[Mesa-dev] [PATCH] nv50/ir: optimise slct(t, f, set) to mov(set) or not(set)

[Karol Herbst]

just noticed I sent out the wrong version of that patch...

On Tue, Mar 27, 2018 at 10:50 PM, Karol Herbst <kherbst at redhat.com> wrote:
> From: Karol Herbst <karolherbst at gmail.com>
>
> helps mainly Feral-ported games
>
> changes in shader-db:
> total instructions in shared programs : 3940749 -> 3935015 (-0.15%)
> total gprs used in shared programs    : 481460 -> 481433 (-0.01%)
> total local used in shared programs   : 27481 -> 27513 (0.12%)
> total bytes used in shared programs   : 36115776 -> 36063344 (-0.15%)
>
>                 local        gpr       inst      bytes
>     helped           6          31         854         854
>       hurt          10           5           1           1
> ---
>  .../drivers/nouveau/codegen/nv50_ir_peephole.cpp   | 26 ++++++++++++++++++++--
>  1 file changed, 24 insertions(+), 2 deletions(-)
>
> diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
> index 48cf74950df..1e3dea95494 100644
> --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
> +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
> @@ -637,9 +637,31 @@ ConstantFolding::expr(Instruction *i,
>        }
>        break;
>     case OP_SLCT:
> -      if (a->data.u32 != b->data.u32)
> +      // slct(a, a, b) -> a
> +      if (a->data.u32 == b->data.u32) {
> +         res.data.u32 = a->data.u32;
> +      } else {
> +         // slct_ne(true, false, bool) -> !bool
> +         CmpInstruction *slct = i->asCmp();
> +         Instruction *set = i->getSrc(2)->getInsn();
> +         if (!set || set->op != OP_SET)
> +            return;
> +         if (isFloatType(set->dType))
> +            return;
> +         if ((slct->getCondition() == CC_NE && imm0.isInteger(-1) && imm1.isInteger(0)) ||
> +             (slct->getCondition() == CC_EQ && imm0.isInteger(0) && imm1.isInteger(-1))) {
> +            bld.setPosition(i, false);
> +            bld.mkOp1(OP_MOV, i->dType, i->getDef(0), i->getSrc(2));
> +            delete_Instruction(prog, i);
> +         } else if (
> +             (slct->getCondition() == CC_EQ && imm0.isInteger(-1) && imm1.isInteger(0)) ||
> +             (slct->getCondition() == CC_NE && imm0.isInteger(0) && imm1.isInteger(-1))) {
> +            bld.setPosition(i, false);
> +            bld.mkOp1(OP_NOT, i->dType, i->getDef(0), i->getSrc(2));
> +            delete_Instruction(prog, i);
> +         }
>           return;
> -      res.data.u32 = a->data.u32;
> +      }
>        break;
>     case OP_EXTBF: {
>        int offset = b->data.u32 & 0xff;
> --
> 2.14.3
>