Add scalar support for both host and device

compiler/rustc_codegen_llvm/src/back/write.rs

@@ -13,6 +13,7 @@ use rustc_codegen_ssa::back::write::{
     TargetMachineFactoryConfig, TargetMachineFactoryFn,
 };
 use rustc_codegen_ssa::base::wants_wasm_eh;
+use rustc_codegen_ssa::common::TypeKind;
 use rustc_codegen_ssa::traits::*;
 use rustc_codegen_ssa::{CompiledModule, ModuleCodegen, ModuleKind};
 use rustc_data_structures::profiling::SelfProfilerRef;
@@ -33,6 +34,8 @@ use crate::back::owned_target_machine::OwnedTargetMachine;
 use crate::back::profiling::{
     LlvmSelfProfiler, selfprofile_after_pass_callback, selfprofile_before_pass_callback,
 };
+use crate::builder::SBuilder;
+use crate::builder::gpu_offload::scalar_width;
 use crate::common::AsCCharPtr;
 use crate::errors::{
     CopyBitcode, FromLlvmDiag, FromLlvmOptimizationDiag, LlvmError, UnknownCompression,
@@ -669,7 +672,17 @@ pub(crate) unsafe fn llvm_optimize(
         // Create the new parameter list, with ptr as the first argument
         let mut new_param_types = Vec::with_capacity(old_param_count as usize + 1);
         new_param_types.push(cx.type_ptr());
-        new_param_types.extend(old_param_types);
+
+        // This relies on undocumented LLVM knowledge that scalars must be passed as i64
+        for &old_ty in &old_param_types {
+            let new_ty = match cx.type_kind(old_ty) {
+                TypeKind::Half | TypeKind::Float | TypeKind::Double | TypeKind::Integer => {
+                    cx.type_i64()
+                }
+                _ => old_ty,
+            };
+            new_param_types.push(new_ty);
+        }
 
         // Create the new function type
         let ret_ty = unsafe { llvm::LLVMGetReturnType(old_fn_ty) };
@@ -682,10 +695,33 @@ pub(crate) unsafe fn llvm_optimize(
         let a0 = llvm::get_param(new_fn, 0);
         llvm::set_value_name(a0, CString::new("dyn_ptr").unwrap().as_bytes());
 
+        let bb = SBuilder::append_block(cx, new_fn, "entry");
+        let mut builder = SBuilder::build(cx, bb);
+
+        let mut old_args_rebuilt = Vec::with_capacity(old_param_types.len());
+
+        for (i, &old_ty) in old_param_types.iter().enumerate() {
+            let new_arg = llvm::get_param(new_fn, (i + 1) as u32);
+
+            let rebuilt = match cx.type_kind(old_ty) {
+                TypeKind::Half | TypeKind::Float | TypeKind::Double | TypeKind::Integer => {
+                    let num_bits = scalar_width(cx, old_ty);
+
+                    let trunc = builder.trunc(new_arg, cx.type_ix(num_bits));
+                    builder.bitcast(trunc, old_ty)
+                }
+                _ => new_arg,
+            };
+
+            old_args_rebuilt.push(rebuilt);
+        }
+
+        builder.ret_void();
+
         // Here we map the old arguments to the new arguments, with an offset of 1 to make sure
         // that we don't use the newly added `%dyn_ptr`.
         unsafe {
-            llvm::LLVMRustOffloadMapper(old_fn, new_fn);
+            llvm::LLVMRustOffloadMapper(old_fn, new_fn, old_args_rebuilt.as_ptr());
         }
 
         llvm::set_linkage(new_fn, llvm::get_linkage(old_fn));

compiler/rustc_codegen_llvm/src/builder.rs

@@ -97,6 +97,21 @@ impl<'a, 'll, CX: Borrow<SCx<'ll>>> GenericBuilder<'a, 'll, CX> {
         GenericBuilder { llbuilder, cx: scx }
     }
 
