nir/algebraic: Reassociate add-and-shift to be shift-and-add

A common thing in many shaders:

    uniform vs { vec4 bones[...]; };

    ...

    x = some_calculation(bones[i + 0]);
    y = some_calculation(bones[i + 1]);
    z = some_calculation(bones[i + 2]);

This turns into stuff like

    vec1 32 ssa_12 = iadd ssa_11, ssa_0
    vec1 32 ssa_13 = ishl ssa_12, ssa_3
    vec1 32 ssa_14 = intrinsic load_ssbo (ssa_7, ssa_13) (16, 4, 0)
    vec1 32 ssa_15 = iadd ssa_11, ssa_1
    vec1 32 ssa_16 = ishl ssa_15, ssa_3
    vec1 32 ssa_17 = intrinsic load_ssbo (ssa_7, ssa_16) (16, 4, 0)
    vec1 32 ssa_18 = iadd ssa_11, ssa_2
    vec1 32 ssa_19 = ishl ssa_18, ssa_3
    vec1 32 ssa_20 = intrinsic load_ssbo (ssa_7, ssa_19) (16, 4, 0)

By reassociating the shift and the add, we can reduce this to

    vec1 32 ssa_12 = ishl ssa_11, ssa_3
    vec1 32 ssa_13 = iadd ssa_12, ssa_0
    vec1 32 ssa_14 = intrinsic load_ssbo (ssa_7, ssa_13) (16, 4, 0)
    vec1 32 ssa_16 = iadd ssa_12, ssa_1
    vec1 32 ssa_17 = intrinsic load_ssbo (ssa_7, ssa_16) (16, 4, 0)
    vec1 32 ssa_19 = iadd ssa_12, ssa_2
    vec1 32 ssa_20 = intrinsic load_ssbo (ssa_7, ssa_19) (16, 4, 0)

v2: Add some commentary from Rhys Perry's nearly identical patch.

All Intel platforms had similar results. (Ice Lake shown)
total instructions in shared programs: 16277758 -> 16250704 (-0.17%)
instructions in affected programs: 1440284 -> 1413230 (-1.88%)
helped: 4920
HURT: 6
helped stats (abs) min: 1 max: 69 x̄: 5.50 x̃: 4
helped stats (rel) min: 0.10% max: 18.33% x̄: 2.21% x̃: 1.79%
HURT stats (abs)   min: 1 max: 12 x̄: 4.50 x̃: 3
HURT stats (rel)   min: 0.18% max: 3.23% x̄: 1.91% x̃: 2.55%
95% mean confidence interval for instructions value: -5.67 -5.31
95% mean confidence interval for instructions %-change: -2.26% -2.16%
Instructions are helped.

total cycles in shared programs: 367118526 -> 365895358 (-0.33%)
cycles in affected programs: 93504145 -> 92280977 (-1.31%)
helped: 2754
HURT: 1269
helped stats (abs) min: 1 max: 47039 x̄: 460.66 x̃: 16
helped stats (rel) min: <.01% max: 34.93% x̄: 3.77% x̃: 1.12%
HURT stats (abs)   min: 1 max: 1500 x̄: 35.85 x̃: 9
HURT stats (rel)   min: 0.01% max: 17.35% x̄: 2.18% x̃: 0.75%
95% mean confidence interval for cycles value: -387.31 -220.78
95% mean confidence interval for cycles %-change: -2.11% -1.68%
Cycles are helped.

LOST:   1
GAINED: 1

Reviewed-by: Alyssa Rosenzweig <[email protected]>
Reviewed-by: Jason Ekstrand <[email protected]>

1 files changed, 5 insertions, 0 deletions

diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index a3a244c41fb..bf5c69314c5 100644
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -225,6 +225,11 @@ optimizations = [
    # a * (#b << #c)
    (('ishl', ('imul', a, '#b'), '#c'), ('imul', a, ('ishl', b, c))),
 
+   # This is common for address calculations.  Reassociating may enable the
+   # 'a<<c' to be CSE'd.  It also helps architectures that have an ISHLADD
+   # instruction or a constant offset field for in load / store instructions.
+   (('ishl', ('iadd', a, '#b'), '#c'), ('iadd', ('ishl', a, c), ('ishl', b, c))),
+
    # Comparison simplifications
    (('~inot', ('flt', a, b)), ('fge', a, b)),
    (('~inot', ('fge', a, b)), ('flt', a, b)),