Currently, we can't use add_load_index() or add_load_component() on the loads to the `"retval"` variables that come from `add_call()` because these functions assume that the loaded value (`var_instr`) won't change between its location and the location and the new load to be created.
We can't get rid of this assumptions either, because, at least `add_load_index()` may be used in the lhs of an assignment, and in that case we can't store to the "deref" synthetic: ``` x[0] = 20; ```
Here I implemented the alternative solution of copying `"retval"` into a synthetic variable after each function call, but we may want to discuss this approach.
The second problem this that is covered in 3/3 is fixing the number of components of the swizzle instructions generated by copy-prop, without it, the following assertion in `hlsl_sm4.c` fails for me
```c /* Narrowing casts were already lowered. */ assert(src_type->dimx == dst_type->dimx); ```
for some tests, in one of my branches, because of narrowing casts may end up appearing if the node is replaced with a node with more components. This error happens under quite complex conditions (involving casts and matrices) that I still try to narrow down, but I think that the fix is logical enough, without it, the asserts introduced in the patch fail in many tests.