Auto merge of #149170 - ZuseZ4:automate-offload-packager, r=<try>

automate gpu offloading - part 1

try-job: dist-ohos-aarch64

Author: rust-bors[bot]

compiler/rustc/Cargo.toml

@@ -38,6 +38,7 @@ check_only = ['rustc_driver_impl/check_only']
 jemalloc = ['dep:tikv-jemalloc-sys']
 llvm = ['rustc_driver_impl/llvm']
 llvm_enzyme = ['rustc_driver_impl/llvm_enzyme']
+llvm_offload = ['rustc_driver_impl/llvm_offload']
 max_level_info = ['rustc_driver_impl/max_level_info']
 rustc_randomized_layouts = ['rustc_driver_impl/rustc_randomized_layouts']
 # tidy-alphabetical-end

compiler/rustc_codegen_llvm/Cargo.toml

@@ -47,5 +47,6 @@ tracing = "0.1"
 # tidy-alphabetical-start
 check_only = ["rustc_llvm/check_only"]
 llvm_enzyme = []
+llvm_offload = []
 # tidy-alphabetical-end
 

compiler/rustc_codegen_llvm/src/back/write.rs

@@ -771,6 +771,13 @@ 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

@@ -1718,6 +1718,20 @@ unsafe extern "C" {
     ) -> &'a Value;
 }
 
+#[cfg(feature = "llvm_offload")]
+unsafe extern "C" {
+    /// 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;
+}
+
+#[cfg(not(feature = "llvm_offload"))]
+#[allow(unused_unsafe)]
+/// Processes the module and writes it in an offload compatible way into a "host.out" file.
+/// Marked as unsafe to match the real offload wrapper which is unsafe due to FFI.
+pub(crate) unsafe fn LLVMRustBundleImages<'a>(_M: &'a Module, _TM: &'a TargetMachine) -> bool {
+    unimplemented!("This rustc version was not built with LLVM Offload support!");
+}
+
 // FFI bindings for `DIBuilder` functions in the LLVM-C API.
 // Try to keep these in the same order as in `llvm/include/llvm-c/DebugInfo.h`.
 //

compiler/rustc_driver_impl/Cargo.toml

@@ -75,6 +75,7 @@ ctrlc = "3.4.4"
 check_only = ['rustc_interface/check_only']
 llvm = ['rustc_interface/llvm']
 llvm_enzyme = ['rustc_interface/llvm_enzyme']
+llvm_offload = ['rustc_interface/llvm_offload']
 max_level_info = ['rustc_log/max_level_info']
 rustc_randomized_layouts = [
     'rustc_index/rustc_randomized_layouts',

compiler/rustc_interface/Cargo.toml

@@ -59,4 +59,5 @@ rustc_abi = { path = "../rustc_abi" }
 check_only = ['rustc_codegen_llvm?/check_only']
 llvm = ['dep:rustc_codegen_llvm']
 llvm_enzyme = ['rustc_builtin_macros/llvm_enzyme', 'rustc_codegen_llvm/llvm_enzyme']
+llvm_offload = ['rustc_codegen_llvm/llvm_offload']
 # tidy-alphabetical-end

compiler/rustc_llvm/build.rs

@@ -214,6 +214,10 @@ fn main() {
         cfg.define("ENZYME", None);
     }
 
+    if tracked_env_var_os("LLVM_OFFLOAD").is_some() {
+        cfg.define("OFFLOAD", None);
+    }
+
     if tracked_env_var_os("LLVM_RUSTLLVM").is_some() {
         cfg.define("LLVM_RUSTLLVM", None);
     }

compiler/rustc_llvm/llvm-wrapper/RustWrapper.cpp

@@ -39,6 +39,14 @@
 #include "llvm/Transforms/Utils/ValueMapper.h"
 #include <iostream>
 
+// Some of the functions below rely on LLVM modules that may not always be
+// available. As such, we only try to build it in the first place, if
+// llvm.offload is enabled.
+#ifdef OFFLOAD
+#include "llvm/Object/OffloadBinary.h"
+#include "llvm/Target/TargetMachine.h"
+#endif
+
 // for raw `write` in the bad-alloc handler
 #ifdef _MSC_VER
 #include <io.h>
@@ -144,6 +152,55 @@ extern "C" void LLVMRustPrintStatistics(RustStringRef OutBuf) {
   llvm::PrintStatistics(OS);
 }
 
+// Some of the functions here rely on LLVM modules that may not always be
+// available. As such, we only try to build it in the first place, if
+// llvm.offload is enabled.
+#ifdef OFFLOAD
+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);
@@ -163,6 +220,7 @@ extern "C" void LLVMRustOffloadMapper(LLVMValueRef OldFn, LLVMValueRef NewFn) {
                           llvm::CloneFunctionChangeType::LocalChangesOnly,
                           returns);
 }
+#endif
 
 extern "C" LLVMValueRef LLVMRustGetNamedValue(LLVMModuleRef M, const char *Name,
                                               size_t NameLen) {

src/bootstrap/src/core/build_steps/compile.rs

@@ -1436,6 +1436,9 @@ fn rustc_llvm_env(builder: &Builder<'_>, cargo: &mut Cargo, target: TargetSelect
     if builder.config.llvm_enzyme {
         cargo.env("LLVM_ENZYME", "1");
     }
+    if builder.config.llvm_offload {
+        cargo.env("LLVM_OFFLOAD", "1");
+    }
     let llvm::LlvmResult { host_llvm_config, .. } = builder.ensure(llvm::Llvm { target });
     cargo.env("LLVM_CONFIG", &host_llvm_config);
 

src/bootstrap/src/lib.rs

@@ -850,6 +850,10 @@ impl Build {
             features.insert("llvm_enzyme");
         }
 
+        //if self.config.llvm_offload {
+        //    features.insert("llvm_offload");
+        //}
+
         features.into_iter().collect::<Vec<_>>().join(" ")
     }
 
@@ -873,6 +877,9 @@ impl Build {
         if self.config.llvm_enzyme {
             features.push("llvm_enzyme");
         }
+        if self.config.llvm_offload {
+            features.push("llvm_offload");
+        }
         // keep in sync with `bootstrap/compile.rs:rustc_cargo_env`
         if self.config.rust_randomize_layout && check("rustc_randomized_layouts") {
             features.push("rustc_randomized_layouts");