Nouveau - Apr 2017 - [PATCH v2 0/3] nv50/ir: Preapre for running Opts inside a loop

Slowly we are getting to the point, that we miss enough optimization
opportunities as the result of our own passes.

For this we need to fix AlgebraicOpt to be able to handle mods on sources
without creating new issues.

The last patch enables looping opts.

v2: update commit author

Karol Herbst (3):
  nv50/ir: fix AlgebraicOpt for slcts with mods
  nv50/ir: handle logops with NOT in AlgebraicOpt
  nv50/ir: run some passes multiple times

 .../drivers/nouveau/codegen/nv50_ir_peephole.cpp   | 29 +++++++++++++++-------
 1 file changed, 20 insertions(+), 9 deletions(-)

-- 
2.12.2

Signed-off-by: Karol Herbst <karolherbst at gmail.com>
---
 src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
index 4c92a1efb5..bd60a84998 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
@@ -1797,10 +1797,10 @@ AlgebraicOpt::handleSLCT(Instruction *slct)
       if
(slct->getSrc(2)->asImm()->compare(slct->asCmp()->setCond, 0.0f))
          slct->setSrc(0, slct->getSrc(1));
    } else
-   if (slct->getSrc(0) != slct->getSrc(1)) {
+   if (slct->getSrc(0) != slct->getSrc(1) || slct->src(0).mod !=
slct->src(1).mod)
       return;
-   }
-   slct->op = OP_MOV;
+   slct->op = slct->src(0).mod.getOp();
+   slct->src(0).mod = slct->src(0).mod ^ Modifier(slct->op);
    slct->setSrc(1, NULL);
    slct->setSrc(2, NULL);
 }
-- 
2.12.2

Signed-off-by: Karol Herbst <karolherbst at gmail.com>
---
 src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
index bd60a84998..0de84fe9fc 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
@@ -1856,6 +1856,12 @@ AlgebraicOpt::handleLOGOP(Instruction *logop)
 
       set0 = cloneForward(func, set0);
       set1 = cloneShallow(func, set1);
+
+      if (logop->src(0).mod == Modifier(NV50_IR_MOD_NOT))
+         set0->asCmp()->setCond =
inverseCondCode(set0->asCmp()->setCond);
+      if (logop->src(1).mod == Modifier(NV50_IR_MOD_NOT))
+         set1->asCmp()->setCond =
inverseCondCode(set1->asCmp()->setCond);
+
       logop->bb->insertAfter(logop, set1);
       logop->bb->insertAfter(logop, set0);
 
-- 
2.12.2

With the shader cache, compilation time matters less.

As a side effect we can write more optimizations to produce better optimized
code.

total instructions in shared programs : 3931743 -> 3917512 (-0.36%)
total gprs used in shared programs    : 481460 -> 481680 (0.05%)
total local used in shared programs   : 27481 -> 26761 (-2.62%)
total bytes used in shared programs   : 36032672 -> 35902648 (-0.36%)

                local        gpr       inst      bytes
    helped          48         133        3843        3843
      hurt           1         295          75          75

Signed-off-by: Karol Herbst <karolherbst at gmail.com>
---
 .../drivers/nouveau/codegen/nv50_ir_peephole.cpp        | 17 +++++++++++------
 1 file changed, 11 insertions(+), 6 deletions(-)

diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
index 0de84fe9fc..505de08573 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
@@ -3729,12 +3729,17 @@ Program::optimizeSSA(int level)
    RUN_PASS(1, CopyPropagation, run);
    RUN_PASS(1, MergeSplits, run);
    RUN_PASS(2, GlobalCSE, run);
-   RUN_PASS(1, LocalCSE, run);
-   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);
+   for (int i = 0; i < 2; ++i) {
+      RUN_PASS(1, LocalCSE, run);
+      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);
+      // only once
+      if (i == 0)
+         RUN_PASS(1, Split64BitOpPreRA, run);
+      RUN_PASS(1, DeadCodeElim, buryAll);
+   }
    RUN_PASS(1, LoadPropagation, run);
    RUN_PASS(1, IndirectPropagation, run);
    RUN_PASS(2, MemoryOpt, run);
-- 
2.12.2

On Mon, Apr 3, 2017 at 11:58 AM, Karol Herbst <karolherbst at gmail.com>
wrote:
> Signed-off-by: Karol Herbst <karolherbst at gmail.com>
> ---
>  src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp | 6 +++---
>  1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
> index 4c92a1efb5..bd60a84998 100644
> --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
> +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
> @@ -1797,10 +1797,10 @@ AlgebraicOpt::handleSLCT(Instruction *slct)
>        if
(slct->getSrc(2)->asImm()->compare(slct->asCmp()->setCond, 0.0f))
>           slct->setSrc(0, slct->getSrc(1));
>     } else
> -   if (slct->getSrc(0) != slct->getSrc(1)) {
> +   if (slct->getSrc(0) != slct->getSrc(1) || slct->src(0).mod !=
slct->src(1).mod)
SLCT can't have mods on src0/src1. Only on src2. I'd be just as happy
to assert that they're both == 0 here. You can also add a helper to
ValueRef to see if it's == to another ValueRef, which compares both
the Value ptr as well as any modifiers, indirects, etc. But it again
doesn't ultimately need to be used here.
>        return;
> -   }
> -   slct->op = OP_MOV;
> +   slct->op = slct->src(0).mod.getOp();
> +   slct->src(0).mod = slct->src(0).mod ^ Modifier(slct->op);
>     slct->setSrc(1, NULL);
>     slct->setSrc(2, NULL);
>  }
> --
> 2.12.2
>

On Mon, Apr 3, 2017 at 11:58 AM, Karol Herbst <karolherbst at gmail.com>
wrote:
> Signed-off-by: Karol Herbst <karolherbst at gmail.com>
> ---
>  src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp | 6 ++++++
>  1 file changed, 6 insertions(+)
>
> diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
> index bd60a84998..0de84fe9fc 100644
> --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
> +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_peephole.cpp
> @@ -1856,6 +1856,12 @@ AlgebraicOpt::handleLOGOP(Instruction *logop)
>
>        set0 = cloneForward(func, set0);
>        set1 = cloneShallow(func, set1);
> +
> +      if (logop->src(0).mod == Modifier(NV50_IR_MOD_NOT))
> +         set0->asCmp()->setCond =
inverseCondCode(set0->asCmp()->setCond);
> +      if (logop->src(1).mod == Modifier(NV50_IR_MOD_NOT))
> +         set1->asCmp()->setCond =
inverseCondCode(set1->asCmp()->setCond);
set0/set1 may have been swapped further up, so you need to keep track of that.

Also, I don't think this will work if one of the sets is a SET_AND --
the condcode applies to the set bit, not to the AND bit. I think you'd
also have to flip AND <-> OR and flip the neg.

  -ilia