[Mesa-dev] [PATCH] nv50/ir: handle SHLADD in IndirectPropagation

[Karol Herbst]

On Mon, Jun 11, 2018 at 2:39 PM, Rhys Perry <pendingchaos02 at gmail.com> wrote:
> An alternative solution to the problem fixed in
> 0bd83d0 ("nv50/ir: move LateAlgebraicOpt to the very end"). Should be
> useful in the future and seems to make dolphin ubershaders a bit smaller.
>
> total instructions in shared programs : 226722 -> 226464 (-0.11%)
> total gprs used in shared programs    : 19378 -> 19378 (0.00%)
> total shared used in shared programs  : 0 -> 0 (0.00%)
> total local used in shared programs   : 0 -> 0 (0.00%)
>
>                 local     shared        gpr       inst      bytes
>     helped           0           0           0          51          51
>       hurt           0           0           0           0           0
>
> Signed-off-by: Rhys Perry <pendingchaos02 at gmail.com>
> ---
>  src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp | 14 +++++++++++++-
>  1 file changed, 13 insertions(+), 1 deletion(-)
>
> diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
> index 39177bd044..4d0589214d 100644
> --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
> +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
> @@ -283,6 +283,8 @@ class IndirectPropagation : public Pass
>  {
>  private:
>     virtual bool visit(BasicBlock *);
> +
> +   BuildUtil bld;
>  };
>
>  bool
> @@ -294,6 +296,8 @@ IndirectPropagation::visit(BasicBlock *bb)
>     for (Instruction *i = bb->getEntry(); i; i = next) {
>        next = i->next;
>
> +      bld.setPosition(i, false);
> +
>        for (int s = 0; i->srcExists(s); ++s) {
>           Instruction *insn;
>           ImmediateValue imm;
> @@ -325,6 +329,14 @@ IndirectPropagation::visit(BasicBlock *bb)
>              i->setIndirect(s, 0, NULL);
>              i->setSrc(s, cloneShallow(func, i->getSrc(s)));
>              i->src(s).get()->reg.data.offset += imm.reg.data.u32;
> +         } else if (insn->op == OP_SHLADD) {
> +            if (!insn->src(2).getImmediate(imm) ||
> +                !targ->insnCanLoadOffset(i, s, imm.reg.data.s32))
> +               continue;
> +            i->setIndirect(s, 0, bld.mkOp2v(
> +               OP_SHL, TYPE_U32, bld.getSSA(), insn->getSrc(0), insn->getSrc(1)));
> +            i->setSrc(s, cloneShallow(func, i->getSrc(s)));
> +            i->src(s).get()->reg.data.offset += imm.reg.data.u32;
>           }
>        }
>     }
> @@ -3797,11 +3809,11 @@ Program::optimizeSSA(int level)
>     RUN_PASS(2, AlgebraicOpt, run);
>     RUN_PASS(2, ModifierFolding, run); // before load propagation -> less checks
>     RUN_PASS(1, ConstantFolding, foldAll);
> +   RUN_PASS(2, LateAlgebraicOpt, run);
>     RUN_PASS(1, Split64BitOpPreRA, run);
>     RUN_PASS(1, LoadPropagation, run);
>     RUN_PASS(1, IndirectPropagation, run);
>     RUN_PASS(2, MemoryOpt, run);
> -   RUN_PASS(2, LateAlgebraicOpt, run);
>     RUN_PASS(2, LocalCSE, run);
>     RUN_PASS(0, DeadCodeElim, buryAll);

well no, it is late for a good reason actually. those opts have the
potential to increase instruction count in other shaders. This should
be tested on more shaders than just the default shader-db ones.

>
> --
> 2.14.4
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev