Refactor the way cg_ssa handles indirect returns (returns via sret) — Take 2

compiler/rustc_codegen_gcc/src/builder.rs

@@ -18,7 +18,7 @@ use rustc_codegen_ssa::mir::operand::{OperandRef, OperandValue};
 use rustc_codegen_ssa::mir::place::PlaceRef;
 use rustc_codegen_ssa::traits::{
     BackendTypes, BaseTypeCodegenMethods, BuilderMethods, ConstCodegenMethods,
-    LayoutTypeCodegenMethods, OverflowOp, StaticBuilderMethods,
+    LayoutTypeCodegenMethods, OverflowOp, ReturnSlot, StaticBuilderMethods,
 };
 use rustc_data_structures::fx::FxHashSet;
 use rustc_middle::bug;
@@ -603,6 +603,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
         fn_attrs: Option<&CodegenFnAttrs>,
         fn_abi: Option<&FnAbi<'tcx, Ty<'tcx>>>,
         func: RValue<'gcc>,
+        return_slot: ReturnSlot<RValue<'gcc>>,
         args: &[RValue<'gcc>],
         then: Block<'gcc>,
         catch: Block<'gcc>,
@@ -613,7 +614,8 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
 
         let current_block = self.block;
         self.block = try_block;
-        let call = self.call(typ, fn_attrs, fn_abi, func, args, None, instance); // FIXME(antoyo): use funclet here?
+        // FIXME(antoyo): use funclet here?
+        let call = self.call(typ, fn_attrs, fn_abi, func, return_slot, args, None, instance);
         self.block = current_block;
 
         let return_value =
@@ -641,13 +643,14 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
         fn_attrs: Option<&CodegenFnAttrs>,
         fn_abi: Option<&FnAbi<'tcx, Ty<'tcx>>>,
         func: RValue<'gcc>,
+        return_slot: ReturnSlot<RValue<'gcc>>,
         args: &[RValue<'gcc>],
         then: Block<'gcc>,
         catch: Block<'gcc>,
         _funclet: Option<&Funclet>,
         instance: Option<Instance<'tcx>>,
     ) -> RValue<'gcc> {
-        let call_site = self.call(typ, fn_attrs, fn_abi, func, args, None, instance);
+        let call_site = self.call(typ, fn_attrs, fn_abi, func, return_slot, args, None, instance);
         let condition = self.context.new_rvalue_from_int(self.bool_type, 1);
         self.llbb().end_with_conditional(self.location, condition, then, catch);
         if let Some(_fn_abi) = fn_abi {
@@ -1772,19 +1775,30 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
         _fn_attrs: Option<&CodegenFnAttrs>,
         fn_abi: Option<&FnAbi<'tcx, Ty<'tcx>>>,
         func: RValue<'gcc>,
+        return_slot: ReturnSlot<RValue<'gcc>>,
         args: &[RValue<'gcc>],
         funclet: Option<&Funclet>,
         _instance: Option<Instance<'tcx>>,
     ) -> RValue<'gcc> {
+        // FIXME: change this in the `rustc_codegen_gcc` repo after the sync, to use the `libgccjit` indirect return suppport.
+        let args = match return_slot {
+            ReturnSlot::Direct => args.to_vec(),
+            ReturnSlot::Indirect(sret_ptr) => {
+                let mut args = args.to_vec();
+                // Prepend the indirect return pointer
+                args.insert(0, sret_ptr);
+                args
+            }
+        };
         // FIXME(antoyo): remove when having a proper API.
         let gcc_func = unsafe { std::mem::transmute::<RValue<'gcc>, Function<'gcc>>(func) };
         let call = if self.functions.borrow().values().any(|value| *value == gcc_func) {
             // FIXME(antoyo): remove when the API supports a different type for functions.
             let func: Function<'gcc> = self.cx.rvalue_as_function(func);
-            self.function_call(func, args, funclet)
+            self.function_call(func, &args, funclet)
         } else {
             // If it's a not function that was defined, it's a function pointer.
-            self.function_ptr_call(typ, fn_abi, func, args, funclet)
+            self.function_ptr_call(typ, fn_abi, func, &args, funclet)
         };
         if let Some(_fn_abi) = fn_abi {
             // FIXME(bjorn3): Apply function attributes
@@ -1798,6 +1812,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
         _fn_attrs: Option<&CodegenFnAttrs>,
         _fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
         _llfn: Self::Value,
+        _return_slot: ReturnSlot<Self::Value>,
         _args: &[Self::Value],
         _funclet: Option<&Self::Funclet>,
         _instance: Option<Instance<'tcx>>,

compiler/rustc_codegen_gcc/src/intrinsic/mod.rs

@@ -16,7 +16,7 @@ use rustc_codegen_ssa::mir::place::{PlaceRef, PlaceValue};
 use rustc_codegen_ssa::traits::MiscCodegenMethods;
 use rustc_codegen_ssa::traits::{
     ArgAbiBuilderMethods, BaseTypeCodegenMethods, BuilderMethods, ConstCodegenMethods,
-    IntrinsicCallBuilderMethods, LayoutTypeCodegenMethods,
+    IntrinsicCallBuilderMethods, LayoutTypeCodegenMethods, ReturnSlot,
 };
 use rustc_codegen_ssa::{MemFlags, RetagInfo};
 use rustc_data_structures::fx::FxHashSet;
@@ -656,7 +656,8 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tc
         }
 
         // FIXME directly use the llvm intrinsic adjustment functions here
-        let llret = self.call(fn_ty, None, None, fn_ptr, &call_args, None, None);
+        let llret =
+            self.call(fn_ty, None, None, fn_ptr, ReturnSlot::Direct, &call_args, None, None);
         if is_cleanup {
             self.apply_attrs_to_cleanup_callsite(llret);
         }
@@ -667,7 +668,7 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tc
     fn abort(&mut self) {
         let func = self.context.get_builtin_function("abort");
         let func: RValue<'gcc> = unsafe { std::mem::transmute(func) };
-        self.call(self.type_void(), None, None, func, &[], None, None);
+        self.call(self.type_void(), None, None, func, ReturnSlot::Direct, &[], None, None);
     }
 
     fn assume(&mut self, value: Self::Value) {
@@ -1348,7 +1349,7 @@ fn try_intrinsic<'a, 'b, 'gcc, 'tcx>(
         let param_type = bx.u8_type.make_pointer();
         let fn_type =
             bx.context.new_function_pointer_type(None, bx.type_void(), &[param_type], false);
-        bx.call(fn_type, None, None, try_func, &[data], None, None);
+        bx.call(fn_type, None, None, try_func, ReturnSlot::Direct, &[data], None, None);
         // Return 0 unconditionally from the intrinsic call;
         // we can never unwind.
         OperandValue::Immediate(bx.const_bool(false)).store(bx, dest);
@@ -1421,21 +1422,41 @@ fn codegen_gnu_try<'gcc, 'tcx>(
         let zero = bx.cx.context.new_rvalue_zero(bx.int_type);
         let ptr = bx.cx.context.new_call(None, eh_pointer_builtin, &[zero]);
         let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
-        bx.call(catch_ty, None, None, catch_func, &[data, ptr], None, None);
+        bx.call(catch_ty, None, None, catch_func, ReturnSlot::Direct, &[data, ptr], None, None);
         bx.ret(bx.const_bool(true));
 
         // NOTE: the blocks must be filled before adding the try/catch, otherwise gcc will not
         // generate a try/catch.
         // FIXME(antoyo): add a check in the libgccjit API to prevent this.
         bx.switch_to_block(current_block);
-        bx.invoke(try_func_ty, None, None, try_func, &[data], then, catch, None, None);
+        bx.invoke(
+            try_func_ty,
+            None,
+            None,
+            try_func,
+            ReturnSlot::Direct,
+            &[data],
+            then,
+            catch,
+            None,
+            None,
+        );
     });
 
     let func = unsafe { std::mem::transmute::<Function<'gcc>, RValue<'gcc>>(func) };
 
     // Note that no invoke is used here because by definition this function
     // can't panic (that's what it's catching).
-    let ret = bx.call(llty, None, None, func, &[try_func, data, catch_func], None, None);
+    let ret = bx.call(
+        llty,
+        None,
+        None,
+        func,
+        ReturnSlot::Direct,
+        &[try_func, data, catch_func],
+        None,
+        None,
+    );
     OperandValue::Immediate(ret).store(bx, dest);
 }
 

compiler/rustc_codegen_llvm/src/asm.rs

@@ -562,9 +562,20 @@ pub(crate) fn inline_asm_call<'ll>(
         assert!(catch_funclet.is_none());
         bx.callbr(fty, None, None, v, inputs, dest.unwrap(), labels, None, None)
     } else if let Some((catch, funclet)) = catch_funclet {
-        bx.invoke(fty, None, None, v, inputs, dest.unwrap(), catch, funclet, None)
+        bx.invoke(
+            fty,
+            None,
+            None,
+            v,
+            ReturnSlot::Direct,
+            inputs,
+            dest.unwrap(),
+            catch,
+            funclet,
+            None,
+        )
     } else {
-        bx.call(fty, None, None, v, inputs, None, None)
+        bx.call(fty, None, None, v, ReturnSlot::Direct, inputs, None, None)
     };
 
     // Store mark in a metadata node so we can map LLVM errors

compiler/rustc_codegen_llvm/src/builder.rs

@@ -451,15 +451,25 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
         fn_attrs: Option<&CodegenFnAttrs>,
         fn_abi: Option<&FnAbi<'tcx, Ty<'tcx>>>,
         llfn: &'ll Value,
+        return_slot: ReturnSlot<&'ll Value>,
         args: &[&'ll Value],
         then: &'ll BasicBlock,
         catch: &'ll BasicBlock,
         funclet: Option<&Funclet<'ll>>,
         instance: Option<Instance<'tcx>>,
     ) -> &'ll Value {
+        // If this function returns indirectly (`PassMode::Indirect`),
+        // the `return_slot` should be the first argument.
+        let args = match return_slot {
+            ReturnSlot::Direct => args.to_vec(),
+            ReturnSlot::Indirect(sret_ptr) => {
+                let mut args = args.to_vec();
+                args.insert(0, sret_ptr);
+                args
+            }
+        };
         debug!("invoke {:?} with args ({:?})", llfn, args);
-
-        let args = self.check_call("invoke", llty, llfn, args);
+        let args = self.check_call("invoke", llty, llfn, &args);
         let funclet_bundle = funclet.map(|funclet| funclet.bundle());
         let mut bundles: SmallVec<[_; 2]> = SmallVec::new();
         if let Some(funclet_bundle) = funclet_bundle {
@@ -1452,13 +1462,23 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
         caller_attrs: Option<&CodegenFnAttrs>,
         fn_abi: Option<&FnAbi<'tcx, Ty<'tcx>>>,
         llfn: &'ll Value,
+        return_slot: ReturnSlot<&'ll Value>,
         args: &[&'ll Value],
         funclet: Option<&Funclet<'ll>>,
         callee_instance: Option<Instance<'tcx>>,
     ) -> &'ll Value {
+        // If this function returns indirectly (`PassMode::Indirect`),
+        // the `return_slot` should be the first argument.
+        let args = match return_slot {
+            ReturnSlot::Direct => args.to_vec(),
+            ReturnSlot::Indirect(sret_ptr) => {
+                let mut args = args.to_vec();
+                args.insert(0, sret_ptr);
+                args
+            }
+        };
         debug!("call {:?} with args ({:?})", llfn, args);
-
-        let args = self.check_call("call", llty, llfn, args);
+        let args = self.check_call("call", llty, llfn, &args);
         let funclet_bundle = funclet.map(|funclet| funclet.bundle());
         let mut bundles: SmallVec<[_; 2]> = SmallVec::new();
         if let Some(funclet_bundle) = funclet_bundle {
@@ -1519,12 +1539,21 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
         caller_attrs: Option<&CodegenFnAttrs>,
         fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
         llfn: Self::Value,
+        return_slot: ReturnSlot<Self::Value>,
         args: &[Self::Value],
         funclet: Option<&Self::Funclet>,
         callee_instance: Option<Instance<'tcx>>,
     ) {
-        let call =
-            self.call(llty, caller_attrs, Some(fn_abi), llfn, args, funclet, callee_instance);
+        let call = self.call(
+            llty,
+            caller_attrs,
+            Some(fn_abi),
+            llfn,
+            return_slot,
+            args,
+            funclet,
+            callee_instance,
+        );
         llvm::LLVMSetTailCallKind(call, llvm::TailCallKind::MustTail);
 
         match &fn_abi.ret.mode {
@@ -1819,7 +1848,8 @@ impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> {
         args: &[&'ll Value],
     ) -> &'ll Value {
         let (ty, f) = self.cx.get_intrinsic(base_name.into(), type_params);
-        self.call(ty, None, None, f, args, None, None)
+        // No LLVM intrinsic returns its data indirectly (via `sret`).
+        self.call(ty, None, None, f, ReturnSlot::Direct, args, None, None)
     }
 
     fn call_lifetime_intrinsic(&mut self, intrinsic: &'static str, ptr: &'ll Value, size: Size) {

compiler/rustc_codegen_llvm/src/builder/autodiff.rs

@@ -6,7 +6,7 @@ use rustc_codegen_ssa::common::TypeKind;
 use rustc_codegen_ssa::mir::IntrinsicResult;
 use rustc_codegen_ssa::mir::operand::{OperandRef, OperandValue};
 use rustc_codegen_ssa::mir::place::PlaceValue;
-use rustc_codegen_ssa::traits::{BaseTypeCodegenMethods, BuilderMethods};
+use rustc_codegen_ssa::traits::{BaseTypeCodegenMethods, BuilderMethods, ReturnSlot};
 use rustc_data_structures::thin_vec::ThinVec;
 use rustc_hir::attrs::RustcAutodiff;
 use rustc_middle::ty::{PseudoCanonicalInput, Ty, TyCtxt, TypingEnv};
@@ -377,7 +377,7 @@ pub(crate) fn generate_enzyme_call<'ll, 'tcx>(
         crate::typetree::add_tt(&bx, fn_to_diff, fnc_tree);
     }
 
