[PATCH vkd3d v7 4/6] vkd3d-shader/hlsl: Perform a copy propagation pass.

[Matteo Bruni]

On Thu, Nov 18, 2021 at 7:23 PM Zebediah Figura <zfigura at codeweavers.com> wrote:
>
> On 11/18/21 12:18 PM, Matteo Bruni wrote:
> > On Thu, Nov 18, 2021 at 6:43 PM Zebediah Figura (she/her)
> > <zfigura at codeweavers.com> wrote:
> >>
> >> On 11/18/21 11:28, Matteo Bruni wrote:
> >>> This and copy_propagation_transform_block() are a bit of a
> >>> reimplementation of transform_ir(). It should be possible to extend
> >>> transform_ir() to handle this new pass instead.
> >>>
> >>> I'd introduce a struct like
> >>>
> >>> struct transform_pass
> >>> {
> >>>       bool (*initialize)(struct hlsl_ctx *hlsl_ctx, void **ctx);
> >>>       bool (*process_instruction)(void *ctx, struct hlsl_ir_node *, void *);
> >>>       void (*destroy)(void *ctx);
> >>> };
> >>>
> >>> and update transform_ir() and the existing passes to make use of it. I
> >>> imagine we could have some generic initialize() implementation simply
> >>> doing *ctx = hlsl_ctx; and an empty destroy() implementation to be
> >>> used when there is no particular context to carry around (like in the
> >>> existing passes).
> >>>
> >>
> >> Pretty much the whole point of transform_ir() is that it recursively
> >> iterates through CF blocks for cases where you don't care about
> >> transforming the CF blocks themselves. That's not the case here. What
> >> you're describing sounds to me like a framework that doesn't actually
> >> serve any purpose to its use case.
> >
> > Well, transform_ir() could start calling process_instruction() for the
> > HLSL_IR_IF and HLSL_IR_LOOP instructions before recursing, allowing
> > the pass to do what it needs for those, and existing passes would keep
> > ignoring those instructions. Would that not be enough? Or are you
> > saying that transform_ir() should not become more complex?
> >
>
> It could, but we also need to do things after the else/loop block, and
> between the if and else blocks. At that point, as far as I see it, the
> copy propagation function has to manually handle enough logic that
> transform_ir() doesn't actually save us any work.

Fair enough.