+    pub(crate) fn append_block(
+        cx: &'a GenericCx<'ll, CX>,
+        llfn: &'ll Value,
+        name: &str,
+    ) -> &'ll BasicBlock {
+        unsafe {
+            let name = SmallCStr::new(name);
+            llvm::LLVMAppendBasicBlockInContext(cx.llcx(), llfn, name.as_ptr())
+        }
+    }
+
+    pub(crate) fn trunc(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value {
+        unsafe { llvm::LLVMBuildTrunc(self.llbuilder, val, dest_ty, UNNAMED) }
+    }
+
     pub(crate) fn bitcast(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value {
         unsafe { llvm::LLVMBuildBitCast(self.llbuilder, val, dest_ty, UNNAMED) }
     }

compiler/rustc_codegen_llvm/src/builder/gpu_offload.rs

@@ -2,6 +2,7 @@ use std::ffi::CString;
 
 use llvm::Linkage::*;
 use rustc_abi::Align;
+use rustc_codegen_ssa::common::TypeKind;
 use rustc_codegen_ssa::mir::operand::{OperandRef, OperandValue};
 use rustc_codegen_ssa::traits::{BaseTypeCodegenMethods, BuilderMethods};
 use rustc_middle::bug;
@@ -357,7 +358,6 @@ pub(crate) fn add_global<'ll>(
 pub(crate) fn gen_define_handling<'ll>(
     cx: &CodegenCx<'ll, '_>,
     metadata: &[OffloadMetadata],
-    types: &[&'ll Type],
     symbol: String,
     offload_globals: &OffloadGlobals<'ll>,
 ) -> OffloadKernelGlobals<'ll> {
@@ -367,25 +367,18 @@ pub(crate) fn gen_define_handling<'ll>(
 
     let offload_entry_ty = offload_globals.offload_entry_ty;
 
-    // It seems like non-pointer values are automatically mapped. So here, we focus on pointer (or
-    // reference) types.
-    let ptr_meta = types.iter().zip(metadata).filter_map(|(&x, meta)| match cx.type_kind(x) {
-        rustc_codegen_ssa::common::TypeKind::Pointer => Some(meta),
-        _ => None,
-    });
-
     // FIXME(Sa4dUs): add `OMP_MAP_TARGET_PARAM = 0x20` only if necessary
-    let (ptr_sizes, ptr_transfer): (Vec<_>, Vec<_>) =
-        ptr_meta.map(|m| (m.payload_size, m.mode.bits() | 0x20)).unzip();
+    let (sizes, transfer): (Vec<_>, Vec<_>) =
+        metadata.iter().map(|m| (m.payload_size, m.mode.bits() | 0x20)).unzip();
 
-    let offload_sizes = add_priv_unnamed_arr(&cx, &format!(".offload_sizes.{symbol}"), &ptr_sizes);
+    let offload_sizes = add_priv_unnamed_arr(&cx, &format!(".offload_sizes.{symbol}"), &sizes);
     // Here we figure out whether something needs to be copied to the gpu (=1), from the gpu (=2),
     // or both to and from the gpu (=3). Other values shouldn't affect us for now.
     // A non-mutable reference or pointer will be 1, an array that's not read, but fully overwritten
     // will be 2. For now, everything is 3, until we have our frontend set up.
     // 1+2+32: 1 (MapTo), 2 (MapFrom), 32 (Add one extra input ptr per function, to be used later).
     let memtransfer_types =
-        add_priv_unnamed_arr(&cx, &format!(".offload_maptypes.{symbol}"), &ptr_transfer);
+        add_priv_unnamed_arr(&cx, &format!(".offload_maptypes.{symbol}"), &transfer);
 
     // Next: For each function, generate these three entries. A weak constant,
     // the llvm.rodata entry name, and  the llvm_offload_entries value
@@ -441,11 +434,24 @@ fn declare_offload_fn<'ll>(
     )
 }
 
+pub(crate) fn scalar_width<'ll>(cx: &'ll SimpleCx<'_>, ty: &'ll Type) -> u64 {
+    match cx.type_kind(ty) {
+        TypeKind::Half
+        | TypeKind::Float
+        | TypeKind::Double
+        | TypeKind::X86_FP80
+        | TypeKind::FP128
+        | TypeKind::PPC_FP128 => cx.float_width(ty) as u64,
+        TypeKind::Integer => cx.int_width(ty),
+        other => bug!("scalar_width was called on a non scalar type {other:?}"),
+    }
+}
+
 // For each kernel *call*, we now use some of our previous declared globals to move data to and from
 // the gpu. For now, we only handle the data transfer part of it.
 // If two consecutive kernels use the same memory, we still move it to the host and back to the gpu.
 // Since in our frontend users (by default) don't have to specify data transfer, this is something
-// we should optimize in the future! In some cases we can directly zero-allocate ont he device and
+// we should optimize in the future! In some cases we can directly zero-allocate on the device and
 // only move data back, or if something is immutable, we might only copy it to the device.
 //
 // Current steps:
@@ -533,8 +539,34 @@ pub(crate) fn gen_call_handling<'ll, 'tcx>(
     let mut geps = vec![];
     let i32_0 = cx.get_const_i32(0);
     for &v in args {
-        let gep = builder.inbounds_gep(cx.type_f32(), v, &[i32_0]);
-        vals.push(v);
+        let ty = cx.val_ty(v);
+        let ty_kind = cx.type_kind(ty);
+        let (base_val, gep_base) = match ty_kind {
+            TypeKind::Pointer => (v, v),
+            TypeKind::Half | TypeKind::Float | TypeKind::Double | TypeKind::Integer => {
+                // FIXME(Sa4dUs): check for `f128` support, latest NVIDIA cards support it
+                let num_bits = scalar_width(cx, ty);
+
+                let bb = builder.llbb();
+                unsafe {
+                    llvm::LLVMRustPositionBuilderPastAllocas(builder.llbuilder, builder.llfn());
+                }
+                let addr = builder.direct_alloca(cx.type_i64(), Align::EIGHT, "addr");
+                unsafe {
+                    llvm::LLVMPositionBuilderAtEnd(builder.llbuilder, bb);
+                }
+
+                let cast = builder.bitcast(v, cx.type_ix(num_bits));
+                let value = builder.zext(cast, cx.type_i64());
+                builder.store(value, addr, Align::EIGHT);
+                (value, addr)
+            }
+            other => bug!("offload does not support {other:?}"),
+        };
+
+        let gep = builder.inbounds_gep(cx.type_f32(), gep_base, &[i32_0]);
+
+        vals.push(base_val);
         geps.push(gep);
     }
 

compiler/rustc_codegen_llvm/src/intrinsic.rs

@@ -1388,7 +1388,7 @@ fn codegen_offload<'ll, 'tcx>(
     let args = get_args_from_tuple(bx, args[3], fn_target);
     let target_symbol = symbol_name_for_instance_in_crate(tcx, fn_target, LOCAL_CRATE);
 
-    let sig = tcx.fn_sig(fn_target.def_id()).instantiate_identity();
+    let sig = tcx.fn_sig(fn_target.def_id()).skip_binder();
     let sig = tcx.instantiate_bound_regions_with_erased(sig);
     let inputs = sig.inputs();
 
@@ -1404,7 +1404,7 @@ fn codegen_offload<'ll, 'tcx>(
             return;
         }
     };
-    let offload_data = gen_define_handling(&cx, &metadata, &types, target_symbol, offload_globals);
+    let offload_data = gen_define_handling(&cx, &metadata, target_symbol, offload_globals);
     gen_call_handling(bx, &offload_data, &args, &types, &metadata, offload_globals, &offload_dims);
 }
 

compiler/rustc_codegen_llvm/src/llvm/ffi.rs

@@ -1675,7 +1675,11 @@ mod Offload {
             _M: &'a Module,
             _host_out: *const c_char,
         ) -> bool;
-        pub(crate) fn LLVMRustOffloadMapper<'a>(OldFn: &'a Value, NewFn: &'a Value);
+        pub(crate) fn LLVMRustOffloadMapper<'a>(
+            OldFn: &'a Value,
+            NewFn: &'a Value,
+            RebuiltArgs: *const &Value,
+        );
     }
 }
 
