Merge pull request #2208 from odin-lang/multiple-return-abi-experiment

Multiple Return ABI Changes and Improvements

1 files changed, 165 insertions, 42 deletions

diff --git a/src/llvm_backend_proc.cpp b/src/llvm_backend_proc.cpp
index 6e2345436..f8dffa9b6 100644
--- a/src/llvm_backend_proc.cpp
+++ b/src/llvm_backend_proc.cpp
@@ -1,4 +1,3 @@
-
 LLVMValueRef lb_call_intrinsic(lbProcedure *p, const char *name, LLVMValueRef* args, unsigned arg_count, LLVMTypeRef* types, unsigned type_count)
 {
 	unsigned id = LLVMLookupIntrinsicID(name, gb_strlen(name));
@@ -123,6 +122,7 @@ lbProcedure *lb_create_procedure(lbModule *m, Entity *entity, bool ignore_body)
 	p->scope_stack.allocator   = a;
 	map_init(&p->selector_values,  a, 0);
 	map_init(&p->selector_addr,    a, 0);
+	map_init(&p->tuple_fix_map,    a, 0);
 
 	if (p->is_foreign) {
 		lb_add_foreign_library_path(p->module, entity->Procedure.foreign_library);
@@ -501,8 +501,23 @@ void lb_begin_procedure_body(lbProcedure *p) {
 			// NOTE(bill): this must be parameter 0
 
 			String name = str_lit("agg.result");
+			if (ft->multiple_return_original_type &&
+			    p->type->Proc.has_named_results) {
+				auto const &variables = p->type->Proc.results->Tuple.variables;
+				Entity *e = variables[variables.count-1];
+				if (!is_blank_ident(e->token)) {
+					name = e->token.string;
+				}
+			}
 
-			Type *ptr_type = alloc_type_pointer(reduce_tuple_to_single_type(p->type->Proc.results));
+			Type *return_ptr_type = reduce_tuple_to_single_type(p->type->Proc.results);
+			bool split_returns = ft->multiple_return_original_type != nullptr;
+			if (split_returns) {
+				GB_ASSERT(is_type_tuple(return_ptr_type));
+				auto const &variables = return_ptr_type->Tuple.variables;
+				return_ptr_type = variables[variables.count-1]->type;
+			}
+			Type *ptr_type = alloc_type_pointer(return_ptr_type);
 			Entity *e = alloc_entity_param(nullptr, make_token_ident(name), ptr_type, false, false);
 			e->flags |= EntityFlag_NoAlias;
 
@@ -580,14 +595,70 @@ void lb_begin_procedure_body(lbProcedure *p) {
 				if (e->token.string != "") {
 					GB_ASSERT(!is_blank_ident(e->token));
 
