postpone wrapper generation

Author: ZuseZ4

compiler/rustc_codegen_llvm/src/builder.rs

@@ -4,6 +4,7 @@ use std::{iter, ptr};
 
 pub(crate) mod autodiff;
 pub(crate) mod gpu_offload;
+pub(crate) mod gpu_wrapper;
 
 use libc::{c_char, c_uint, size_t};
 use rustc_abi as abi;

compiler/rustc_codegen_llvm/src/builder/gpu_offload.rs

@@ -12,7 +12,7 @@ use crate::llvm::{self, Linkage};
 use crate::{LlvmCodegenBackend, SimpleCx, attributes};
 
 pub(crate) fn handle_gpu_code<'ll>(
-    _cgcx: &CodegenContext<LlvmCodegenBackend>,
+    cgcx: &CodegenContext<LlvmCodegenBackend>,
     cx: &'ll SimpleCx<'_>,
 ) {
     let (offload_entry_ty, at_one, begin, update, end, tgt_bin_desc, fn_ty) = gen_globals(&cx);
@@ -27,6 +27,7 @@ pub(crate) fn handle_gpu_code<'ll>(
         }
     }
     gen_call_handling(&cx, &kernels, at_one, begin, update, end, tgt_bin_desc, fn_ty, &o_types);
+    crate::builder::gpu_wrapper::gen_image_wrapper_module(&cgcx);
 }
 
 // The meaning of the __tgt_offload_entry (as per llvm docs) is

compiler/rustc_codegen_llvm/src/builder/gpu_wrapper.rs

@@ -0,0 +1,119 @@
+use std::ffi::CString;
+
+use llvm::Linkage::*;
+use rustc_abi::Align;
+use rustc_codegen_ssa::back::write::CodegenContext;
+use rustc_codegen_ssa::traits::BaseTypeCodegenMethods;
+
+use crate::builder::gpu_offload::*;
+use crate::llvm::{self, Visibility};
+use crate::{LlvmCodegenBackend, ModuleLlvm, SimpleCx};
+
+pub(crate) fn create_struct_ty<'ll>(
+    cx: &'ll SimpleCx<'_>,
+    name: &str,
+    tys: &[&'ll llvm::Type],
+) -> &'ll llvm::Type {
+    let entry_struct_name = CString::new(name).unwrap();
+    unsafe {
+        let entry_struct = llvm::LLVMStructCreateNamed(cx.llcx, entry_struct_name.as_ptr());
+        llvm::LLVMStructSetBody(entry_struct, tys.as_ptr(), tys.len() as u32, 0);
+        entry_struct
+    }
+}
+
+// We don't copy types from other functions because we generate a new module and context.
+// Bringing in types from other contexts would likely cause issues.
+pub(crate) fn gen_image_wrapper_module(cgcx: &CodegenContext<LlvmCodegenBackend>) {
+    let dl_cstr = CString::new("e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-p7:160:256:256:32-p8:128:128-p9:192:256:256:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1-ni:7:8:9").unwrap();
+    let target_cstr = CString::new("amdgcn-amd-amdhsa").unwrap();
+    let name = "offload.wrapper.module";
+    let m: crate::ModuleLlvm =
+        ModuleLlvm::new_simple(name, dl_cstr.into_raw(), target_cstr.into_raw(), &cgcx).unwrap();
+    let cx = SimpleCx::new(m.llmod(), m.llcx, cgcx.pointer_size);
+    let tptr = cx.type_ptr();
+    let ti64 = cx.type_i64();
+    let ti32 = cx.type_i32();
+    let ti16 = cx.type_i16();
+
+    let entry_fields = [ti64, ti16, ti16, ti32, tptr, tptr, ti64, ti64, tptr];
+    create_struct_ty(&cx, "__tgt_offload_entry", &entry_fields);
+    create_struct_ty(&cx, "__tgt_device_image", &[tptr, tptr, tptr, tptr]);
+    create_struct_ty(&cx, "__tgt_bin_desc", &[ti32, tptr, tptr, tptr]);
+
+    let offload_entry_ty = add_tgt_offload_entry(&cx);
+    let offload_entry_arr = cx.type_array(offload_entry_ty, 0);
+
+    let c_name = CString::new("__start_omp_offloading_entries").unwrap();
+    let llglobal = llvm::add_global(cx.llmod, offload_entry_arr, &c_name);
+    llvm::set_global_constant(llglobal, true);
+    llvm::set_linkage(llglobal, ExternalLinkage);
+    llvm::set_visibility(llglobal, Visibility::Hidden);
+    let c_name = CString::new("__stop_omp_offloading_entries").unwrap();
+    let llglobal = llvm::add_global(cx.llmod, offload_entry_arr, &c_name);
+    llvm::set_global_constant(llglobal, true);
+    llvm::set_linkage(llglobal, ExternalLinkage);
+    llvm::set_visibility(llglobal, Visibility::Hidden);
+
+    let c_name = CString::new("__dummy.omp_offloading_entries").unwrap();
+    let llglobal = llvm::add_global(cx.llmod, offload_entry_arr, &c_name);
+    llvm::set_global_constant(llglobal, true);
+    llvm::set_linkage(llglobal, InternalLinkage);
+    let c_section_name = CString::new("omp_offloading_entries").unwrap();
+    llvm::set_section(llglobal, &c_section_name);
+    let zeroinit = cx.const_null(offload_entry_arr);
+    llvm::set_initializer(llglobal, zeroinit);
+
+    CString::new("llvm.compiler.used").unwrap();
+    let arr_val = cx.const_array(tptr, &[llglobal]);
+    let c_section_name = CString::new("llvm.metadata").unwrap();
+    let llglobal = add_global(&cx, "llvm.compiler.used", arr_val, AppendingLinkage);
+    llvm::set_section(llglobal, &c_section_name);
+    llvm::set_global_constant(llglobal, false);
+
+    //@llvm.compiler.used = appending global [1 x ptr] [ptr @__dummy.omp_offloading_entries], section "llvm.metadata"
+
+    let mapper_fn_ty = cx.type_func(&[tptr], cx.type_void());
+    crate::declare::declare_simple_fn(
+        &cx,
+        &"__tgt_unregister_lib",
+        llvm::CallConv::CCallConv,
+        llvm::UnnamedAddr::No,
+        llvm::Visibility::Default,
+        mapper_fn_ty,
+    );
+    crate::declare::declare_simple_fn(
+        &cx,
+        &"__tgt_register_lib",
+        llvm::CallConv::CCallConv,
+        llvm::UnnamedAddr::No,
+        llvm::Visibility::Default,
+        mapper_fn_ty,
+    );
+    crate::declare::declare_simple_fn(
+        &cx,
+        &"atexit",
+        llvm::CallConv::CCallConv,
+        llvm::UnnamedAddr::No,
+        llvm::Visibility::Default,
+        cx.type_func(&[tptr], ti32),
+    );
+
+    let unknown_txt = "11111111111111";
+    let c_entry_name = CString::new(unknown_txt).unwrap();
+    let c_val = c_entry_name.as_bytes_with_nul();
+    let initializer = crate::common::bytes_in_context(cx.llcx, c_val);
+    let llglobal =
+        add_unnamed_global(&cx, &".omp_offloading.device_image", initializer, InternalLinkage);
+    let c_section_name = CString::new(".llvm.offloading").unwrap();
+    llvm::set_section(llglobal, &c_section_name);
+    llvm::set_alignment(llglobal, Align::EIGHT);
+
+    unsafe {
+        llvm::LLVMPrintModuleToFile(
+            cx.llmod,
+            CString::new("rustmagic.openmp.image.wrapper.ll").unwrap().as_ptr(),
+            std::ptr::null_mut(),
+        );
+    }
+}

compiler/rustc_codegen_llvm/src/context.rs

@@ -160,6 +160,23 @@ fn to_llvm_tls_model(tls_model: TlsModel) -> llvm::ThreadLocalMode {
     }
 }
 
+// FIXME(offload): This method is not relying on a tcx. We might still want to try to share some of
+// the logic with create_module, e.g. the target_data_layout handling.
+pub(crate) unsafe fn create_simple_module<'ll>(
+    llcx: &'ll llvm::Context,
+    target_data_layout: *const i8,
+    target_triple: *const i8,
+    mod_name: &str,
+) -> &'ll llvm::Module {
+    let mod_name = SmallCStr::new(mod_name);
+    let llmod = unsafe { llvm::LLVMModuleCreateWithNameInContext(mod_name.as_ptr(), llcx) };
+    unsafe {
+        llvm::LLVMSetDataLayout(llmod, target_data_layout);
+        llvm::LLVMSetTarget(llmod, target_triple);
+    }
+    llmod
+}
+
 pub(crate) unsafe fn create_module<'ll>(
     tcx: TyCtxt<'_>,
     llcx: &'ll llvm::Context,

compiler/rustc_codegen_llvm/src/lib.rs

@@ -399,6 +399,21 @@ unsafe impl Send for ModuleLlvm {}
 unsafe impl Sync for ModuleLlvm {}
 
 impl ModuleLlvm {
+    fn new_simple(
+        name: &str,
+        dl_cstr: *const i8,
+        target_cstr: *const i8,
+        cgcx: &CodegenContext<LlvmCodegenBackend>,
+    ) -> Result<Self, FatalError> {
+        unsafe {
+            let llcx = llvm::LLVMRustContextCreate(false);
+            let llmod_raw = context::create_simple_module(llcx, dl_cstr, target_cstr, name);
+            let dcx = cgcx.create_dcx();
+            let tm = ModuleLlvm::tm_from_cgcx(cgcx, name, dcx.handle())?;
+            Ok(ModuleLlvm { llmod_raw, llcx, tm: ManuallyDrop::new(tm) })
+        }
+    }
+
     fn new(tcx: TyCtxt<'_>, mod_name: &str) -> Self {
         unsafe {
             let llcx = llvm::LLVMRustContextCreate(tcx.sess.fewer_names());

compiler/rustc_codegen_llvm/src/llvm/ffi.rs

@@ -1009,12 +1009,19 @@ unsafe extern "C" {
         ModuleID: *const c_char,
         C: &Context,
     ) -> &Module;
+    pub(crate) fn LLVMPrintModuleToFile(
+        M: &Module,
+        Name: *const c_char,
+        Error_message: *mut c_char,
+    );
     pub(crate) fn LLVMCloneModule(M: &Module) -> &Module;
 
     /// Data layout. See Module::getDataLayout.
     pub(crate) fn LLVMGetDataLayoutStr(M: &Module) -> *const c_char;
     pub(crate) fn LLVMSetDataLayout(M: &Module, Triple: *const c_char);
 
+    pub(crate) fn LLVMSetTarget(M: &Module, Name: *const c_char);
+
     /// Append inline assembly to a module. See `Module::appendModuleInlineAsm`.
     pub(crate) fn LLVMAppendModuleInlineAsm(
         M: &Module,