Unrolled build for #142097

Rollup merge of #142097 - ZuseZ4:offload-host1, r=oli-obk

gpu offload host code generation

r? ghost

This will generate most of the host side code to use llvm's offload feature.
The first PR will only handle automatic mem-transfers to and from the device.
So if a user calls a kernel, we will copy inputs back and forth, but we won't do the actual kernel launch.
Before merging, we will use LLVM's Info infrastructure to verify that the memcopies match what openmp offloa generates in C++. `LIBOMPTARGET_INFO=-1 ./my_rust_binary` should print that a memcpy to and later from the device is happening.

A follow-up PR will generate the actual device-side kernel which will then do computations on the GPU.
A third PR will implement manual host2device and device2host functionality, but the goal is to minimize cases where a user has to overwrite our default handling due to performance issues.

I'm trying to get a full MVP out first, so this just recognizes GPU functions based on magic names. The final frontend will obviously move this over to use proper macros, like I'm already doing it for the autodiff work.
This work will also be compatible with std::autodiff, so one can differentiate GPU kernels.

Tracking:
- #131513

Author: rust-timer

compiler/rustc_codegen_llvm/src/back/lto.rs

@@ -654,6 +654,7 @@ pub(crate) fn run_pass_manager(
     // We then run the llvm_optimize function a second time, to optimize the code which we generated
     // in the enzyme differentiation pass.
     let enable_ad = config.autodiff.contains(&config::AutoDiff::Enable);
+    let enable_gpu = config.offload.contains(&config::Offload::Enable);
     let stage = if thin {
         write::AutodiffStage::PreAD
     } else {
@@ -668,6 +669,12 @@ pub(crate) fn run_pass_manager(
         write::llvm_optimize(cgcx, dcx, module, None, config, opt_level, opt_stage, stage)?;
     }
 
+    if enable_gpu && !thin {
+        let cx =
+            SimpleCx::new(module.module_llvm.llmod(), &module.module_llvm.llcx, cgcx.pointer_size);
+        crate::builder::gpu_offload::handle_gpu_code(cgcx, &cx);
+    }
+
     if cfg!(llvm_enzyme) && enable_ad && !thin {
         let cx =
             SimpleCx::new(module.module_llvm.llmod(), &module.module_llvm.llcx, cgcx.pointer_size);

compiler/rustc_codegen_llvm/src/builder.rs

@@ -3,6 +3,7 @@ use std::ops::Deref;
 use std::{iter, ptr};
 
 pub(crate) mod autodiff;
+pub(crate) mod gpu_offload;
 
 use libc::{c_char, c_uint, size_t};
 use rustc_abi as abi;
@@ -117,6 +118,74 @@ impl<'a, 'll, CX: Borrow<SCx<'ll>>> GenericBuilder<'a, 'll, CX> {
         }
         bx
     }
+
+    // The generic builder has less functionality and thus (unlike the other alloca) we can not
+    // easily jump to the beginning of the function to place our allocas there. We trust the user
+    // to manually do that. FIXME(offload): improve the genericCx and add more llvm wrappers to
+    // handle this.
+    pub(crate) fn direct_alloca(&mut self, ty: &'ll Type, align: Align, name: &str) -> &'ll Value {
+        let val = unsafe {
+            let alloca = llvm::LLVMBuildAlloca(self.llbuilder, ty, UNNAMED);
+            llvm::LLVMSetAlignment(alloca, align.bytes() as c_uint);
+            // Cast to default addrspace if necessary
+            llvm::LLVMBuildPointerCast(self.llbuilder, alloca, self.cx.type_ptr(), UNNAMED)
+        };
+        if name != "" {
+            let name = std::ffi::CString::new(name).unwrap();
+            llvm::set_value_name(val, &name.as_bytes());
+        }
+        val
+    }
+
+    pub(crate) fn inbounds_gep(
+        &mut self,
+        ty: &'ll Type,
+        ptr: &'ll Value,
+        indices: &[&'ll Value],
+    ) -> &'ll Value {
+        unsafe {
+            llvm::LLVMBuildGEPWithNoWrapFlags(
+                self.llbuilder,
+                ty,
+                ptr,
+                indices.as_ptr(),
+                indices.len() as c_uint,
+                UNNAMED,
+                GEPNoWrapFlags::InBounds,
+            )
+        }
+    }
+
+    pub(crate) fn store(&mut self, val: &'ll Value, ptr: &'ll Value, align: Align) -> &'ll Value {
+        debug!("Store {:?} -> {:?}", val, ptr);
+        assert_eq!(self.cx.type_kind(self.cx.val_ty(ptr)), TypeKind::Pointer);
+        unsafe {
+            let store = llvm::LLVMBuildStore(self.llbuilder, val, ptr);
+            llvm::LLVMSetAlignment(store, align.bytes() as c_uint);
+            store
+        }
+    }
+
+    pub(crate) fn load(&mut self, ty: &'ll Type, ptr: &'ll Value, align: Align) -> &'ll Value {
+        unsafe {
+            let load = llvm::LLVMBuildLoad2(self.llbuilder, ty, ptr, UNNAMED);
+            llvm::LLVMSetAlignment(load, align.bytes() as c_uint);
+            load
+        }
+    }
+
+    fn memset(&mut self, ptr: &'ll Value, fill_byte: &'ll Value, size: &'ll Value, align: Align) {
+        unsafe {
+            llvm::LLVMRustBuildMemSet(
+                self.llbuilder,
+                ptr,
+                align.bytes() as c_uint,
+                fill_byte,
+                size,
+                false,
+            );
+        }
+    }
 }
 
 /// Empty string, to be used where LLVM expects an instruction name, indicating