-    let call = bx.call(enzyme_ty, None, None, ad_fn, &args, None, None);
+    let call = bx.call(enzyme_ty, None, None, ad_fn, ReturnSlot::Direct, &args, None, None);
 
     let fn_ret_ty = bx.cx.val_ty(call);
     if fn_ret_ty == bx.cx.type_void() || fn_ret_ty == bx.cx.type_struct(&[], false) {

compiler/rustc_codegen_llvm/src/builder/gpu_offload.rs

@@ -6,7 +6,7 @@ use rustc_abi::Align;
 use rustc_codegen_ssa::MemFlags;
 use rustc_codegen_ssa::common::TypeKind;
 use rustc_codegen_ssa::mir::operand::{OperandRef, OperandValue};
-use rustc_codegen_ssa::traits::{BaseTypeCodegenMethods, BuilderMethods};
+use rustc_codegen_ssa::traits::{BaseTypeCodegenMethods, BuilderMethods, ReturnSlot};
 use rustc_middle::bug;
 use rustc_middle::ty::offload_meta::{MappingFlags, OffloadMetadata, OffloadSize};
 
@@ -109,9 +109,18 @@ pub(crate) fn register_offload<'ll>(cx: &CodegenCx<'ll, '_>) {
     // }
     let bb = Builder::append_block(cx, desc_reg_fn, "entry");
     let mut a = Builder::build(cx, bb);
-    a.call(reg_lib_decl, None, None, register_lib, &[omp_descriptor], None, None);
-    a.call(init_ty, None, None, init_rtls, &[], None, None);
-    a.call(atexit, None, None, atexit_fn, &[desc_unreg_fn], None, None);
+    a.call(
+        reg_lib_decl,
+        None,
+        None,
+        register_lib,
+        ReturnSlot::Direct,
+        &[omp_descriptor],
+        None,
+        None,
+    );
+    a.call(init_ty, None, None, init_rtls, ReturnSlot::Direct, &[], None, None);
+    a.call(atexit, None, None, atexit_fn, ReturnSlot::Direct, &[desc_unreg_fn], None, None);
     a.ret_void();
 
     // define internal void @.omp_offloading.descriptor_unreg() section ".text.startup" {
@@ -121,7 +130,16 @@ pub(crate) fn register_offload<'ll>(cx: &CodegenCx<'ll, '_>) {
     // }
     let bb = Builder::append_block(cx, desc_unreg_fn, "entry");
     let mut a = Builder::build(cx, bb);
-    a.call(reg_lib_decl, None, None, unregister_lib, &[omp_descriptor], None, None);
+    a.call(
+        reg_lib_decl,
+        None,
+        None,
+        unregister_lib,
+        ReturnSlot::Direct,
+        &[omp_descriptor],
+        None,
+        None,
+    );
     a.ret_void();
 
     // @llvm.global_ctors = appending global [1 x { i32, ptr, ptr }] [{ i32, ptr, ptr } { i32 101, ptr @.omp_offloading.descriptor_reg, ptr null }]
@@ -740,7 +758,7 @@ pub(crate) fn gen_call_handling<'ll, 'tcx>(
         let num_args = cx.get_const_i32(num_args);
         let args =
             vec![s_ident_t, i64_max, num_args, geps[0], geps[1], geps[2], o_type, nullptr, nullptr];
-        builder.call(fn_ty, None, None, fn_to_call, &args, None, None);
+        builder.call(fn_ty, None, None, fn_to_call, ReturnSlot::Direct, &args, None, None);
     }
 
     // Step 2)
@@ -784,7 +802,7 @@ pub(crate) fn gen_call_handling<'ll, 'tcx>(
         region_id,
         a5,
     ];
-    builder.call(tgt_target_kernel_ty, None, None, tgt_decl, &args, None, None);
+    builder.call(tgt_target_kernel_ty, None, None, tgt_decl, ReturnSlot::Direct, &args, None, None);
     // %41 = call i32 @__tgt_target_kernel(ptr @1, i64 -1, i32 2097152, i32 256, ptr @.kernel_1.region_id, ptr %kernel_args)
 
     // Step 4)