-					// NOTE(bill): Don't even bother trying to optimize this with the return ptr value
-					// This will violate the defer rules if you do:
-					//         foo :: proc() -> (x, y: T) {
-					//                 defer x = ... // defer is executed after the `defer`
-					//                 return // the values returned should be zeroed
-					//         }
-					// NOTE(bill): REALLY, don't even bother.
-					lbAddr res = lb_add_local(p, e->type, e);
+					lbAddr res = {};
+					if (p->entity && p->entity->decl_info &&
+					    p->entity->decl_info->defer_use_checked &&
+					    p->entity->decl_info->defer_used == 0) {
+
+						// NOTE(bill): this is a bodge to get around the issue of the problem BELOW
+						// We check to see if we ever use a defer statement ever within a procedure and if it
+						// if it never happens, see if you can possibly do take the return value pointer
+						//
+						// NOTE(bill): this could be buggy in that I have missed a case where `defer` was used
+						//
+						// TODO(bill): This could be optimized to check to see where a `defer` only uses
+						// the variable in question
+
+						bool has_return_ptr = p->return_ptr.addr.value != nullptr;
+						lbValue ptr = {};
+
+						if (ft->multiple_return_original_type != nullptr) {
+							isize the_offset = -1;
+							if (i+1 < results->variables.count) {
+								the_offset = cast(isize)param_offset + ft->original_arg_count + i;
+							} else if (has_return_ptr) {
+								GB_ASSERT(i+1 == results->variables.count);
+								the_offset = 0;
+							}
+							if (the_offset >= 0) {
+								lbValue ptr = {};
+								ptr.value = LLVMGetParam(p->value, cast(unsigned)the_offset);
+								ptr.type = alloc_type_pointer(e->type);
+
+
+							}
+						} else if (has_return_ptr) {
+							lbValue ptr = p->return_ptr.addr;
+
+							if (results->variables.count > 1) {
+								ptr = lb_emit_tuple_ep(p, ptr, cast(i32)i);
+							}
+							GB_ASSERT(is_type_pointer(ptr.type));
+							GB_ASSERT(are_types_identical(type_deref(ptr.type), e->type));
+						}
+
+						if (ptr.value != nullptr) {
+							lb_add_entity(p->module, e, ptr);
+							lb_add_debug_local_variable(p, ptr.value, e->type, e->token);
+							// NOTE(bill): no need to zero on the callee side as it is zeroed on the caller side
+
+							res = lb_addr(ptr);
+						}
+					}
+
+					if (res.addr.type == nullptr) {
+						// NOTE(bill): Don't even bother trying to optimize this with the return ptr value
+						// This will violate the defer rules if you do:
+						//         foo :: proc() -> (x, y: T) {
+						//                 defer x = ... // defer is executed after the `defer`
+						//                 return // the values returned should be zeroed
+						//         }
+						// NOTE(bill): REALLY, don't even bother.
+						//
+						// IMPORTANT NOTE(bill): REALLY, don't even bother!!!!!!
+						res = lb_add_local(p, e->type, e);
+					}
+
 					if (e->Variable.param_value.kind != ParameterValue_Invalid) {
 						lbValue c = lb_handle_param_value(p, e->type, e->Variable.param_value, e->token.pos);
 						lb_addr_store(p, res, c);
@@ -700,15 +771,8 @@ Array<lbValue> lb_value_to_array(lbProcedure *p, lbValue value) {
 	if (t == nullptr) {
 		// Do nothing
 	} else if (is_type_tuple(t)) {
-		GB_ASSERT(t->kind == Type_Tuple);
-		auto *rt = &t->Tuple;
-		if (rt->variables.count > 0) {
-			array = array_make<lbValue>(permanent_allocator(), rt->variables.count);
-			for_array(i, rt->variables) {
-				lbValue elem = lb_emit_struct_ev(p, value, cast(i32)i);
-				array[i] = elem;
-			}
-		}
+		array = array_make<lbValue>(permanent_allocator(), 0, t->Tuple.variables.count);
+		lb_append_tuple_values(p, &array, value);
 	} else {
 		array = array_make<lbValue>(permanent_allocator(), 1);
 		array[0] = value;
@@ -734,6 +798,7 @@ lbValue lb_emit_call_internal(lbProcedure *p, lbValue value, lbValue return_ptr,
 	if (return_ptr.value != nullptr) {
 		args[arg_index++] = return_ptr.value;
 	}
+
 	for_array(i, processed_args) {
 		lbValue arg = processed_args[i];
 		if (is_type_proc(arg.type)) {
@@ -741,16 +806,23 @@ lbValue lb_emit_call_internal(lbProcedure *p, lbValue value, lbValue return_ptr,
 		}
 		args[arg_index++] = arg.value;
 	}
+
 	if (context_ptr.addr.value != nullptr) {
 		LLVMValueRef cp = context_ptr.addr.value;
 		cp = LLVMBuildPointerCast(p->builder, cp, lb_type(p->module, t_rawptr), "");
 		args[arg_index++] = cp;
 	}
+
+	GB_ASSERT(arg_index == arg_count);
+
 	LLVMBasicBlockRef curr_block = LLVMGetInsertBlock(p->builder);
 	GB_ASSERT(curr_block != p->decl_block->block);
 
