[Mesa-dev] [PATCH] nv50/ir: optimize ADD(SHL(a, b), c) to SHLADD(a, b, c)

[Ilia Mirkin]

On Sat, Oct 8, 2016 at 3:55 PM, Samuel Pitoiset
<samuel.pitoiset at gmail.com> wrote:
> total instructions in shared programs :2286901 -> 2284473 (-0.11%)
> total gprs used in shared programs    :335256 -> 335273 (0.01%)
> total local used in shared programs   :31968 -> 31968 (0.00%)
>
>                 local        gpr       inst      bytes
>     helped           0          41         852         852
>       hurt           0          44          23          23
>
> Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
> ---
>  .../drivers/nouveau/codegen/nv50_ir_peephole.cpp   | 94 ++++++++++++++++++++++
>  1 file changed, 94 insertions(+)
>
> diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
> index 6efb29e..caf5d1d 100644
> --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
> +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
> @@ -2132,6 +2132,99 @@ AlgebraicOpt::visit(BasicBlock *bb)
>
>  // =============================================================================
>
> +// ADD(SHL(a, b), c) -> SHLADD(a, b, c)
> +class LateAlgebraicOpt : public Pass
> +{
> +private:
> +   virtual bool visit(BasicBlock *);
> +
> +   void handleADD(Instruction *);
> +   bool tryADDToSHLADD(Instruction *);
> +
> +   BuildUtil bld;
> +};
> +
> +void
> +LateAlgebraicOpt::handleADD(Instruction *add)
> +{
> +   Value *src0 = add->getSrc(0);
> +   Value *src1 = add->getSrc(1);
> +
> +   if (src0->reg.file != FILE_GPR || src1->reg.file != FILE_GPR)
> +      return;
> +
> +   if (prog->getTarget()->isOpSupported(OP_SHLADD, add->dType))
> +      tryADDToSHLADD(add);
> +}
> +
> +// ADD(SHL(a, b), c) -> SHLADD(a, b, c)
> +bool
> +LateAlgebraicOpt::tryADDToSHLADD(Instruction *add)
> +{
> +   Value *src0 = add->getSrc(0);
> +   Value *src1 = add->getSrc(1);
> +   Modifier mod[2];
> +   Value *src;
> +   int s;
> +
> +   if (add->saturate || add->usesFlags() || typeSizeof(add->dType) == 8)
> +      return false;
> +
> +   if (src0->getUniqueInsn() && src0->getUniqueInsn()->op == OP_SHL)
> +      s = 0;
> +   else
> +   if (src1->getUniqueInsn() && src1->getUniqueInsn()->op == OP_SHL)
> +      s = 1;
> +   else
> +      return false;
> +
> +   src = add->getSrc(s);
> +
> +   if (src->getUniqueInsn()->bb != add->bb)
> +      return false;

You flip between UniqueInsn and Insn... I think UniqueInsn is fine,
but you could also save the result off into another Instruction *shl
and keep reusing that. IMHO it'll be easier to follow the logic.

> +
> +   if (src->getInsn()->usesFlags() || src->getInsn()->subOp)
> +      return false;
> +
> +   if (src->getInsn()->src(1).getFile() != FILE_IMMEDIATE)

This is happening before LoadPropagation, so that may may not always
work (e.g. it could be a mov). I think you want to use getImmediate()
or however that function is called so that you can peer through moves.

> +      return false;
> +
> +   mod[0] = add->src(0).mod;
> +   mod[1] = add->src(1).mod;
> +
> +   add->op = OP_SHLADD;
> +   add->setSrc(2, add->src(s ? 0 : 1));
> +   add->src(2).mod = mod[s];
> +
> +   add->setSrc(0, src->getInsn()->getSrc(0));
> +   add->src(0).mod = mod[0];

mod[!s] presumably?

IMO this would be more readable with judicious use of swapSources and
moveSources.

> +   add->setSrc(1, src->getInsn()->getSrc(1));
> +   add->src(1).mod = Modifier(0);
> +
> +   return true;
> +}
> +
> +bool
> +LateAlgebraicOpt::visit(BasicBlock *bb)
> +{
> +   Instruction *next;
> +
> +   for (Instruction *i = bb->getEntry(); i; i = next) {
> +      next = i->next;
> +      switch (i->op) {
> +      case OP_ADD:
> +         handleADD(i);
> +         break;
> +      default:
> +         break;
> +      }
> +   }
> +
> +   return true;
> +}
> +
> +// =============================================================================
> +
>  static inline void
>  updateLdStOffset(Instruction *ldst, int32_t offset, Function *fn)
>  {
> @@ -3436,6 +3529,7 @@ 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, LoadPropagation, run);
>     RUN_PASS(1, IndirectPropagation, run);
>     RUN_PASS(2, MemoryOpt, run);
> --
> 2.10.0
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev