Add @(align=N) for variables, raising the alignment of globals, locals, @(static) and @(thread_local) variables to at least N

src/check_decl.cpp

@@ -1753,6 +1753,7 @@ gb_internal void check_global_variable_decl(CheckerContext *ctx, Entity *e, Ast
 
 	e->Variable.thread_local_model = ac.thread_local_model;
 	e->Variable.is_export = ac.is_export;
+	e->Variable.custom_align = ac.align;
 	e->flags &= ~EntityFlag_Static;
 	if (ac.is_static) {
 		error(e->token, "@(static) is not supported for global variables, nor required");

src/check_stmt.cpp

@@ -2287,6 +2287,7 @@ gb_internal void check_value_decl_stmt(CheckerContext *ctx, Ast *node, u32 mod_f
 		if (ac.link_name.len > 0) {
 			e->Variable.link_name = ac.link_name;
 		}
+		e->Variable.custom_align = ac.align;
 
 		e->flags &= ~EntityFlag_Static;
 		if (ac.is_static) {

src/checker.cpp

@@ -4653,6 +4653,24 @@ gb_internal DECL_ATTRIBUTE_PROC(var_decl_attribute) {
 			error(elem, "Expected either no value or a string for '%.*s'", LIT(name));
 		}
 		return true;
+	} else if (name == "align") {
+		ExactValue ev = check_decl_attribute_value(c, value);
+		if (ev.kind != ExactValue_Integer) {
+			error(elem, "Expected a constant integer for '%.*s'", LIT(name));
+			return true;
+		}
+		i64 align = exact_value_to_i64(ev);
+		if (ev.value_integer.used > 1 ||
+		    align < 1 ||
+		    align > (1ll<<31) ||
+		    !gb_is_power_of_two(cast(isize)align)) {
+			gbString str = exact_value_to_string(ev);
+			error(elem, "'%.*s' must be a power of 2 no greater than 2^31, got %s", LIT(name), str);
+			gb_string_free(str);
+		} else {
+			ac->align = align;
+		}
+		return true;
 	}
 
 	if (c->curr_proc_decl != nullptr) {

src/checker.hpp

@@ -123,6 +123,7 @@ struct AttributeContext {
 	String  link_section;
 	String  linkage;
 	isize   init_expr_list_count;
+	i64     align;
 	String  thread_local_model;
 	String  deprecated_message;
 	String  warning_message;

src/entity.cpp

@@ -273,6 +273,7 @@ struct Entity {
 			String     link_prefix;
 			String     link_suffix;
 			String     link_section;
+			i64        custom_align;
 			CommentGroup *docs;
 			CommentGroup *comment;
 			bool       is_foreign;

src/llvm_backend.cpp

@@ -2421,7 +2421,7 @@ gb_internal void lb_create_global_variable(lbModule *m, lbGlobalVariable *var) {
 		LLVM_SET_INTERNAL_WEAK_LINKAGE(g.value);
 	}
 	lb_set_linkage_from_entity_flags(m, g.value, e->flags);
-	LLVMSetAlignment(g.value, cast(u32)type_align_of(e->type));
+	LLVMSetAlignment(g.value, cast(u32)gb_max(type_align_of(e->type), e->Variable.custom_align));
 
 	if (e->Variable.link_section.len > 0) {
 		LLVMSetSection(g.value, alloc_cstring(permanent_allocator(), e->Variable.link_section));

src/llvm_backend_general.cpp

@@ -4676,6 +4676,9 @@ gb_internal lbAddr lb_add_local(lbProcedure *p, Type *type, Entity *e, bool zero
 	if (is_type_matrix(type)) {
 		alignment *= 2; // NOTE(bill): Just in case
 	}
+	if (e != nullptr && e->kind == Entity_Variable) {
+		alignment = gb_max(alignment, cast(unsigned)e->Variable.custom_align);
+	}
 
 	LLVMValueRef ptr = llvm_alloca(p, llvm_type, alignment, name);
 

src/llvm_backend_stmt.cpp

@@ -109,7 +109,10 @@ gb_internal void lb_scan_for_sret_rvo(lbProcedure *p) {
 				Entity *e = entity_of_node(vd->names[0]);
 				if (e == ret_entity) {
 					Ast *rhs = unparen_expr(vd->values[0]);
-					if (rhs->kind == Ast_CallExpr && lb_call_sret_eligible(p, rhs, e->type)) {
+					// NOTE(bill): the caller's return slot only has the type's alignment and not @(align=N)'s
+					if (rhs->kind == Ast_CallExpr &&
+					    e->Variable.custom_align == 0 &&
+					    lb_call_sret_eligible(p, rhs, e->type)) {
 						decl_index = i;
 					}
 					goto done_scanning;
@@ -2481,7 +2484,7 @@ gb_internal void lb_build_static_variables(lbProcedure *p, AstValueDecl *vd) {
 		char *c_name = alloc_cstring(permanent_allocator(), mangled_name);
 
 		LLVMValueRef global = LLVMAddGlobal(p->module->mod, lb_type(p->module, e->type), c_name);
-		LLVMSetAlignment(global, cast(u32)type_align_of(e->type));
+		LLVMSetAlignment(global, cast(u32)gb_max(type_align_of(e->type), e->Variable.custom_align));
 		LLVMSetInitializer(global, LLVMConstNull(lb_type(p->module, e->type)));
 
 		if (e->Variable.is_rodata) {
@@ -2528,7 +2531,7 @@ gb_internal void lb_build_static_variables(lbProcedure *p, AstValueDecl *vd) {
 				if (actual_type != expected_type) {
 					LLVMDeleteGlobal(global);
 					global = LLVMAddGlobal(p->module->mod, actual_type, c_name);
-					LLVMSetAlignment(global, cast(u32)type_align_of(e->type));
+					LLVMSetAlignment(global, cast(u32)gb_max(type_align_of(e->type), e->Variable.custom_align));
 					if (e->Variable.is_rodata) {
 						LLVMSetGlobalConstant(global, true);
 					}
@@ -3464,7 +3467,9 @@ gb_internal void lb_build_stmt(lbProcedure *p, Ast *node) {
 					// the literal is one of its variants, so reusing that storage would bind the
 					// variable to a bare `[]int` and never build the union at all
 					if (comp_lit_addr && are_types_identical(lb_addr_type(*comp_lit_addr), type_of_expr(vd->names[lval_index]))) {
-						if (Entity *e = entity_of_node(vd->names[lval_index])) {
+						Entity *e = entity_of_node(vd->names[lval_index]);
+						// NOTE(bill): the literal's storage only has the type's alignment and not @(align=N)'s
+						if (e != nullptr && e->Variable.custom_align == 0) {
 							lbValue val = comp_lit_addr->addr;
 							lb_add_entity(p->module, e, val);
 							lb_add_debug_local_variable(p, val.value, e->type, e->token);