@@ -1702,7 +1706,11 @@ mod Offload_fallback {
         unimplemented!("This rustc version was not built with LLVM Offload support!");
     }
     #[allow(unused_unsafe)]
-    pub(crate) unsafe fn LLVMRustOffloadMapper<'a>(_OldFn: &'a Value, _NewFn: &'a Value) {
+    pub(crate) unsafe fn LLVMRustOffloadMapper<'a>(
+        _OldFn: &'a Value,
+        _NewFn: &'a Value,
+        _RebuiltArgs: *const &Value,
+    ) {
         unimplemented!("This rustc version was not built with LLVM Offload support!");
     }
 }

compiler/rustc_llvm/llvm-wrapper/RustWrapper.cpp

@@ -223,7 +223,12 @@ extern "C" bool LLVMRustOffloadEmbedBufferInModule(LLVMModuleRef HostM,
   return true;
 }
 
-extern "C" void LLVMRustOffloadMapper(LLVMValueRef OldFn, LLVMValueRef NewFn) {
+// Clone OldFn into NewFn, remapping its arguments to RebuiltArgs.
+// Each arg of OldFn is replaced with the corresponding value in RebuiltArgs.
+// For scalars, RebuiltArgs contains the value cast and/or truncated to the
+// original type.
+extern "C" void LLVMRustOffloadMapper(LLVMValueRef OldFn, LLVMValueRef NewFn,
+                                      const LLVMValueRef *RebuiltArgs) {
   llvm::Function *oldFn = llvm::unwrap<llvm::Function>(OldFn);
   llvm::Function *newFn = llvm::unwrap<llvm::Function>(NewFn);
 
@@ -232,15 +237,25 @@ extern "C" void LLVMRustOffloadMapper(LLVMValueRef OldFn, LLVMValueRef NewFn) {
   llvm::ValueToValueMapTy vmap;
   auto newArgIt = newFn->arg_begin();
   newArgIt->setName("dyn_ptr");
-  ++newArgIt; // skip %dyn_ptr
+
+  unsigned i = 0;
   for (auto &oldArg : oldFn->args()) {
-    vmap[&oldArg] = &*newArgIt++;
+    vmap[&oldArg] = unwrap<Value>(RebuiltArgs[i++]);
   }
 
   llvm::SmallVector<llvm::ReturnInst *, 8> returns;
   llvm::CloneFunctionInto(newFn, oldFn, vmap,
                           llvm::CloneFunctionChangeType::LocalChangesOnly,
                           returns);
+
+  BasicBlock &entry = newFn->getEntryBlock();
+  BasicBlock &clonedEntry = *std::next(newFn->begin());
+
+  if (entry.getTerminator())
+    entry.getTerminator()->eraseFromParent();
+
+  IRBuilder<> B(&entry);
+  B.CreateBr(&clonedEntry);
 }
 #endif
 

compiler/rustc_middle/src/ty/offload_meta.rs

@@ -78,16 +78,13 @@ impl MappingFlags {
         use rustc_ast::Mutability::*;
 
         match ty.kind() {
-            ty::Bool
-            | ty::Char
-            | ty::Int(_)
-            | ty::Uint(_)
-            | ty::Float(_)
-            | ty::Adt(_, _)
-            | ty::Tuple(_)
-            | ty::Array(_, _)
-            | ty::Alias(_, _)
-            | ty::Param(_) => MappingFlags::TO,
+            ty::Bool | ty::Char | ty::Int(_) | ty::Uint(_) | ty::Float(_) => {
+                MappingFlags::LITERAL | MappingFlags::IMPLICIT
+            }
+
+            ty::Adt(_, _) | ty::Tuple(_) | ty::Array(_, _) | ty::Alias(_, _) | ty::Param(_) => {
+                MappingFlags::TO
+            }
 
