[PATCH vkd3d 07/12] vkd3d-shader/hlsl: Replace register offsets with index paths in parse code.
Zebediah Figura
zfigura at codeweavers.com
Fri Jul 8 16:28:36 CDT 2022
On 7/1/22 16:24, Francisco Casas wrote:
> +/* Returns the path of a given component within a type, given its index.
> + * *comp_type will be set to the type of the component.
> + * *path_len will be set to the lenght of the path.
> + * Memory should be free afterwards.
> + */
> +unsigned int *hlsl_compute_component_path(struct hlsl_ctx *ctx, struct hlsl_type *type,
> + unsigned int idx, struct hlsl_type **comp_type, unsigned int *path_len)
This is awkward partly because zero is a valid length. Ideally it should
return bool. More than that, though, it's filling fields of struct
hlsl_deref, so what about something like
static bool hlsl_deref_from_component_index(struct hlsl_ctx *ctx,
struct hlsl_deref *deref, struct hlsl_ir_var *var,
unsigned int index)
This obviously doesn't lend itself nicely to recursion. I ended up
playing around with this for a bit and came up with the following
abbreviated code:
unsigned int subtype_index_from_component_index(struct hlsl_ctx *ctx,
inout const struct hlsl_type **type, inout unsigned int *index)
{
assert(index < hlsl_type_component_count(type));
...
}
static bool hlsl_deref_from_component_index(struct hlsl_ctx *ctx,
struct hlsl_deref *deref, struct hlsl_ir_var *var, uint index)
{
const struct hlsl_type *type = var->data_type;
unsigned int *indices;
size_t index_count;
while (/* type is not 1x1 numeric */)
{
indices[index_count++]
= subtype_index_from_component(ctx, &type, &index);
}
init_deref(ctx, deref, var, index_count);
// and so on
}
Thoughts?
> +/* Returns a simple variable derefence, so that the value can be stored and then passed by reference
> + * to load/store functions. It shall not be modified afterwards. */
> +struct hlsl_deref hlsl_get_direct_var_deref(struct hlsl_ir_var *var)
> +{
> + struct hlsl_deref deref = {};
> +
> + deref.var = var;
> + return deref;
> +}
> +
Naming: for consistency with hlsl_new_simple_store() I'd replace
"direct" with "simple". (Also naming: perhaps hlsl_simple_deref_from_var
instead, i.e. follow the "X from Y" pattern given what the function is
doing).
I'm also inclined to say this should take the hlsl_deref as a pointer in
the first argument, rather than returning a struct.
> struct hlsl_ir_resource_load *hlsl_new_resource_load(struct hlsl_ctx *ctx, struct hlsl_type *data_type,
> - enum hlsl_resource_load_type type, struct hlsl_ir_var *resource, struct hlsl_ir_node *resource_offset,
> - struct hlsl_ir_var *sampler, struct hlsl_ir_node *sampler_offset, struct hlsl_ir_node *coords,
> - struct hlsl_ir_node *texel_offset, const struct vkd3d_shader_location *loc)
> + enum hlsl_resource_load_type type, struct hlsl_deref *resource, struct hlsl_deref *sampler,
> + struct hlsl_ir_node *coords, struct hlsl_ir_node *texel_offset, const struct vkd3d_shader_location *loc)
Separate patch?
> @@ -870,10 +1136,8 @@ struct hlsl_ir_resource_load *hlsl_new_resource_load(struct hlsl_ctx *ctx, struc
> return NULL;
> init_node(&load->node, HLSL_IR_RESOURCE_LOAD, data_type, *loc);
> load->load_type = type;
> - load->resource.var = resource;
> - hlsl_src_from_node(&load->resource.offset, resource_offset);
> - load->sampler.var = sampler;
> - hlsl_src_from_node(&load->sampler.offset, sampler_offset);
> + deref_copy(ctx, &load->resource, resource);
> + deref_copy(ctx, &load->sampler, sampler);
Actually that means deref_copy() could be pulled out of this patch,
right? Just to save another hunk :-)
> @@ -334,7 +336,10 @@ static struct hlsl_ir_node *add_cast(struct hlsl_ctx *ctx, struct list *instrs,
> }
> }
>
> - dst_offset = hlsl_compute_component_offset(ctx, dst_type, dst_idx, &dst_scalar_type);
> + dst_path = hlsl_compute_component_path(ctx, dst_type, dst_idx, &dst_scalar_type, &dst_path_len);
> + if (dst_path_len && !dst_path)
> + return NULL;
> + vkd3d_free(dst_path);
As far as I can tell we're just doing this so that we can get the right
destination type to cast into. In that case my inclination is to move
that cast into hlsl_new_store_component() itself. Same in
initialize_var_components().
>
> if (!(load = add_load_component(ctx, instrs, node, src_idx, *loc)))
> return NULL;
> @@ -343,16 +348,12 @@ static struct hlsl_ir_node *add_cast(struct hlsl_ctx *ctx, struct list *instrs,
> return NULL;
> list_add_tail(instrs, &cast->node.entry);
>
> - if (!(c = hlsl_new_uint_constant(ctx, dst_offset, loc)))
> + if (!(store = hlsl_new_store_component(ctx, &block, &var_deref, dst_idx, &cast->node)))
> return NULL;
> - list_add_tail(instrs, &c->node.entry);
> -
> - if (!(store = hlsl_new_store(ctx, var, &c->node, &cast->node, 0, *loc)))
> - return NULL;
> - list_add_tail(instrs, &store->node.entry);
> + list_move_tail(instrs, &block.instrs);
> }
>
> - if (!(load = hlsl_new_load(ctx, var, NULL, dst_type, *loc)))
> + if (!(load = hlsl_new_var_load(ctx, var, *loc)))
> return NULL;
> list_add_tail(instrs, &load->node.entry);
>
...
> @@ -625,22 +626,13 @@ static struct hlsl_ir_jump *add_return(struct hlsl_ctx *ctx, struct list *instrs
> static struct hlsl_ir_load *add_load_index(struct hlsl_ctx *ctx, struct list *instrs, struct hlsl_ir_node *var_node,
> struct hlsl_ir_node *idx, const struct vkd3d_shader_location loc)
> {
> - struct hlsl_type *elem_type;
> - struct hlsl_ir_node *offset;
> struct hlsl_ir_load *load;
> - struct hlsl_block block;
> -
> - elem_type = hlsl_get_type_from_path_index(ctx, var_node->data_type, idx);
>
> if (var_node->type == HLSL_IR_LOAD)
> {
> const struct hlsl_deref *src = &hlsl_ir_load(var_node)->src;
>
> - if (!(offset = hlsl_new_offset_from_path_index(ctx, &block, var_node->data_type, src->offset.node, idx, &loc)))
> - return NULL;
> - list_move_tail(instrs, &block.instrs);
> -
> - if (!(load = hlsl_new_load(ctx, src->var, offset, elem_type, loc)))
> + if (!(load = hlsl_new_load_index(ctx, src, idx, loc)))
> return NULL;
> list_add_tail(instrs, &load->node.entry);
> }
Is it possible to introduce hlsl_new_load_index() and
hlsl_new_load_component() in helper patch(es), the same way we did for
add_load_index()? Same for stores I think.
(I know this is a lot of splitting, and/or busywork, but the more the
patch is split, the easier it is to be sure we've caught all the errors.)
More information about the wine-devel
mailing list