Minor improvement to multi return value reducing stack usage

6 files changed, 90 insertions, 19 deletions

diff --git a/src/check_expr.cpp b/src/check_expr.cpp
index c58aac609..c819267fd 100644
--- a/src/check_expr.cpp
+++ b/src/check_expr.cpp
@@ -6739,6 +6739,9 @@ ExprKind check_call_expr(CheckerContext *c, Operand *operand, Ast *call, Ast *pr
 	if (initial_entity != nullptr && initial_entity->kind == Entity_Procedure) {
 		if (initial_entity->Procedure.deferred_procedure.entity != nullptr) {
 			call->viral_state_flags |= ViralStateFlag_ContainsDeferredProcedure;
+			if (c->decl != nullptr) {
+				c->decl->defer_use_count += 1;
+			}
 		}
 	}
 
diff --git a/src/check_stmt.cpp b/src/check_stmt.cpp
index 9cacb4a35..8271ca451 100644
--- a/src/check_stmt.cpp
+++ b/src/check_stmt.cpp
@@ -2014,6 +2014,10 @@ void check_stmt_internal(CheckerContext *ctx, Ast *node, u32 flags) {
 		if (is_ast_decl(ds->stmt)) {
 			error(ds->token, "You cannot defer a declaration");
 		} else {
+			if (ctx->decl != nullptr) {
+				ctx->decl->defer_use_count += 1;
+			}
+
 			bool out_in_defer = ctx->in_defer;
 			ctx->in_defer = true;
 			check_stmt(ctx, ds->stmt, 0);
diff --git a/src/checker.hpp b/src/checker.hpp
index badcd93d5..ded4a0ad6 100644
--- a/src/checker.hpp
+++ b/src/checker.hpp
@@ -158,6 +158,7 @@ struct DeclInfo {
 	bool          is_using;
 	bool          where_clauses_evaluated;
 	bool          proc_checked;
+	isize         defer_use_count;
 
 	CommentGroup *comment;
 	CommentGroup *docs;
diff --git a/src/llvm_backend_general.cpp b/src/llvm_backend_general.cpp
index f36dc1842..6f2253d01 100644
--- a/src/llvm_backend_general.cpp
+++ b/src/llvm_backend_general.cpp
@@ -936,23 +936,27 @@ void lb_emit_store(lbProcedure *p, lbValue ptr, lbValue value) {
 			LLVMValueRef src_ptr_original = LLVMGetOperand(value.value, 0);
 			LLVMValueRef src_ptr = LLVMBuildPointerCast(p->builder, src_ptr_original, LLVMTypeOf(dst_ptr), "");
 
-			LLVMBuildMemMove(p->builder,
-			                 dst_ptr, lb_try_get_alignment(dst_ptr, 1),
-			                 src_ptr, lb_try_get_alignment(src_ptr_original, 1),
-			                 LLVMConstInt(LLVMInt64TypeInContext(p->module->ctx), lb_sizeof(LLVMTypeOf(value.value)), false));
+			if (dst_ptr != src_ptr && dst_ptr != src_ptr_original) {
+				LLVMBuildMemMove(p->builder,
+				                 dst_ptr, lb_try_get_alignment(dst_ptr, 1),
+				                 src_ptr, lb_try_get_alignment(src_ptr_original, 1),
+				                 LLVMConstInt(LLVMInt64TypeInContext(p->module->ctx), lb_sizeof(LLVMTypeOf(value.value)), false));
+			}
 			return;
 		} else if (LLVMIsConstant(value.value)) {
 			lbAddr addr = lb_add_global_generated(p->module, value.type, value, nullptr);
 			lb_make_global_private_const(addr);
 
 			LLVMValueRef dst_ptr = ptr.value;
-			LLVMValueRef src_ptr = addr.addr.value;
-			src_ptr = LLVMBuildPointerCast(p->builder, src_ptr, LLVMTypeOf(dst_ptr), "");
+			LLVMValueRef src_ptr_original = addr.addr.value;
+			LLVMValueRef src_ptr = LLVMBuildPointerCast(p->builder, src_ptr_original, LLVMTypeOf(dst_ptr), "");
 
-			LLVMBuildMemMove(p->builder,
-			                 dst_ptr, lb_try_get_alignment(dst_ptr, 1),
-			                 src_ptr, lb_try_get_alignment(src_ptr, 1),
-			                 LLVMConstInt(LLVMInt64TypeInContext(p->module->ctx), lb_sizeof(LLVMTypeOf(value.value)), false));
+			if (dst_ptr != src_ptr && dst_ptr != src_ptr_original) {
+				LLVMBuildMemMove(p->builder,
+				                 dst_ptr, lb_try_get_alignment(dst_ptr, 1),
+				                 src_ptr, lb_try_get_alignment(src_ptr, 1),
+				                 LLVMConstInt(LLVMInt64TypeInContext(p->module->ctx), lb_sizeof(LLVMTypeOf(value.value)), false));
+			}
 			return;
 		}
 	}
diff --git a/src/llvm_backend_opt.cpp b/src/llvm_backend_opt.cpp
index e2f51b868..ba5a41871 100644
--- a/src/llvm_backend_opt.cpp
+++ b/src/llvm_backend_opt.cpp
@@ -267,6 +267,8 @@ void lb_populate_module_pass_manager(LLVMTargetMachineRef target_machine, LLVMPa
 **************************************************************************/
 
