On Fri Jul 7 02:54:02 2023 +0000, Francisco Casas wrote:
I'm going to have to push back on 2/4. Along the lines of 215, I don't
want to spend too much effort doing this on the HLSL level when we really should be doing these kinds of transformations over lower level IR. I am not following what on 2/4 requires too much effort. Can you specify? In sm4 relative addressing can only be done with "whole-register granularity", meaning that the node has to be expressed in whole registers and to specify register components we require a constant uint offset. If the node is constant, moving its value to this constant uint is a low-hanging fruit, I think.
I don't like how 4/4 avoids the recursive struct; probably better to
do it like vkd3d_shader_register. The thing is that the `struct sm4_instruction` values are "ephemeral", they are only created to be used as an input to write_sm4_instruction() and currently they don't allocate heap memory. I made the extended register `struct sm4_xregister` to avoid allocating heap memory and having to take it into account and free it everywhere it is used, and its `.r` field can be initialized as a regular register. I guess you are thinking on something more orthodox, along the lines of:
struct sm4_register_index { struct sm4_register *rel_addr; unsigned int offset; }; struct sm4_register { enum vkd3d_sm4_register_type type; struct sm4_register_index idx[2]; unsigned int idx_count; /* ... */ }; static struct sm4_register *sm4_new_register(void); static void sm4_register_add_index(struct sm4_register *reg, struct sm4_register *rel_addr, unsigned int offset); static void sm4_register_cleanup(struct sm4_register *reg);so I am implementing that.
For the record, I implemented this in this branch https://gitlab.winehq.org/fcasas/vkd3d/-/commits/nonconst-offsets-6. As you can see, moving the registers to the heap is costly (https://gitlab.winehq.org/fcasas/vkd3d/-/commit/b304a29e7668493dec3183428510...), but after that it allows for a nice implementation of the writing instructions part (https://gitlab.winehq.org/fcasas/vkd3d/-/commit/83ff8f6b866e19639b5e946055e1...). So it may be worth it.
One thing about this branch is that it requires !269 first, and henri is suggesting making sm4_register a subset of struct vkd3d_shader_register, so I will get to work on that.