm/Remarks/RemarkFormat.h"
 #include "llvm/Remarks/RemarkSerializer.h"
 #include "llvm/Remarks/RemarkStreamer.h"
@@ -35,6 +36,7 @@
 #include "llvm/Support/Signals.h"
 #include "llvm/Support/Timer.h"
 #include "llvm/Support/ToolOutputFile.h"
+#include "llvm/Target/TargetMachine.h"
 #include "llvm/Transforms/Utils/Cloning.h"
 #include "llvm/Transforms/Utils/ValueMapper.h"
 #include <iostream>
@@ -144,6 +146,51 @@ extern "C" void LLVMRustPrintStatistics(RustStringRef OutBuf) {
   llvm::PrintStatistics(OS);
 }
 
+static Error writeFile(StringRef Filename, StringRef Data) {
+  Expected<std::unique_ptr<FileOutputBuffer>> OutputOrErr =
+      FileOutputBuffer::create(Filename, Data.size());
+  if (!OutputOrErr)
+    return OutputOrErr.takeError();
+  std::unique_ptr<FileOutputBuffer> Output = std::move(*OutputOrErr);
+  llvm::copy(Data, Output->getBufferStart());
+  if (Error E = Output->commit())
+    return E;
+  return Error::success();
+}
+
+// This is the first of many steps in creating a binary using llvm offload,
+// to run code on the gpu. Concrete, it replaces the following binary use:
+// clang-offload-packager -o host.out
+//  --image=file=device.bc,triple=amdgcn-amd-amdhsa,arch=gfx90a,kind=openmp
+// The input module is the rust code compiled for a gpu target like amdgpu.
+// Based on clang/tools/clang-offload-packager/ClangOffloadPackager.cpp
+extern "C" bool LLVMRustBundleImages(LLVMModuleRef M, TargetMachine &TM) {
+  std::string Storage;
+  llvm::raw_string_ostream OS1(Storage);
+  llvm::WriteBitcodeToFile(*unwrap(M), OS1);
+  OS1.flush();
+  auto MB = llvm::MemoryBuffer::getMemBufferCopy(Storage, "module.bc");
+
+  SmallVector<char, 1024> BinaryData;
+  raw_svector_ostream OS2(BinaryData);
+
+  OffloadBinary::OffloadingImage ImageBinary{};
+  ImageBinary.TheImageKind = object::IMG_Bitcode;
+  ImageBinary.Image = std::move(MB);
+  ImageBinary.TheOffloadKind = object::OFK_OpenMP;
+  ImageBinary.StringData["triple"] = TM.getTargetTriple().str();
+  ImageBinary.StringData["arch"] = TM.getTargetCPU();
+  llvm::SmallString<0> Buffer = OffloadBinary::write(ImageBinary);
+  if (Buffer.size() % OffloadBinary::getAlignment() != 0)
+    // Offload binary has invalid size alignment
+    return false;
+  OS2 << Buffer;
+  if (Error E = writeFile("host.out",
+                          StringRef(BinaryData.begin(), BinaryData.size())))
+    return false;
+  return true;
+}
+
 extern "C" void LLVMRustOffloadMapper(LLVMValueRef OldFn, LLVMValueRef NewFn) {
   llvm::Function *oldFn = llvm::unwrap<llvm::Function>(OldFn);
   llvm::Function *newFn = llvm::unwrap<llvm::Function>(NewFn);

src/doc/rustc-dev-guide/src/offload/usage.md

@@ -79,27 +79,17 @@ Now we generate the device code. Replace the target-cpu with the right code for
 RUSTFLAGS="-Ctarget-cpu=gfx90a --emit=llvm-bc,llvm-ir -Zoffload=Enable -Zunstable-options" cargo +offload build -Zunstable-options -r -v --target amdgcn-amd-amdhsa -Zbuild-std=core
 ```
 
-Now find the `<libname>.ll` under target/amdgcn-amd-amdhsa folder and copy it to a device.ll file (or adjust the file names below).
-If you work on an NVIDIA or Intel gpu, please adjust the names acordingly and open an issue to share your results (either if you succeed or fail).
-First we compile our .ll files (good for manual inspections) to .bc files and clean up leftover artifacts. The cleanup is important, otherwise caching might interfere on following runs.
-```
-opt lib.ll -o lib.bc
-opt device.ll -o device.bc
-rm *.o
-rm bare.amdgcn.gfx90a.img*
-```
 