 	{
-		LLVMTypeRef fnp = lb_type_internal_for_procedures_raw(p->module, value.type);
+		Type *proc_type = base_type(value.type);
+		GB_ASSERT(proc_type->kind == Type_Proc);
+
+		LLVMTypeRef fnp = lb_type_internal_for_procedures_raw(p->module, proc_type);
 		LLVMTypeRef ftp = LLVMPointerType(fnp, 0);
 		LLVMValueRef fn = value.value;
 		if (!lb_is_type_kind(LLVMTypeOf(value.value), LLVMFunctionTypeKind)) {
@@ -775,10 +847,11 @@ lbValue lb_emit_call_internal(lbProcedure *p, lbValue value, lbValue return_ptr,
 				// LLVMTypeKind arg_kind = LLVMGetTypeKind(arg_type);
 				GB_ASSERT_MSG(
 					arg_type == param_type,
-					"Parameter types do not match: %s != %s, argument: %s",
+					"Parameter types do not match: %s != %s, argument: %s\n\t%s",
 					LLVMPrintTypeToString(arg_type),
 					LLVMPrintTypeToString(param_type),
-					LLVMPrintValueToString(args[i])
+					LLVMPrintValueToString(args[i]),
+					LLVMPrintTypeToString(fnp)
 				);
 			}
 		}
@@ -882,7 +955,7 @@ lbValue lb_emit_conjugate(lbProcedure *p, lbValue val, Type *type) {
 	return lb_emit_load(p, res);
 }
 
-lbValue lb_emit_call(lbProcedure *p, lbValue value, Array<lbValue> const &args, ProcInlining inlining, bool use_copy_elision_hint) {
+lbValue lb_emit_call(lbProcedure *p, lbValue value, Array<lbValue> const &args, ProcInlining inlining) {
 	lbModule *m = p->module;
 
 	Type *pt = base_type(value.type);
@@ -915,8 +988,9 @@ lbValue lb_emit_call(lbProcedure *p, lbValue value, Array<lbValue> const &args,
 
 		bool is_odin_cc = is_calling_convention_odin(pt->Proc.calling_convention);
 
-		lbFunctionType *ft = lb_get_function_type(m, p, pt);
+		lbFunctionType *ft = lb_get_function_type(m, pt);
 		bool return_by_pointer = ft->ret.kind == lbArg_Indirect;
+		bool split_returns = ft->multiple_return_original_type != nullptr;
 
 		unsigned param_index = 0;
 		for (isize i = 0; i < param_count; i++) {
@@ -979,18 +1053,19 @@ lbValue lb_emit_call(lbProcedure *p, lbValue value, Array<lbValue> const &args,
 		}
 
 		Type *rt = reduce_tuple_to_single_type(results);
-		if (return_by_pointer) {
-			lbValue return_ptr = {};
-			if (use_copy_elision_hint && p->copy_elision_hint.ptr.value != nullptr) {
-				if (are_types_identical(type_deref(p->copy_elision_hint.ptr.type), rt)) {
-					return_ptr = lb_consume_copy_elision_hint(p);
-				}
-			}
-			if (return_ptr.value == nullptr) {
-				lbAddr r = lb_add_local_generated(p, rt, true);
-				return_ptr = r.addr;
+		Type *original_rt = rt;
+		if (split_returns) {
+			GB_ASSERT(rt->kind == Type_Tuple);
+			for (isize j = 0; j < rt->Tuple.variables.count-1; j++) {
+				Type *partial_return_type = rt->Tuple.variables[j]->type;
+				lbValue partial_return_ptr = lb_add_local(p, partial_return_type, nullptr, true, false).addr;
+				array_add(&processed_args, partial_return_ptr);
 			}
-			GB_ASSERT(is_type_pointer(return_ptr.type));
+			rt = reduce_tuple_to_single_type(rt->Tuple.variables[rt->Tuple.variables.count-1]->type);
+		}
+
+		if (return_by_pointer) {
+			lbValue return_ptr = lb_add_local_generated(p, rt, true).addr;
 			lb_emit_call_internal(p, value, return_ptr, processed_args, nullptr, context_ptr, inlining);
 			result = lb_emit_load(p, return_ptr);
 		} else if (rt != nullptr) {
@@ -1010,6 +1085,47 @@ lbValue lb_emit_call(lbProcedure *p, lbValue value, Array<lbValue> const &args,
 			lb_emit_call_internal(p, value, {}, processed_args, nullptr, context_ptr, inlining);
 		}
 
