diff --git a/src/tilde/tb.h b/src/tilde/tb.h
index 3924bbe8e..c2dc3922b 100644
--- a/src/tilde/tb.h
+++ b/src/tilde/tb.h
@@ -196,6 +196,11 @@ typedef union TB_DataType {
 typedef enum TB_NodeTypeEnum {
     TB_NULL = 0,
 
+    // Immediates
+    TB_INTEGER_CONST,
+    TB_FLOAT32_CONST,
+    TB_FLOAT64_CONST,
+
     // only one per function
     TB_START, // fn()
 
@@ -253,11 +258,6 @@ typedef enum TB_NodeTypeEnum {
     TB_MEMBER_ACCESS,
     TB_ARRAY_ACCESS,
 
-    // Immediates
-    TB_INTEGER_CONST,
-    TB_FLOAT32_CONST,
-    TB_FLOAT64_CONST,
-
     // Conversions
     TB_TRUNCATE,
     TB_FLOAT_EXT,
@@ -295,6 +295,8 @@ typedef enum TB_NodeTypeEnum {
     TB_SHL,
     TB_SHR,
     TB_SAR,
+    TB_ROL,
+    TB_ROR,
     TB_UDIV,
     TB_SDIV,
     TB_UMOD,
@@ -336,9 +338,6 @@ typedef enum TB_NodeTypeEnum {
 } TB_NodeTypeEnum;
 typedef uint8_t TB_NodeType;
 
-#define TB_IS_NODE_SIDE_EFFECT(type) ((type) >= TB_LINE_INFO && (type) <= TB_DEBUGBREAK)
-#define TB_IS_NODE_TERMINATOR(type)  ((type) >= TB_BRANCH && (type) <= TB_TRAP)
-
 typedef int TB_Label;
 
 // just represents some region of bytes, usually in file parsing crap
@@ -980,6 +979,8 @@ TB_API TB_Node* tb_inst_xor(TB_Function* f, TB_Node* a, TB_Node* b);
 TB_API TB_Node* tb_inst_sar(TB_Function* f, TB_Node* a, TB_Node* b);
 TB_API TB_Node* tb_inst_shl(TB_Function* f, TB_Node* a, TB_Node* b, TB_ArithmeticBehavior arith_behavior);
 TB_API TB_Node* tb_inst_shr(TB_Function* f, TB_Node* a, TB_Node* b);
+TB_API TB_Node* tb_inst_rol(TB_Function* f, TB_Node* a, TB_Node* b);
+TB_API TB_Node* tb_inst_ror(TB_Function* f, TB_Node* a, TB_Node* b);
 
 // Atomics
 // By default you can use TB_MEM_ORDER_SEQ_CST for the memory order to get
@@ -1052,10 +1053,7 @@ TB_API void tb_inst_ret(TB_Function* f, size_t count, TB_Node** values);
 ////////////////////////////////
 // Optimizer
 ////////////////////////////////
-
-// Function-level optimizations are managed via TB_FuncOpt, it's tied
-// to a single TB_Function and it'll can be used to run peepholes incrementally
-// between whatever passes TB may have.
+// Function analysis, optimizations, and codegen are all part of this
 typedef struct TB_FuncOpt TB_FuncOpt;
 
 // the arena is used to allocate the nodes
diff --git a/src/tilde/tb.lib b/src/tilde/tb.lib
index 81c925854..5a1cd67dd 100644
Binary files a/src/tilde/tb.lib and b/src/tilde/tb.lib differ
diff --git a/src/tilde_stmt.cpp b/src/tilde_stmt.cpp
index c5aaa9161..6aa906288 100644
--- a/src/tilde_stmt.cpp
+++ b/src/tilde_stmt.cpp
@@ -850,11 +850,39 @@ gb_internal void cg_build_assignment(cgProcedure *p, Array<cgAddr> const &lvals,
 	p->in_multi_assignment = lval_count > 1;
 
 	GB_ASSERT(lvals.count == inits.count);
-	for_array(i, inits) {
+
+
+	if (inits.count > 1) for_array(i, inits) {
 		cgAddr lval = lvals[i];
-		cgValue init = inits[i];
+		cgValue init = cg_flatten_value(p, inits[i]);
+
+		GB_ASSERT(init.kind != cgValue_Multi);
+		if (init.type == nullptr) {
+			continue;
+		}
+
+	    	Type *type = cg_addr_type(lval);
+		GB_ASSERT(are_types_identical(type, init.type));
+		if (init.kind == cgValue_Addr &&
+		    !cg_addr_is_empty(lval)) {
+		    	// NOTE(bill): This is needed for certain constructs such as this:
+		    	// a, b = b, a
+		    	// NOTE(bill): This is a bodge and not necessarily a good way of doing things whatsoever
+		    	TB_CharUnits size  = cast(TB_CharUnits)type_size_of(type);
+		    	TB_CharUnits align = cast(TB_CharUnits)type_align_of(type);
+		    	TB_Node *copy = tb_inst_local(p->func, size, align);
+		    	tb_inst_memcpy(p->func, copy, init.node, tb_inst_uint(p->func, TB_TYPE_INT, size), align, false);
+		    	// use the copy instead
+		    	init.node = copy;
+		}
+		inits[i] = init;
+	}
+
+	for_array(i, inits) {
+		cgAddr  lval = lvals[i];
+		cgValue init = inits[i];
+		GB_ASSERT(init.kind != cgValue_Multi);
 		if (init.type == nullptr) {
-			// TODO(bill): figure out how to do this
 			continue;
 		}
 		cg_addr_store(p, lval, init);