[Mesa-dev] [PATCH v2] gk104/ir: fix tex use generation to be more careful about eliding uses

[Samuel Pitoiset]

Reviewed-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>

On 06/18/2016 10:05 PM, Ilia Mirkin wrote:
> If we have a loop, instructions before the tex might be added as tex
> uses, and those may in fact dominate all other uses of the tex results.
> This however doesn't mean that we don't need a texbar after the tex.
> Only check if uses dominate each other they are dominated by the tex.
>
> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=96565
> Fixes: 7752bbc44 (gk104/ir: simplify and fool-proof texbar algorithm)
> Signed-off-by: Ilia Mirkin <imirkin at alum.mit.edu>
> Cc: "11.2 12.0" <mesa-stable at lists.freedesktop.org>
> ---
>  .../nouveau/codegen/nv50_ir_lowering_nvc0.cpp      | 34 +++++++++++++++-------
>  .../nouveau/codegen/nv50_ir_lowering_nvc0.h        |  5 ++--
>  2 files changed, 27 insertions(+), 12 deletions(-)
>
> diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
> index 71013eb..67bd73b 100644
> --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
> +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
> @@ -172,19 +172,33 @@ NVC0LegalizePostRA::addTexUse(std::list<TexUse> &uses,
>                                Instruction *usei, const Instruction *texi)
>  {
>     bool add = true;
> -   for (std::list<TexUse>::iterator it = uses.begin();
> -        it != uses.end();) {
> -      if (insnDominatedBy(usei, it->insn)) {
> -         add = false;
> -         break;
> -      }
> -      if (insnDominatedBy(it->insn, usei))
> -         it = uses.erase(it);
> -      else
> +   bool dominated = insnDominatedBy(usei, texi);
> +   // Uses before the tex have to all be included. Just because an earlier
> +   // instruction dominates another instruction doesn't mean that there's no
> +   // way to get from the tex to the later instruction. For example you could
> +   // have nested loops, with the tex in the inner loop, and uses before it in
> +   // both loops - even though the outer loop's instruction would dominate the
> +   // inner's, we still want a texbar before the inner loop's instruction.
> +   //
> +   // However we can still use the eliding logic between uses dominated by the
> +   // tex instruction, as that is unambiguously correct.
> +   if (dominated) {
> +      for (std::list<TexUse>::iterator it = uses.begin(); it != uses.end();) {
> +         if (it->after) {
> +            if (insnDominatedBy(usei, it->insn)) {
> +               add = false;
> +               break;
> +            }
> +            if (insnDominatedBy(it->insn, usei)) {
> +               it = uses.erase(it);
> +               continue;
> +            }
> +         }
>           ++it;
> +      }
>     }
>     if (add)
> -      uses.push_back(TexUse(usei, texi));
> +      uses.push_back(TexUse(usei, texi, dominated));
>  }
>
>  // While it might be tempting to use the an algorithm that just looks at tex
> diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.h b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.h
> index 2321956..4b23303 100644
> --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.h
> +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.h
> @@ -55,10 +55,11 @@ private:
>
>     struct TexUse
>     {
> -      TexUse(Instruction *use, const Instruction *tex)
> -         : insn(use), tex(tex), level(-1) { }
> +      TexUse(Instruction *use, const Instruction *tex, bool after)
> +         : insn(use), tex(tex), after(after), level(-1) { }
>        Instruction *insn;
>        const Instruction *tex; // or split / mov
> +      bool after;
>        int level;
>     };
>     struct Limits
>