+		if (original_rt != rt) {
+			GB_ASSERT(split_returns);
+			GB_ASSERT(is_type_tuple(original_rt));
+
+			// IMPORTANT NOTE(bill, 2022-11-24)
+			// result_ptr is a dummy value which is only used to reference a tuple
+			// value for the "tuple-fix"
+			//
+			// The reason for the fake stack allocation is to have a unique pointer
+			// for the value to be used as a key within the procedure itself
+
+			lbValue result_ptr = lb_add_local_generated(p, original_rt, false).addr;
+			isize ret_count = original_rt->Tuple.variables.count;
+
+			auto tuple_fix_values = slice_make<lbValue>(permanent_allocator(), ret_count);
+			auto tuple_geps = slice_make<lbValue>(permanent_allocator(), ret_count);
+
+			isize offset = ft->original_arg_count;
+			for (isize j = 0; j < ret_count-1; j++) {
+				lbValue ret_arg_ptr = processed_args[offset + j];
+				lbValue ret_arg = lb_emit_load(p, ret_arg_ptr);
+				tuple_fix_values[j] = ret_arg;
+			}
+			tuple_fix_values[ret_count-1] = result;
+
+		#if 0
+			for (isize j = 0; j < ret_count; j++) {
+				tuple_geps[j] = lb_emit_struct_ep(p, result_ptr, cast(i32)j);
+			}
+			for (isize j = 0; j < ret_count; j++) {
+				lb_emit_store(p, tuple_geps[j], tuple_fix_values[j]);
+			}
+		#endif
+
+			result = lb_emit_load(p, result_ptr);
+
+			lbTupleFix tf = {tuple_fix_values};
+			map_set(&p->tuple_fix_map, result_ptr.value, tf);
+			map_set(&p->tuple_fix_map, result.value, tf);
+		}
+
 	}
 
 	Entity **found = map_get(&p->module->procedure_values, value.value);
@@ -2300,7 +2416,7 @@ lbValue lb_build_builtin_proc(lbProcedure *p, Ast *expr, TypeAndValue const &tv,
 		);
 		LLVMSetWeak(value, weak);
 
-		if (tv.type->kind == Type_Tuple) {
+		if (is_type_tuple(tv.type)) {
 			Type *fix_typed = alloc_type_tuple();
 			slice_init(&fix_typed->Tuple.variables, permanent_allocator(), 2);
 			fix_typed->Tuple.variables[0] = tv.type->Tuple.variables[0];
@@ -3032,7 +3148,7 @@ lbValue lb_build_call_expr_internal(lbProcedure *p, Ast *expr) {
 			}
 		}
 
-		return lb_emit_call(p, value, args, ce->inlining, p->copy_elision_hint.ast == expr);
+		return lb_emit_call(p, value, args, ce->inlining);
 	}
 
 	isize arg_index = 0;
@@ -3044,7 +3160,7 @@ lbValue lb_build_call_expr_internal(lbProcedure *p, Ast *expr) {
 		GB_ASSERT_MSG(tav.mode != Addressing_Invalid, "%s %s %d", expr_to_string(arg), expr_to_string(expr), tav.mode);
 		GB_ASSERT_MSG(tav.mode != Addressing_ProcGroup, "%s", expr_to_string(arg));
 		Type *at = tav.type;
-		if (at->kind == Type_Tuple) {
+		if (is_type_tuple(at)) {
 			arg_count += at->Tuple.variables.count;
 		} else {
 			arg_count++;
@@ -3084,9 +3200,16 @@ lbValue lb_build_call_expr_internal(lbProcedure *p, Ast *expr) {
 			lbValue a = lb_build_expr(p, arg);
 			Type *at = a.type;
 			if (at->kind == Type_Tuple) {
-				for_array(i, at->Tuple.variables) {
-					lbValue v = lb_emit_struct_ev(p, a, cast(i32)i);
-					args[arg_index++] = v;
+				lbTupleFix *tf = map_get(&p->tuple_fix_map, a.value);
+				if (tf) {
+					for_array(j, tf->values) {
+						args[arg_index++] = tf->values[j];
+					}
+				} else {
+					for_array(j, at->Tuple.variables) {
+						lbValue v = lb_emit_struct_ev(p, a, cast(i32)j);
+						args[arg_index++] = v;
+					}
 				}
 			} else {
 				args[arg_index++] = a;
@@ -3213,6 +3336,6 @@ lbValue lb_build_call_expr_internal(lbProcedure *p, Ast *expr) {
 	}
 
 	auto call_args = array_slice(args, 0, final_count);
-	return lb_emit_call(p, value, call_args, ce->inlining, p->copy_elision_hint.ast == expr);
+	return lb_emit_call(p, value, call_args, ce->inlining);
 }