 ```
-"clang-offload-packager" "-o" "host.out" "--image=file=device.bc,triple=amdgcn-amd-amdhsa,arch=gfx90a,kind=openmp"
-
 "clang-21" "-cc1" "-triple" "x86_64-unknown-linux-gnu" "-S" "-save-temps=cwd" "-disable-free" "-clear-ast-before-backend" "-main-file-name" "lib.rs" "-mrelocation-model" "pic" "-pic-level" "2" "-pic-is-pie" "-mframe-pointer=all" "-fmath-errno" "-ffp-contract=on" "-fno-rounding-math" "-mconstructor-aliases" "-funwind-tables=2" "-target-cpu" "x86-64" "-tune-cpu" "generic" "-resource-dir" "/<ABSOLUTE_PATH_TO>/rust/build/x86_64-unknown-linux-gnu/llvm/lib/clang/21" "-ferror-limit" "19" "-fopenmp" "-fopenmp-offload-mandatory" "-fgnuc-version=4.2.1" "-fskip-odr-check-in-gmf" "-fembed-offload-object=host.out" "-fopenmp-targets=amdgcn-amd-amdhsa" "-faddrsig" "-D__GCC_HAVE_DWARF2_CFI_ASM=1" "-o" "host.s" "-x" "ir" "lib.bc"
 
 "clang-21" "-cc1as" "-triple" "x86_64-unknown-linux-gnu" "-filetype" "obj" "-main-file-name" "lib.rs" "-target-cpu" "x86-64" "-mrelocation-model" "pic" "-o" "host.o" "host.s"
 
 "clang-linker-wrapper" "--should-extract=gfx90a" "--device-compiler=amdgcn-amd-amdhsa=-g" "--device-compiler=amdgcn-amd-amdhsa=-save-temps=cwd" "--device-linker=amdgcn-amd-amdhsa=-lompdevice" "--host-triple=x86_64-unknown-linux-gnu" "--save-temps" "--linker-path=/ABSOlUTE_PATH_TO/rust/build/x86_64-unknown-linux-gnu/lld/bin/ld.lld" "--hash-style=gnu" "--eh-frame-hdr" "-m" "elf_x86_64" "-pie" "-dynamic-linker" "/lib64/ld-linux-x86-64.so.2" "-o" "bare" "/lib/../lib64/Scrt1.o" "/lib/../lib64/crti.o" "/ABSOLUTE_PATH_TO/crtbeginS.o" "-L/ABSOLUTE_PATH_TO/rust/build/x86_64-unknown-linux-gnu/llvm/bin/../lib/x86_64-unknown-linux-gnu" "-L/ABSOLUTE_PATH_TO/rust/build/x86_64-unknown-linux-gnu/llvm/lib/clang/21/lib/x86_64-unknown-linux-gnu" "-L/lib/../lib64" "-L/usr/lib64" "-L/lib" "-L/usr/lib" "host.o" "-lstdc++" "-lm" "-lomp" "-lomptarget" "-L/ABSOLUTE_PATH_TO/rust/build/x86_64-unknown-linux-gnu/llvm/lib" "-lgcc_s" "-lgcc" "-lpthread" "-lc" "-lgcc_s" "-lgcc" "/ABSOLUTE_PATH_TO/crtendS.o" "/lib/../lib64/crtn.o"
 ```
 
-Especially for the last command I recommend to not fix the paths, but rather just re-generate them by copying a bare-mode openmp example and compiling it with your clang. By adding `-###` to your clang invocation, you can see the invidual steps.
+Especially for the last three commands I recommend to not fix the paths, but rather just re-generate them by copying a bare-mode openmp example and compiling it with your clang. By adding `-###` to your clang invocation, you can see the invidual steps.
+You can ignore other steps, e.g. the invocation of a "clang-offload-packager".
 ```
 myclang++ -fuse-ld=lld -O3 -fopenmp  -fopenmp-offload-mandatory --offload-arch=gfx90a omp_bare.cpp -o main -###
 ```