diff --git a/py/asmx64.c b/py/asmx64.c
index f82a2ea423..f2ad16d15e 100644
--- a/py/asmx64.c
+++ b/py/asmx64.c
@@ -49,6 +49,7 @@
 #define OPCODE_MOV_I8_TO_R8      (0xb0) /* +rb */
 #define OPCODE_MOV_I64_TO_R64    (0xb8) /* +rq */
 #define OPCODE_MOV_I32_TO_RM32   (0xc7)
+#define OPCODE_MOV_R8_TO_RM8     (0x88) /* /r */
 #define OPCODE_MOV_R64_TO_RM64   (0x89) /* /r */
 #define OPCODE_MOV_RM64_TO_R64   (0x8b)
 #define OPCODE_LEA_MEM_TO_R64    (0x8d) /* /r */
@@ -85,6 +86,8 @@
 #define MODRM_RM_REG    (0xc0)
 #define MODRM_RM_R64(x) ((x) & 0x7)
 
+#define OP_SIZE_PREFIX (0x66)
+
 #define REX_PREFIX  (0x40)
 #define REX_W       (0x08)  // width
 #define REX_R       (0x04)  // register
@@ -298,18 +301,28 @@ STATIC void asm_x64_ret(asm_x64_t *as) {
     asm_x64_write_byte_1(as, OPCODE_RET);
 }
 
-void asm_x64_mov_r32_to_r32(asm_x64_t *as, int src_r32, int dest_r32) {
-    // defaults to 32 bit operation
-    assert(src_r32 < 8);
-    assert(dest_r32 < 8);
-    asm_x64_write_byte_2(as, OPCODE_MOV_R64_TO_RM64, MODRM_R64(src_r32) | MODRM_RM_REG | MODRM_RM_R64(dest_r32));
-}
-
 void asm_x64_mov_r64_to_r64(asm_x64_t *as, int src_r64, int dest_r64) {
     // use REX prefix for 64 bit operation
     asm_x64_write_byte_3(as, REX_PREFIX | REX_W | (src_r64 < 8 ? 0 : REX_R) | (dest_r64 < 8 ? 0 : REX_B), OPCODE_MOV_R64_TO_RM64, MODRM_R64(src_r64) | MODRM_RM_REG | MODRM_RM_R64(dest_r64));
 }
 
+void asm_x64_mov_r8_to_disp(asm_x64_t *as, int src_r64, int dest_r64, int dest_disp) {
+    assert(dest_r64 < 8);
+    if (src_r64 < 8) {
+        asm_x64_write_byte_1(as, OPCODE_MOV_R8_TO_RM8);
+    } else {
+        asm_x64_write_byte_2(as, REX_PREFIX | REX_R, OPCODE_MOV_R8_TO_RM8);
+    }
+    asm_x64_write_r64_disp(as, src_r64, dest_r64, dest_disp);
+}
+
+void asm_x64_mov_r16_to_disp(asm_x64_t *as, int src_r64, int dest_r64, int dest_disp) {
+    assert(src_r64 < 8);
+    assert(dest_r64 < 8);
+    asm_x64_write_byte_2(as, OP_SIZE_PREFIX, OPCODE_MOV_R64_TO_RM64);
+    asm_x64_write_r64_disp(as, src_r64, dest_r64, dest_disp);
+}
+
 void asm_x64_mov_r64_to_disp(asm_x64_t *as, int src_r64, int dest_r64, int dest_disp) {
     // use REX prefix for 64 bit operation
     assert(dest_r64 < 8);
@@ -356,6 +369,7 @@ void asm_x64_mov_i64_to_r64(asm_x64_t *as, int64_t src_i64, int dest_r64) {
 }
 
 void asm_x64_mov_i64_to_r64_optimised(asm_x64_t *as, int64_t src_i64, int dest_r64) {
+    // TODO use movzx, movsx if possible
     if (UNSIGNED_FIT32(src_i64)) {
         // 5 bytes
         asm_x64_mov_i32_to_r64(as, src_i64 & 0xffffffff, dest_r64);
diff --git a/py/asmx64.h b/py/asmx64.h
index 3a9aa832ec..440f573677 100644
--- a/py/asmx64.h
+++ b/py/asmx64.h
@@ -58,6 +58,7 @@
 #define ASM_X64_CC_JNZ (0x5)
 #define ASM_X64_CC_JNE (0x5)
 #define ASM_X64_CC_JL  (0xc) // less, signed
+#define ASM_X64_CC_JG  (0xf) // greater, signed
 
 typedef struct _asm_x64_t asm_x64_t;
 
@@ -75,6 +76,9 @@ void asm_x64_mov_r64_to_r64(asm_x64_t* as, int src_r64, int dest_r64);
 void asm_x64_mov_i64_to_r64(asm_x64_t* as, int64_t src_i64, int dest_r64);
 void asm_x64_mov_i64_to_r64_optimised(asm_x64_t *as, int64_t src_i64, int dest_r64);
 void asm_x64_mov_i64_to_r64_aligned(asm_x64_t *as, int64_t src_i64, int dest_r64);
+void asm_x64_mov_r8_to_disp(asm_x64_t *as, int src_r64, int dest_r64, int dest_disp);
+void asm_x64_mov_r16_to_disp(asm_x64_t *as, int src_r64, int dest_r64, int dest_disp);
+void asm_x64_mov_r64_to_disp(asm_x64_t *as, int src_r64, int dest_r64, int dest_disp);
 void asm_x64_xor_r64_to_r64(asm_x64_t *as, int src_r64, int dest_r64);
 void asm_x64_add_r64_to_r64(asm_x64_t* as, int src_r64, int dest_r64);
 void asm_x64_cmp_r64_with_r64(asm_x64_t* as, int src_r64_a, int src_r64_b);
diff --git a/py/asmx86.c b/py/asmx86.c
index 1570d9f7c0..ffe3e5e439 100644
--- a/py/asmx86.c
+++ b/py/asmx86.c
@@ -49,6 +49,7 @@
 //#define OPCODE_MOV_I8_TO_R8      (0xb0) /* +rb */
 #define OPCODE_MOV_I32_TO_R32    (0xb8)
 //#define OPCODE_MOV_I32_TO_RM32   (0xc7)
+#define OPCODE_MOV_R8_TO_RM8     (0x88) /* /r */
 #define OPCODE_MOV_R32_TO_RM32   (0x89)
 #define OPCODE_MOV_RM32_TO_R32   (0x8b)
 #define OPCODE_LEA_MEM_TO_R32    (0x8d) /* /r */
@@ -85,6 +86,8 @@
 #define MODRM_RM_REG    (0xc0)
 #define MODRM_RM_R32(x) (x)
 
+#define OP_SIZE_PREFIX (0x66)
+
 #define IMM32_L0(x) ((x) & 0xff)
 #define IMM32_L1(x) (((x) >> 8) & 0xff)
 #define IMM32_L2(x) (((x) >> 16) & 0xff)
@@ -232,7 +235,17 @@ void asm_x86_mov_r32_to_r32(asm_x86_t *as, int src_r32, int dest_r32) {
     asm_x86_write_byte_2(as, OPCODE_MOV_R32_TO_RM32, MODRM_R32(src_r32) | MODRM_RM_REG | MODRM_RM_R32(dest_r32));
 }
 
-STATIC void asm_x86_mov_r32_to_disp(asm_x86_t *as, int src_r32, int dest_r32, int dest_disp) {
+void asm_x86_mov_r8_to_disp(asm_x86_t *as, int src_r32, int dest_r32, int dest_disp) {
+    asm_x86_write_byte_1(as, OPCODE_MOV_R8_TO_RM8);
+    asm_x86_write_r32_disp(as, src_r32, dest_r32, dest_disp);
+}
+
+void asm_x86_mov_r16_to_disp(asm_x86_t *as, int src_r32, int dest_r32, int dest_disp) {
+    asm_x86_write_byte_2(as, OP_SIZE_PREFIX, OPCODE_MOV_R32_TO_RM32);
+    asm_x86_write_r32_disp(as, src_r32, dest_r32, dest_disp);
+}
+
+void asm_x86_mov_r32_to_disp(asm_x86_t *as, int src_r32, int dest_r32, int dest_disp) {
     asm_x86_write_byte_1(as, OPCODE_MOV_R32_TO_RM32);
     asm_x86_write_r32_disp(as, src_r32, dest_r32, dest_disp);
 }
diff --git a/py/asmx86.h b/py/asmx86.h
index 6099325c7e..76df7bf2b8 100644
--- a/py/asmx86.h
+++ b/py/asmx86.h
@@ -59,6 +59,7 @@
 #define ASM_X86_CC_JNZ (0x5)
 #define ASM_X86_CC_JNE (0x5)
 #define ASM_X86_CC_JL  (0xc) // less, signed
+#define ASM_X86_CC_JG  (0xf) // greater, signed
 
 typedef struct _asm_x86_t asm_x86_t;
 
@@ -72,6 +73,9 @@ void* asm_x86_get_code(asm_x86_t* as);
 void asm_x86_mov_r32_to_r32(asm_x86_t* as, int src_r32, int dest_r32);
 void asm_x86_mov_i32_to_r32(asm_x86_t *as, int32_t src_i32, int dest_r32);
 void asm_x86_mov_i32_to_r32_aligned(asm_x86_t *as, int32_t src_i32, int dest_r32);
+void asm_x86_mov_r8_to_disp(asm_x86_t *as, int src_r32, int dest_r32, int dest_disp);
+void asm_x86_mov_r16_to_disp(asm_x86_t *as, int src_r32, int dest_r32, int dest_disp);
+void asm_x86_mov_r32_to_disp(asm_x86_t *as, int src_r32, int dest_r32, int dest_disp);
 void asm_x86_xor_r32_to_r32(asm_x86_t *as, int src_r32, int dest_r32);
 void asm_x86_add_r32_to_r32(asm_x86_t* as, int src_r32, int dest_r32);
 void asm_x86_cmp_r32_with_r32(asm_x86_t* as, int src_r32_a, int src_r32_b);