2 files changed, 20 insertions, 5 deletions

diff --git a/executor/x86_64/asm.c b/executor/x86_64/asm.c
index e29130e..c633db8 100644
--- a/executor/x86_64/asm.c
+++ b/executor/x86_64/asm.c
@@ -210,16 +210,21 @@ static void asm_print_code_hex(Asm *self) {
 }
 #endif
 
+typedef union {
+    u8 *data;
+    int (*func)(void);
+} RawFuncCallPtr;
+
 int asm_execute(Asm *self, u32 offset) {
-    void (*func)();
+    RawFuncCallPtr raw_func_ptr;
     if(mprotect(self->code, self->allocated_size, PROT_READ | PROT_EXEC) != 0)
         return -errno;
 
     /*asm_print_code_hex(self);*/
 
     /* TODO: Verify if this is valid on all platforms. According to ISO C standard it isn't? */
-    *(void**)(&func) = (u8*)self->code + offset;
-    func();
+    raw_func_ptr.data = (u8*)self->code + offset;
+    raw_func_ptr.func();
     return 0;
 }
 
diff --git a/executor/x86_64/executor.c b/executor/x86_64/executor.c
index c442da8..f747e4a 100644
--- a/executor/x86_64/executor.c
+++ b/executor/x86_64/executor.c
@@ -284,6 +284,7 @@ int amal_exec_call(amal_executor *self, u32 code_offset, u8 num_args, i8 dst_reg
     isize asm_offset = asm_get_size(&impl->asm);
     ASM_ENSURE_CAPACITY
 
+    assert(num_args % 2 == 0 && "TODO: Align stack to 16-bytes before calling functions");
     assert(code_offset < asm_offset);
     asm_call_rel32(&impl->asm, (isize)code_offset - asm_offset);
 
@@ -318,7 +319,8 @@ int amal_exec_calle(amal_executor *self, void *func, u8 num_args, i8 dst_reg) {
     AsmPtr dst;
     IMPL_START
 
-    /* TODO: Support R and XMM registers so more than 5 arguments can be used for functions */
+    assert(num_args % 2 == 0 && "TODO: Align stack to 16-bytes before calling functions");
+    /* TODO: Support R and XMM registers so more than 4 arguments can be used for functions */
     assert(num_args < 5);
     {
         /*
@@ -427,6 +429,10 @@ int amal_exec_ret(amal_executor *self, i8 reg) {
     return amal_exec_func_end(self);
 }
 
+static u32 get_next_uneven_number(u32 value) {
+    return value + !(value & 1);
+}
+
 int amal_exec_func_start(amal_executor *self, u16 num_regs) {
     /*
         TODO: Validate stack size, or maybe remove all validation? do we really need validation?
@@ -443,7 +449,11 @@ int amal_exec_func_start(amal_executor *self, u16 num_regs) {
     asm_pushr(&impl->asm, RBX);
     asm_pushr(&impl->asm, RBP);
     asm_mov_rr(&impl->asm, RBP, RSP);
-    asm_sub_rm64_imm(&impl->asm, RSP, num_regs * sizeof(isize));
+    /*
+        Functions are entered with a stack alignment of 8 (because of call return address is pushed to stack).
+        Make sure to align to to next 16-byte even if the extra bytes are not used.
+    */
+    asm_sub_rm64_imm(&impl->asm, RSP, get_next_uneven_number(num_regs) * sizeof(isize));
     return 0;
 }