[Mesa-dev] [PATCH] nv50/ir: make a copy of tex src if it's referenced multiple times

[Ilia Mirkin]

On Mon, Apr 9, 2018 at 10:23 PM, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
> For nv50 we coalesce the srcs and defs into a single node (some ops
> can't take a separate src and dst, while others can only handle one reg
> in the short encoding mode).
>
> [Side-note: the ops for which there is an encoding that can take a src
> should instead use the RA's register preference facility.]

For posterity, turns out that I was mistaken. All the tex ops
apparently can only take a single register, even when it's a wide
encoding (LTDST vs LTSRC in envydis, but both reference the same
bitfield). I'm going to reword the description accordingly.

>
> Signed-off-by: Ilia Mirkin <imirkin at alum.mit.edu>
> ---
>  src/gallium/drivers/nouveau/codegen/nv50_ir_ra.cpp | 10 ++++++++++
>  1 file changed, 10 insertions(+)
>
> diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_ra.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_ra.cpp
> index 3a0e56e1385..f5ce1b4a0b0 100644
> --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_ra.cpp
> +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_ra.cpp
> @@ -2216,6 +2216,16 @@ RegAlloc::InsertConstraintsPass::texConstraintNV50(TexInstruction *tex)
>     for (c = 0; tex->srcExists(c) || tex->defExists(c); ++c) {
>        if (!tex->srcExists(c))
>           tex->setSrc(c, new_LValue(func, tex->getSrc(0)->asLValue()));
> +      else if (tex->getSrc(c)->refCount() > 1) {
> +         // Disconnect source from the single definition since it's about to
> +         // get merged with the defs (due to JOIN_MASK_TEX).
> +         LValue *lval = new_LValue(func, FILE_GPR);
> +         Instruction *mov = new_Instruction(func, OP_MOV, TYPE_U32);
> +         mov->setSrc(0, tex->getSrc(c));
> +         mov->setDef(0, lval);
> +         tex->bb->insertBefore(tex, mov);
> +         tex->setSrc(c, lval);
> +      }
>        if (!tex->defExists(c))
>           tex->setDef(c, new_LValue(func, tex->getDef(0)->asLValue()));
>     }
> --
> 2.16.1
>