             ty::RawPtr(_, Not) | ty::Ref(_, _, Not) => MappingFlags::TO,
 

tests/codegen-llvm/gpu_offload/scalar_device.rs

@@ -0,0 +1,36 @@
+//@ add-minicore
+//@ revisions: amdgpu nvptx
+//@[nvptx] compile-flags: -Copt-level=0 -Zunstable-options -Zoffload=Device --target nvptx64-nvidia-cuda --crate-type=rlib
+//@[nvptx] needs-llvm-components: nvptx
+//@[amdgpu] compile-flags: -Copt-level=0 -Zunstable-options -Zoffload=Device --target amdgcn-amd-amdhsa -Ctarget-cpu=gfx900 --crate-type=rlib
+//@[amdgpu] needs-llvm-components: amdgpu
+//@ no-prefer-dynamic
+//@ needs-offload
+
+// This test verifies that the offload intrinsic is properly handling scalar args on the device,
+// replacing the args by i64 and then trunc and cast them to the original type
+
+#![feature(abi_gpu_kernel, rustc_attrs, no_core)]
+#![no_core]
+
+extern crate minicore;
+
+// CHECK: ; Function Attrs
+// nvptx-NEXT: define ptx_kernel void @foo(ptr %dyn_ptr, ptr %0, i64 %1)
+// amdgpu-NEXT: define amdgpu_kernel void @foo(ptr %dyn_ptr, ptr %0, i64 %1)
+// CHECK-NEXT: entry:
+// CHECK-NEXT: %2 = trunc i64 %1 to i32
+// CHECK-NEXT: %3 = bitcast i32 %2 to float
+// CHECK-NEXT: br label %start
+// CHECK: start:
+// CHECK-NEXT: store float %3, ptr %0, align 4
+// CHECK-NEXT: ret void
+// CHECK-NEXT: }
+
+#[unsafe(no_mangle)]
+#[rustc_offload_kernel]
+pub unsafe extern "gpu-kernel" fn foo(x: *mut f32, k: f32) {
+    unsafe {
+        *x = k;
+    };
+}

tests/codegen-llvm/gpu_offload/scalar_host.rs

@@ -0,0 +1,37 @@
+//@ compile-flags: -Zoffload=Test -Zunstable-options -C opt-level=1 -Clto=fat
+//@ no-prefer-dynamic
+//@ needs-offload
+
+// This test verifies that the offload intrinsic is properly handling scalar args, passing them to
+// the kernel as i64
+
+#![feature(abi_gpu_kernel)]
+#![feature(rustc_attrs)]
+#![feature(core_intrinsics)]
+#![no_main]
+
+// CHECK: define{{( dso_local)?}} void @main()
+// CHECK-NOT: define
+// CHECK: %addr = alloca i64, align 8
+// CHECK: store double 4.200000e+01, ptr %0, align 8
+// CHECK: %_0.i = load double, ptr %0, align 8
+// CHECK: store double %_0.i, ptr %addr, align 8
+// CHECK: %1 = getelementptr inbounds nuw i8, ptr %.offload_baseptrs, i64 8
+// CHECK-NEXT: store double %_0.i, ptr %1, align 8
+// CHECK-NEXT: %2 = getelementptr inbounds nuw i8, ptr %.offload_ptrs, i64 8
+// CHECK-NEXT: store ptr %addr, ptr %2, align 8
+// CHECK-NEXT: %3 = getelementptr inbounds nuw i8, ptr %.offload_sizes, i64 8
+// CHECK-NEXT: store i64 4, ptr %3, align 8
+// CHECK-NEXT: call void @__tgt_target_data_begin_mapper
+
+#[unsafe(no_mangle)]
+fn main() {
+    let mut x = 0.0;
+    let k = core::hint::black_box(42.0);
+
+    core::intrinsics::offload::<_, _, ()>(foo, [1, 1, 1], [1, 1, 1], (&mut x, k));
+}
+
+unsafe extern "C" {
+    pub fn foo(x: *mut f32, k: f32);
+}