 void lb_run_remove_dead_instruction_pass(lbProcedure *p) {
+	LLVMTypeRef llvm_void = LLVMVoidTypeInContext(p->module->ctx);
+
 	isize removal_count = 0;
 	isize pass_count = 0;
 	isize const max_pass_count = 10;
@@ -322,7 +324,7 @@ void lb_run_remove_dead_instruction_pass(lbProcedure *p) {
 				case LLVMOr:
 				case LLVMXor:
 				case LLVMAlloca:
-				case LLVMLoad:
+				case LLVMLoad: // TODO: should LLVMLoad be removed?
 				case LLVMGetElementPtr:
 				case LLVMTrunc:
 				case LLVMZExt:
@@ -347,6 +349,37 @@ void lb_run_remove_dead_instruction_pass(lbProcedure *p) {
 					LLVMInstructionEraseFromParent(curr_instr);
 					was_dead_instructions = true;
 					break;
+
+				case LLVMCall:
+					if (LLVMTypeOf(curr_instr) == llvm_void) {
+						LLVMValueRef the_proc = LLVMGetCalledValue(curr_instr);
+						unsigned id = LLVMGetIntrinsicID(the_proc);
+						if (id != 0) {
+							size_t text_len = 0;
+							char const *text = LLVMIntrinsicGetName(id, &text_len);
+							String name = make_string(cast(u8 const *)text, cast(isize)text_len);
+							if (name == "llvm.memmove" || name == "llvm.memcpy") {
+								LLVMValueRef dst = LLVMGetOperand(curr_instr, 0);
+								LLVMValueRef src = LLVMGetOperand(curr_instr, 1);
+								LLVMValueRef sz = LLVMGetOperand(curr_instr, 2);
+								if ((dst == src) || (LLVMIsConstant(sz) && LLVMConstIntGetZExtValue(sz) == 0)) {
+									removal_count += 1;
+									LLVMInstructionEraseFromParent(curr_instr);
+									was_dead_instructions = true;
+									break;
+								}
+							} else if (name == "llvm.memset") {
+								LLVMValueRef sz = LLVMGetOperand(curr_instr, 2);
+								if (LLVMIsConstant(sz) && LLVMConstIntGetZExtValue(sz) == 0) {
+									removal_count += 1;
+									LLVMInstructionEraseFromParent(curr_instr);
+									was_dead_instructions = true;
+									break;
+								}
+							}
+						}
+					}
+					break;
 				}
 			}
 		}
diff --git a/src/llvm_backend_proc.cpp b/src/llvm_backend_proc.cpp
index eaff6edc0..2e508a939 100644
--- a/src/llvm_backend_proc.cpp
+++ b/src/llvm_backend_proc.cpp
@@ -9,9 +9,15 @@ LLVMValueRef lb_call_intrinsic(lbProcedure *p, const char *name, LLVMValueRef* a
 }
 
 void lb_mem_copy_overlapping(lbProcedure *p, lbValue dst, lbValue src, lbValue len, bool is_volatile) {
+	if (dst.value == src.value) {
+		return;
+	}
 	dst = lb_emit_conv(p, dst, t_rawptr);
 	src = lb_emit_conv(p, src, t_rawptr);
 	len = lb_emit_conv(p, len, t_int);
+	if (dst.value == src.value) {
+		return;
+	}
 	
 	char const *name = "llvm.memmove";
 	if (LLVMIsConstant(len.value)) {
@@ -38,9 +44,16 @@ void lb_mem_copy_overlapping(lbProcedure *p, lbValue dst, lbValue src, lbValue l
 
 
 void lb_mem_copy_non_overlapping(lbProcedure *p, lbValue dst, lbValue src, lbValue len, bool is_volatile) {
+	if (dst.value == src.value) {
+		return;
+	}
 	dst = lb_emit_conv(p, dst, t_rawptr);
 	src = lb_emit_conv(p, src, t_rawptr);
 	len = lb_emit_conv(p, len, t_int);
+	if (dst.value == src.value) {
+		return;
+	}
+
 	
 	char const *name = "llvm.memcpy";
 	if (LLVMIsConstant(len.value)) {
@@ -580,18 +593,31 @@ 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
-					//         }
-					lbAddr res = lb_add_local(p, e->type, e);
+					lbAddr res = {};
+					if (p->return_ptr.addr.value != nullptr &&
+					    p->entity                != nullptr &&
+					    p->entity->decl_info     != nullptr &&
+					    p->entity->decl_info->defer_use_count == 0) {
+						lbValue val = lb_emit_struct_ep(p, p->return_ptr.addr, cast(i32)i);
+
+						lb_add_entity(p->module, e, val);
+						lb_add_debug_local_variable(p, val.value, e->type, e->token);
+
+						// NOTE(bill): no need to zero initialize due to caller will zero return value
+						res = lb_addr(val);
+					} else {
+						// 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
+						//         }
+						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);
 					}
-
 				}
 			}